Load initial app

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.keras filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,5 @@
+__pycache__/
+*.pyc
+.venv/
+venv/
+

EDSR_x4.pb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd35ce3cae53ecee2d16045e08a932c3e7242d641bb65cb971d123e06904347f
+size 38573255

LICENSE

@@ -0,0 +1,674 @@
+                    GNU GENERAL PUBLIC LICENSE
+                       Version 3, 29 June 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU General Public License is a free, copyleft license for
+software and other kinds of works.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+the GNU General Public License is intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.  We, the Free Software Foundation, use the
+GNU General Public License for most of our software; it applies also to
+any other work released this way by its authors.  You can apply it to
+your programs, too.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  To protect your rights, we need to prevent others from denying you
+these rights or asking you to surrender the rights.  Therefore, you have
+certain responsibilities if you distribute copies of the software, or if
+you modify it: responsibilities to respect the freedom of others.
+
+  For example, if you distribute copies of such a program, whether
+gratis or for a fee, you must pass on to the recipients the same
+freedoms that you received.  You must make sure that they, too, receive
+or can get the source code.  And you must show them these terms so they
+know their rights.
+
+  Developers that use the GNU GPL protect your rights with two steps:
+(1) assert copyright on the software, and (2) offer you this License
+giving you legal permission to copy, distribute and/or modify it.
+
+  For the developers' and authors' protection, the GPL clearly explains
+that there is no warranty for this free software.  For both users' and
+authors' sake, the GPL requires that modified versions be marked as
+changed, so that their problems will not be attributed erroneously to
+authors of previous versions.
+
+  Some devices are designed to deny users access to install or run
+modified versions of the software inside them, although the manufacturer
+can do so.  This is fundamentally incompatible with the aim of
+protecting users' freedom to change the software.  The systematic
+pattern of such abuse occurs in the area of products for individuals to
+use, which is precisely where it is most unacceptable.  Therefore, we
+have designed this version of the GPL to prohibit the practice for those
+products.  If such problems arise substantially in other domains, we
+stand ready to extend this provision to those domains in future versions
+of the GPL, as needed to protect the freedom of users.
+
+  Finally, every program is threatened constantly by software patents.
+States should not allow patents to restrict development and use of
+software on general-purpose computers, but in those that do, we wish to
+avoid the special danger that patents applied to a free program could
+make it effectively proprietary.  To prevent this, the GPL assures that
+patents cannot be used to render the program non-free.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Use with the GNU Affero General Public License.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU Affero General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the special requirements of the GNU Affero General Public License,
+section 13, concerning interaction through a network will apply to the
+combination as such.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU General Public License from time to time.  Such new versions will
+be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If the program does terminal interaction, make it output a short
+notice like this when it starts in an interactive mode:
+
+    <program>  Copyright (C) <year>  <name of author>
+    This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
+    This is free software, and you are welcome to redistribute it
+    under certain conditions; type `show c' for details.
+
+The hypothetical commands `show w' and `show c' should show the appropriate
+parts of the General Public License.  Of course, your program's commands
+might be different; for a GUI interface, you would use an "about box".
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU GPL, see
+<https://www.gnu.org/licenses/>.
+
+  The GNU General Public License does not permit incorporating your program
+into proprietary programs.  If your program is a subroutine library, you
+may consider it more useful to permit linking proprietary applications with
+the library.  If this is what you want to do, use the GNU Lesser General
+Public License instead of this License.  But first, please read
+<https://www.gnu.org/licenses/why-not-lgpl.html>.

README.md

@@ -1,6 +1,6 @@
 ---
-title: Deki
-emoji: 🐢
+title: deki
+emoji: 📱
 colorFrom: blue
 colorTo: blue
 sdk: gradio

app.py

@@ -1,7 +1,322 @@
 import gradio as gr
+import os
+import base64
+import time
+import json
+import logging
+import tempfile
+import uuid
+import io
 
-def greet(name):
-    return "Hello " + name + "!!"
+from PIL import Image
+from openai import OpenAI
+from ultralytics import YOLO
 
-demo = gr.Interface(fn=greet, inputs="text", outputs="text")
+from wrapper import process_image_description
+from utils.pills import preprocess_image
+import cv2
+import cv2.dnn_superres as dnn_superres
+import easyocr
+from spellchecker import SpellChecker
+
+logging.basicConfig(level=logging.INFO, format='%(asctime)s %(levelname)s %(message)s')
+
+GLOBAL_SR = None
+GLOBAL_READER = None
+GLOBAL_SPELL = None
+YOLO_MODEL = None
+
+def load_models():
+    """
+    Called once to load all necessary models into memory.
+    """
+    global GLOBAL_SR, GLOBAL_READER, GLOBAL_SPELL, YOLO_MODEL
+
+    logging.info("Loading all models...")
+    start_time_total = time.perf_counter()
+
+    # Super-resolution
+    logging.info("Loading super-resolution model...")
+    start_time = time.perf_counter()
+    sr = None
+    model_path = "EDSR_x4.pb"
+    if os.path.exists(model_path):
+        if hasattr(cv2, 'dnn_superres'):
+            try:
+                sr = dnn_superres.DnnSuperResImpl_create()
+            except AttributeError:
+                sr = dnn_superres.DnnSuperResImpl()
+            sr.readModel(model_path)
+            sr.setModel('edsr', 4)
+            GLOBAL_SR = sr
+            logging.info("Super-resolution model loaded.")
+        else:
+            logging.warning("cv2.dnn_superres module not available.")
+    else:
+        logging.warning(f"Super-resolution model file not found: {model_path}. Skipping SR.")
+    logging.info(f"Super-resolution init took {time.perf_counter()-start_time:.3f}s.")
+
+    # EasyOCR + SpellChecker
+    logging.info("Loading OCR + SpellChecker...")
+    start_time = time.perf_counter()
+    GLOBAL_READER = easyocr.Reader(['en'], gpu=True)
+    GLOBAL_SPELL = SpellChecker()
+    logging.info(f"OCR + SpellChecker init took {time.perf_counter()-start_time:.3f}s.")
+
+    # YOLO Model
+    logging.info("Loading YOLO model...")
+    start_time = time.perf_counter()
+    yolo_weights = "best.pt"
+    if os.path.exists(yolo_weights):
+        YOLO_MODEL = YOLO(yolo_weights)
+        logging.info("YOLO model loaded.")
+    else:
+        logging.error(f"YOLO weights file '{yolo_weights}' not found! Endpoints will fail.")
+    logging.info(f"YOLO init took {time.perf_counter()-start_time:.3f}s.")
+    
+    logging.info(f"Total model loading time: {time.perf_counter()-start_time_total:.3f}s.")
+
+
+def pil_to_base64_str(pil_image, format="PNG"):
+    """Converts a PIL Image to a base64 string with a data URI header."""
+    buffered = io.BytesIO()
+    pil_image.save(buffered, format=format)
+    img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
+    return f"data:image/{format.lower()};base64,{img_str}"
+
+def save_base64_image(image_data: str, file_path: str):
+    """Saves a base64 encoded image to a file."""
+    if image_data.startswith("data:image"):
+        _, image_data = image_data.split(",", 1)
+    img_bytes = base64.b64decode(image_data)
+    with open(file_path, "wb") as f:
+        f.write(img_bytes)
+    return img_bytes
+
+def run_wrapper(image_path: str, output_dir: str, skip_ocr: bool = False, skip_spell: bool = False, json_mini=False) -> str:
+    """Calls the main processing script and returns the result."""
+    process_image_description(
+        input_image=image_path,
+        weights_file="best.pt",
+        output_dir=output_dir,
+        no_captioning=True,
+        output_json=True,
+        json_mini=json_mini,
+        model_obj=YOLO_MODEL,
+        sr=GLOBAL_SR,
+        spell=None if skip_ocr else GLOBAL_SPELL,
+        reader=None if skip_ocr else GLOBAL_READER,
+        skip_ocr=skip_ocr,
+        skip_spell=skip_spell,
+    )
+    base_name = os.path.splitext(os.path.basename(image_path))[0]
+    result_dir = os.path.join(output_dir, "result")
+    json_file = os.path.join(result_dir, f"{base_name}.json")
+    if os.path.exists(json_file):
+        with open(json_file, "r", encoding="utf-8") as f:
+            return f.read()
+    else:
+        raise FileNotFoundError(f"Result file not generated: {json_file}")
+
+def handle_action(openai_key, image, prompt):
+    if not openai_key: return "Error: OpenAI API Key is required for /action."
+    if image is None: return "Error: Please upload an image."
+    if not prompt: return "Error: Please provide a prompt."
+
+    try:
+        llm_client = OpenAI(api_key=openai_key)
+        image_b64 = pil_to_base64_str(image)
+        
+        with tempfile.TemporaryDirectory() as temp_dir:
+            request_id = str(uuid.uuid4())
+            original_image_path = os.path.join(temp_dir, f"{request_id}.png")
+            yolo_updated_image_path = os.path.join(temp_dir, f"{request_id}_yolo_updated.png")
+            save_base64_image(image_b64, original_image_path)
+            
+            image_description = run_wrapper(original_image_path, temp_dir, skip_ocr=False, skip_spell=True, json_mini=True)
+            
+            with open(yolo_updated_image_path, "rb") as f:
+                yolo_updated_img_bytes = f.read()
+            
+            _, new_b64 = preprocess_image(yolo_updated_img_bytes, threshold=2000, scale=0.5, fmt="png")
+
+        base64_image_url = f"data:image/png;base64,{new_b64}"
+        prompt_text = f"""You are an AI agent... (rest of your long prompt)
+        The user said: "{prompt}"
+        Description: "{image_description}" """
+        
+        messages = [{"role": "user", "content": [{"type": "text", "text": prompt_text}, {"type": "image_url", "image_url": {"url": base64_image_url, "detail": "high"}}]}]
+        
+        response = llm_client.chat.completions.create(model="gpt-4.1", messages=messages, temperature=0.2)
+        return response.choices[0].message.content.strip()
+
+    except Exception as e:
+        logging.error(f"Error in /action endpoint: {e}", exc_info=True)
+        return f"An error occurred: {e}"
+
+def handle_analyze(image, output_style):
+    if image is None: return "Error: Please upload an image."
+    
+    try:
+        image_b64 = pil_to_base64_str(image)
+        with tempfile.TemporaryDirectory() as temp_dir:
+            image_path = os.path.join(temp_dir, "image_to_analyze.png")
+            save_base64_image(image_b64, image_path)
+            
+            is_mini = (output_style == "Mini JSON")
+            description_str = run_wrapper(image_path=image_path, output_dir=temp_dir, json_mini=is_mini)
+            
+            parsed_json = json.loads(description_str)
+            return json.dumps(parsed_json, indent=2)
+
+    except Exception as e:
+        logging.error(f"Error in /analyze endpoint: {e}", exc_info=True)
+        return f"An error occurred: {e}"
+
+def handle_analyze_yolo(image, output_style):
+    if image is None: return None, "Error: Please upload an image."
+
+    try:
+        image_b64 = pil_to_base64_str(image)
+        with tempfile.TemporaryDirectory() as temp_dir:
+            request_id = str(uuid.uuid4())
+            image_path = os.path.join(temp_dir, f"{request_id}.png")
+            yolo_image_path = os.path.join(temp_dir, f"{request_id}_yolo_updated.png")
+            save_base64_image(image_b64, image_path)
+
+            is_mini = (output_style == "Mini JSON")
+            description_str = run_wrapper(image_path=image_path, output_dir=temp_dir, json_mini=is_mini)
+
+            parsed_json = json.loads(description_str)
+            description_output = json.dumps(parsed_json, indent=2)
+            
+            yolo_image_result = Image.open(yolo_image_path)
+            return yolo_image_result, description_output
+
+    except Exception as e:
+        logging.error(f"Error in /analyze_and_get_yolo: {e}", exc_info=True)
+        return None, f"An error occurred: {e}"
+
+def handle_generate(openai_key, image, prompt):
+    if not openai_key: return "Error: OpenAI API Key is required for /generate."
+    if image is None: return "Error: Please upload an image."
+    if not prompt: return "Error: Please provide a prompt."
+
+    try:
+        llm_client = OpenAI(api_key=openai_key)
+        image_b64 = pil_to_base64_str(image)
+
+        with tempfile.TemporaryDirectory() as temp_dir:
+            request_id = str(uuid.uuid4())
+            original_image_path = os.path.join(temp_dir, f"{request_id}.png")
+            yolo_updated_image_path = os.path.join(temp_dir, f"{request_id}_yolo_updated.png")
+            save_base64_image(image_b64, original_image_path)
+
+            image_description = run_wrapper(image_path=original_image_path, output_dir=temp_dir, json_mini=False)
+            
+            with open(yolo_updated_image_path, "rb") as f:
+                yolo_updated_img_bytes = f.read()
+            
+            _, new_b64 = preprocess_image(yolo_updated_img_bytes, threshold=1500, scale=0.5, fmt="png")
+
+        base64_image_url = f"data:image/png;base64,{new_b64}"
+        messages = [
+            {"role": "user", "content": [
+                {"type": "text", "text": f'"Prompt: {prompt}"\nImage description:\n"{image_description}"'},
+                {"type": "image_url", "image_url": {"url": base64_image_url, "detail": "high"}}
+            ]}
+        ]
+    
+        response = llm_client.chat.completions.create(model="gpt-4.1", messages=messages, temperature=0.2)
+        return response.choices[0].message.content.strip()
+
+    except Exception as e:
+        logging.error(f"Error in /generate endpoint: {e}", exc_info=True)
+        return f"An error occurred: {e}"
+
+default_image_1 = Image.open("./res/bb_1.jpeg")
+default_image_2 = Image.open("./res/mfa_1.jpeg")
+
+def load_example_action_1(): return default_image_1, "Open and read Umico partner"
+def load_example_action_2(): return default_image_2, "Sign up in the application"
+def load_example_analyze_1(): return default_image_1
+def load_example_analyze_2(): return default_image_2
+def load_example_yolo_1(): return default_image_1
+def load_example_yolo_2(): return default_image_2
+def load_example_generate_1(): return default_image_1, "Generate the code for this screen for Android XML. Try to use constraint layout"
+def load_example_generate_2(): return default_image_2, "Generate the code for this screen for Android XML. Try to use constraint layout"
+
+
+with gr.Blocks(theme=gr.themes.Soft()) as demo:
+    gr.Markdown("# Deki Automata: UI Analysis and Generation")
+    gr.Markdown("Provide your API keys below. The OpenAI key is only required for the 'Action' and 'Generate' tabs.")
+    
+    with gr.Row():
+        openai_key_input = gr.Textbox(label="OpenAI API Key", placeholder="Enter your OpenAI API Key", type="password", scale=1)
+    
+    with gr.Tabs():
+        with gr.TabItem("Action"):
+            gr.Markdown("### Control a device with natural language.")
+            with gr.Row():
+                image_input_action = gr.Image(type="pil", label="Upload Screen Image")
+                prompt_input_action = gr.Textbox(lines=2, placeholder="e.g., 'Open whatsapp and text my friend...'", label="Prompt")
+            action_output = gr.Textbox(label="Response Command")
+            action_button = gr.Button("Run Action", variant="primary")
+            with gr.Row():
+                example_action_btn1 = gr.Button("Load Example 1")
+                example_action_btn2 = gr.Button("Load Example 2")
+        
+        with gr.TabItem("Analyze"):
+            gr.Markdown("### Get a structured JSON description of the UI elements.")
+            with gr.Row():
+                image_input_analyze = gr.Image(type="pil", label="Upload Screen Image")
+                with gr.Column():
+                    output_style_analyze = gr.Radio(["Standard JSON", "Mini JSON"], label="Output Format", value="Standard JSON")
+                    analyze_button = gr.Button("Analyze Image", variant="primary")
+            analyze_output = gr.JSON(label="JSON Description")
+            with gr.Row():
+                example_analyze_btn1 = gr.Button("Load Example 1")
+                example_analyze_btn2 = gr.Button("Load Example 2")
+
+        with gr.TabItem("Analyze & Get YOLO"):
+            gr.Markdown("### Get a JSON description and the image with detected elements.")
+            with gr.Row():
+                image_input_yolo = gr.Image(type="pil", label="Upload Screen Image")
+                with gr.Column():
+                     output_style_yolo = gr.Radio(["Standard JSON", "Mini JSON"], label="Output Format", value="Standard JSON")
+                     yolo_button = gr.Button("Analyze and Visualize", variant="primary")
+            with gr.Row():
+                yolo_image_output = gr.Image(label="YOLO Annotated Image")
+                description_output_yolo = gr.JSON(label="JSON Description")
+            with gr.Row():
+                example_yolo_btn1 = gr.Button("Load Example 1")
+                example_yolo_btn2 = gr.Button("Load Example 2")
+
+        with gr.TabItem("Generate"):
+            gr.Markdown("### Generate code or text based on a screenshot.")
+            with gr.Row():
+                image_input_generate = gr.Image(type="pil", label="Upload Screen Image")
+                prompt_input_generate = gr.Textbox(lines=2, placeholder="e.g., 'Generate the Android XML for this screen'", label="Prompt")
+            generate_output = gr.Code(label="Generated Output", language="xml")
+            generate_button = gr.Button("Generate", variant="primary")
+            with gr.Row():
+                example_generate_btn1 = gr.Button("Load Example 1")
+                example_generate_btn2 = gr.Button("Load Example 2")
+
+    action_button.click(fn=handle_action, inputs=[openai_key_input, image_input_action, prompt_input_action], outputs=action_output)
+    analyze_button.click(fn=handle_analyze, inputs=[image_input_analyze, output_style_analyze], outputs=analyze_output)
+    yolo_button.click(fn=handle_analyze_yolo, inputs=[image_input_yolo, output_style_yolo], outputs=[yolo_image_output, description_output_yolo])
+    generate_button.click(fn=handle_generate, inputs=[openai_key_input, image_input_generate, prompt_input_generate], outputs=generate_output)
+
+    example_action_btn1.click(fn=load_example_action_1, outputs=[image_input_action, prompt_input_action])
+    example_action_btn2.click(fn=load_example_action_2, outputs=[image_input_action, prompt_input_action])
+    example_analyze_btn1.click(fn=load_example_analyze_1, outputs=image_input_analyze)
+    example_analyze_btn2.click(fn=load_example_analyze_2, outputs=image_input_analyze)
+    example_yolo_btn1.click(fn=load_example_yolo_1, outputs=image_input_yolo)
+    example_yolo_btn2.click(fn=load_example_yolo_2, outputs=image_input_yolo)
+    example_generate_btn1.click(fn=load_example_generate_1, outputs=[image_input_generate, prompt_input_generate])
+    example_generate_btn2.click(fn=load_example_generate_2, outputs=[image_input_generate, prompt_input_generate])
+
+load_models()
 demo.launch()
+

best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:879fd9bd951bce6910815fbe54b66a1970d903326d276c3d7cad19db798d0c2c
+size 33455351

icon-image-detection-model.keras

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd08501457e8dee65d173483026744d52d0d12c58589a3ee03cc6ad3c2e9cdd3
+size 245597405

main.py

@@ -0,0 +1,472 @@
+import os
+import base64
+import time
+import easyocr
+from spellchecker import SpellChecker
+import cv2
+import cv2.dnn_superres as dnn_superres
+import json
+import asyncio
+import functools
+from fastapi.responses import JSONResponse
+from fastapi import FastAPI, HTTPException, Depends
+from fastapi.security import HTTPBearer, HTTPAuthorizationCredentials
+from pydantic import BaseModel
+import openai
+from openai import OpenAI
+from ultralytics import YOLO
+from wrapper import process_image_description
+from utils.pills import preprocess_image
+import logging
+import tempfile
+import uuid
+
+logging.basicConfig(level=logging.INFO, format='%(asctime)s %(levelname)s %(message)s')
+
+app = FastAPI(title="deki-automata API")
+
+# Global semaphore limit is set to 1 because the ML tasks take almost all RAM for the current server.
+# Can be increased in the future
+CONCURRENT_LIMIT = 2
+concurrency_semaphore = asyncio.Semaphore(CONCURRENT_LIMIT)
+
+def with_semaphore(timeout: float = 20):
+    """
+    Decorator to limit concurrent access by acquiring the semaphore
+    before the function runs, and releasing it afterward.
+    """
+    def decorator(func):
+        @functools.wraps(func)
+        async def wrapper(*args, **kwargs):
+            try:
+                await asyncio.wait_for(concurrency_semaphore.acquire(), timeout=timeout)
+            except asyncio.TimeoutError:
+                raise HTTPException(status_code=503, detail="Service busy, please try again later.")
+            try:
+                return await func(*args, **kwargs)
+            finally:
+                concurrency_semaphore.release()
+        return wrapper
+    return decorator
+
+OPENAI_API_KEY = os.environ.get("OPENAI_API_KEY")
+API_TOKEN = os.environ.get("API_TOKEN")
+
+if not OPENAI_API_KEY or not API_TOKEN:
+    logging.error("OPENAI_API_KEY and API_TOKEN must be set in environment variables.")
+    raise RuntimeError("OPENAI_API_KEY and API_TOKEN must be set in environment variables.")
+
+openai.api_key = OPENAI_API_KEY
+
+GLOBAL_SR = None
+GLOBAL_READER = None
+GLOBAL_SPELL = None
+LLM_CLIENT = None
+
+os.makedirs("./res", exist_ok=True)
+os.makedirs("./result", exist_ok=True)
+os.makedirs("./output", exist_ok=True)
+
+
+# for action step tracking
+ACTION_STEPS_LIMIT = 10 # can be updated
+
+@app.on_event("startup")
+def load_models():
+    """
+    Called once when FastAPI starts.
+    """
+    global GLOBAL_SR, GLOBAL_READER, GLOBAL_SPELL, LLM_CLIENT
+
+    # Super-resolution
+    logging.info("Loading super-resolution model ...")
+    start_time = time.perf_counter()
+    sr = None
+    model_path = "EDSR_x4.pb"
+    if hasattr(cv2, 'dnn_superres'):
+        logging.info("dnn_superres module is available.")
+        try:
+            sr = dnn_superres.DnnSuperResImpl_create()
+            logging.info("Using DnnSuperResImpl_create()")
+        except AttributeError:
+            sr = dnn_superres.DnnSuperResImpl()
+            logging.info("Using DnnSuperResImpl()")
+
+        if os.path.exists(model_path):
+            sr.readModel(model_path)
+            sr.setModel('edsr', 4)
+            GLOBAL_SR = sr
+            logging.info("Super-resolution model loaded.")
+        else:
+            logging.warning(f"Super-resolution model file not found: {model_path}. Skipping SR.")
+            GLOBAL_SR = None
+    else:
+        logging.info("dnn_superres module is NOT available; skipping super-resolution.")
+        GLOBAL_SR = None
+    logging.info(f"Super-resolution initialization took {time.perf_counter()-start_time:.3f}s.")
+
+    # EasyOCR + SpellChecker
+    logging.info("Loading OCR + SpellChecker ...")
+    start_time = time.perf_counter()
+    GLOBAL_READER = easyocr.Reader(['en'], gpu=True)
+    GLOBAL_SPELL = SpellChecker()
+    logging.info(f"OCR + SpellChecker init took {time.perf_counter()-start_time:.3f}s.")
+    LLM_CLIENT = OpenAI()
+
+class ActionRequest(BaseModel):
+    image: str  # Base64-encoded image
+    prompt: str  # User prompt (like "Open whatsapp and write my friend user_name that I will be late for 15 minutes")
+    history: list[str] = []
+
+class ActionResponse(BaseModel):
+    response: str
+    history: list[str]
+
+
+class AnalyzeRequest(BaseModel):
+    image: str  # Base64-encoded image
+
+security = HTTPBearer()
+
+def verify_token(credentials: HTTPAuthorizationCredentials = Depends(security)):
+    if credentials.credentials != API_TOKEN:
+        logging.warning("Invalid API token attempt.")
+        raise HTTPException(status_code=401, detail="Invalid API token")
+    return credentials.credentials
+
+def save_base64_image(image_data: str, file_path: str) -> bytes:
+    """
+    Decode base64 image data (removing any data URI header) and save to the specified file.
+    Returns the raw image bytes.
+    """
+    if image_data.startswith("data:image"):
+        _, image_data = image_data.split(",", 1)
+    try:
+        img_bytes = base64.b64decode(image_data)
+    except Exception as e:
+        logging.exception("Error decoding base64 image data.")
+        raise HTTPException(status_code=400, detail=f"Invalid base64 image data: {e}")
+    
+    try:
+        with open(file_path, "wb") as f:
+            f.write(img_bytes)
+    except Exception as e:
+        logging.exception("Error saving image file.")
+        raise HTTPException(status_code=500, detail=f"Failed to save image: {e}")
+    
+    return img_bytes
+
+def log_request_data(request, endpoint: str):
+    """
+    Log the user prompt (if any) and a preview of the image data.
+    """
+    logging.info(f"{endpoint} request received:")
+    if hasattr(request, 'prompt'):
+        logging.info(f"User prompt: {request.prompt}")
+    image_preview = request.image[:100] + "..." if len(request.image) > 100 else request.image
+    logging.info(f"User image data (base64 preview): {image_preview}")
+
+def run_wrapper(image_path: str, output_dir: str, skip_ocr: bool = False, skip_spell: bool = False, json_mini = False) -> str:
+    """
+    Calls process_image_description() to perform YOLO detection and image description,
+    then reads the resulting JSON or text file from ./result.
+    """
+
+    weights_file = "best.pt"
+    no_captioning = True
+    output_json = True
+
+    process_image_description(
+        input_image=image_path,
+        weights_file=weights_file,
+        output_dir=output_dir,
+        no_captioning=no_captioning,
+        output_json=output_json,
+        json_mini=json_mini,
+        sr=GLOBAL_SR,
+        spell=None if skip_ocr else GLOBAL_SPELL,
+        reader=None if skip_ocr else GLOBAL_READER,
+        skip_ocr=skip_ocr,
+        skip_spell=skip_spell,
+    )
+
+    base_name = os.path.splitext(os.path.basename(image_path))[0]
+    result_dir = os.path.join(output_dir, "result") 
+    json_file = os.path.join(result_dir, f"{base_name}.json")
+    txt_file = os.path.join(result_dir, f"{base_name}.txt")
+
+    if os.path.exists(json_file):
+        try:
+            with open(json_file, "r", encoding="utf-8") as f:
+                return f.read()
+        except Exception as e:
+            logging.exception("Failed to read JSON description file.")
+            raise Exception(f"Failed to read JSON description file: {e}")
+    elif os.path.exists(txt_file):
+        try:
+            with open(txt_file, "r", encoding="utf-8") as f:
+                return f.read()
+        except Exception as e:
+            logging.exception("Failed to read TXT description file.")
+            raise Exception(f"Failed to read TXT description file: {e}")
+    else:
+        logging.error("No image description file was generated.")
+        raise FileNotFoundError("No image description file was generated.")
+
+@app.get("/")
+async def root():
+    return {"message": "deki"}
+
+@app.post("/action", response_model=ActionResponse)
+@with_semaphore(timeout=60)
+async def action(request: ActionRequest, token: str = Depends(verify_token)):
+    """
+    Processes the input image (in base64 format) and a user prompt:
+    1. Decodes and saves the original image.
+    2. Runs the wrapper to generate an image description and the YOLO-updated image file.
+    3. Reads the YOLO-updated image file.
+    4. Preprocesses the YOLO-updated image.
+    5. Constructs a prompt for ChatGPT (using description + preprocessed YOLO image) and sends it.
+    6. Returns the command response.
+    """
+    start_time = time.perf_counter()
+    logging.info("action endpoint start")
+    log_request_data(request, "/action")
+
+    action_step_history = request.history
+    action_step_count = len(action_step_history)
+
+    # Check if the step limit is reached.
+    if action_step_count >= ACTION_STEPS_LIMIT:
+        logging.warning(f"Step limit of {ACTION_STEPS_LIMIT} reached. Resetting history.")
+        # Return a clear response and an empty history to reset the client.
+        return ActionResponse(response="Step limit is reached", history=[])
+
+    # Use a temporary directory to isolate all files for this request.
+    with tempfile.TemporaryDirectory() as temp_dir:
+        request_id = str(uuid.uuid4())
+        
+        original_image_path = os.path.join(temp_dir, f"{request_id}.png")
+        yolo_updated_image_path = os.path.join(temp_dir, f"{request_id}_yolo_updated.png")
+
+        save_base64_image(request.image, original_image_path)
+
+        try:
+            loop = asyncio.get_running_loop()
+            image_description = await loop.run_in_executor(
+                None,
+                run_wrapper,
+                original_image_path,
+                temp_dir,
+                False,
+                True,
+                True,
+            )
+        except Exception as e:
+            logging.exception("Image processing failed in action endpoint.")
+            raise HTTPException(status_code=500, detail=f"Image processing failed: {e}")
+
+        try:
+            if not os.path.exists(yolo_updated_image_path):
+                logging.error(f"YOLO updated image not found at {yolo_updated_image_path}")
+                raise HTTPException(status_code=500, detail="YOLO updated image generation failed or not found.")
+            with open(yolo_updated_image_path, "rb") as f:
+                yolo_updated_img_bytes = f.read()
+        except Exception as e:
+            logging.exception(f"Error reading YOLO updated image from {yolo_updated_image_path}")
+            raise HTTPException(status_code=500, detail=f"Failed to read YOLO updated image: {e}")
+        
+        try:
+            _, new_b64 = preprocess_image(yolo_updated_img_bytes, threshold=2000, scale=0.5, fmt="png")
+        except Exception as e:
+            logging.exception("YOLO updated image preprocessing failed.")
+            raise HTTPException(status_code=500, detail=f"YOLO updated image preprocessing failed: {e}")
+
+    base64_image_url = f"data:image/png;base64,{new_b64}"
+
+    current_step = action_step_count + 1
+    previous_steps_text = ""
+    if action_step_history:
+        previous_steps_text = "\nPrevious steps:\n" + "\n".join(f"{i+1}. {step}" for i, step in enumerate(action_step_history))
+
+    prompt_text = f"""You are an AI agent that controls a mobile device and sees the content of screen.
+User can ask you about some information or to do some task and you need to do these tasks.
+You can only respond with one of these commands (in quotes) but some variables are dynamic
+and can be changed based on the context:
+1. "Swipe left. From start coordinates 300, 400" (or other coordinates) (Goes right)
+2. "Swipe right. From start coordinates 500, 650" (or other coordinates) (Goes left)
+3. "Swipe top. From start coordinates 600, 510" (or other coordinates) (Goes bottom)
+4. "Swipe bottom. From start coordinates 640, 500" (or other coordinates) (Goes top)
+5. "Go home"
+6. "Go back"
+8. "Open com.whatsapp" (or other app)
+9. "Tap coordinates 160, 820" (or other coordinates)
+10. "Insert text 210, 820:Hello world" (or other coordinates and text)
+11. "Screen is in a loading state. Try again" (send image again)
+12. "Answer: There are no new important mails today" (or other answer)
+13. "Finished" (task is finished)
+14. "Can't proceed" (can't understand what to do or image has problem etc.)
+
+The user said: "{request.prompt}"
+
+Current step: {current_step}
+{previous_steps_text}
+
+I will share the screenshot of the current state of the phone (with UI elements highlighted and the corresponding 
+index of these UI elements) and the description (sizes, coordinates and indexes) of UI elements.
+Description:
+"{image_description}" """
+    
+    messages = [
+        {"role": "user", "content": [{"type": "text", "text": prompt_text}, {"type": "image_url", "image_url": {"url": base64_image_url, "detail": "high"}}]}
+    ]
+
+    try:
+        response = LLM_CLIENT.chat.completions.create(model="gpt-4.1", messages=messages, temperature=0.2)
+    except Exception as e:
+        logging.exception("OpenAI API error.")
+        raise HTTPException(status_code=500, detail=f"OpenAI API error: {e}")
+
+    command_response = response.choices[0].message.content.strip()
+
+    action_step_history.append(command_response)
+
+    command_lower = command_response.strip().strip('\'"').lower()
+    if command_lower.startswith(("answer:", "finished", "can't proceed")):
+        logging.info(f"Terminal command received ('{command_response}'). Resetting history for next turn.")
+        final_history = []
+    else:
+        final_history = action_step_history
+    
+    logging.info(f"action endpoint total processing time: {time.perf_counter()-start_time:.3f} seconds.")
+    logging.info(f"Response: {command_response}, History length for next turn: {len(final_history)}")
+
+    return ActionResponse(response=command_response, history=final_history)
+
+
+@app.post("/generate")
+@with_semaphore(timeout=60)
+async def generate(request: ActionRequest, token: str = Depends(verify_token)):
+    """
+    Processes the input image (in base64 format) and a user prompt:
+    1. Decodes and saves the original image.
+    2. Runs the wrapper to generate an image description and the YOLO updated image file.
+    3. Reads the YOLO updated image file.
+    4. Preprocesses the YOLO-updated image.
+    5. Constructs a prompt for GPT (using description + preprocessed YOLO image) and sends it.
+    6. Returns the command response.
+    """
+    start_time = time.perf_counter()
+    logging.info("generate endpoint start")
+    log_request_data(request, "/generate")
+
+    with tempfile.TemporaryDirectory() as temp_dir:
+        request_id = str(uuid.uuid4())
+        original_image_path = os.path.join(temp_dir, f"{request_id}.png")
+        yolo_updated_image_path = os.path.join(temp_dir, f"{request_id}_yolo_updated.png")
+
+        save_base64_image(request.image, original_image_path)
+        
+        try:
+            image_description = run_wrapper(image_path=original_image_path, output_dir=temp_dir)
+        except Exception as e:
+            logging.exception("Image processing failed in generate endpoint")
+            raise HTTPException(status_code=500, detail=f"Image processing failed: {e}")
+
+        try:
+            if not os.path.exists(yolo_updated_image_path):
+                raise HTTPException(status_code=500, detail="YOLO updated image generation failed or not found")
+            with open(yolo_updated_image_path, "rb") as f:
+                yolo_updated_img_bytes = f.read()
+        except Exception as e:
+            raise HTTPException(status_code=500, detail=f"Failed to read YOLO updated image: {e}")
+        
+        try:
+            _, new_b64 = preprocess_image(yolo_updated_img_bytes, threshold=1500, scale=0.5, fmt="png")
+        except Exception as e:
+            raise HTTPException(status_code=500, detail=f"Image preprocessing failed: {e}")
+
+    base64_image_url = f"data:image/png;base64,{new_b64}"
+
+    messages = [
+        {"role": "user", "content": [
+            {"type": "text", "text": f'"Prompt: {request.prompt}"\nImage description:\n"{image_description}"'},
+            {"type": "image_url", "image_url": {"url": base64_image_url, "detail": "high"}}
+        ]}
+    ]
+    
+    try:
+        response = LLM_CLIENT.chat.completions.create(model="gpt-4.1", messages=messages, temperature=0.2)
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=f"OpenAI API error: {e}")
+
+    command_response = response.choices[0].message.content.strip()
+    logging.info(f"generate endpoint total processing time: {time.perf_counter()-start_time:.3f} seconds")
+    return {"response": command_response}
+
+@app.post("/analyze")
+@with_semaphore(timeout=60)
+async def analyze(request: AnalyzeRequest, token: str = Depends(verify_token)):
+    """
+    Processes the input image (in base64 format) to return the image description as a JSON object.
+    """
+    logging.info("analyze endpoint start")
+    log_request_data(request, "/analyze")
+    
+    with tempfile.TemporaryDirectory() as temp_dir:
+        image_path = os.path.join(temp_dir, "image_to_analyze.png")
+        save_base64_image(request.image, image_path)
+
+        try:
+            image_description = run_wrapper(image_path=image_path, output_dir=temp_dir)
+            analyzed_description = json.loads(image_description)
+        except Exception as e:
+            logging.exception("Image processing failed in analyze endpoint.")
+            raise HTTPException(status_code=500, detail=f"Image processing failed: {e}")
+
+    return JSONResponse(content={"description": analyzed_description})
+
+
+@app.post("/analyze_and_get_yolo")
+@with_semaphore(timeout=60)
+async def analyze_and_get_yolo(request: AnalyzeRequest, token: str = Depends(verify_token)):
+    """
+    Processes the input image (in base64 format) to:
+      1. Return the image description as a JSON object.
+      2. Return the YOLO-updated image (base64 encoded). 
+    """
+    logging.info("analyze_and_get_yolo endpoint start")
+    log_request_data(request, "/analyze_and_get_yolo")
+    
+    with tempfile.TemporaryDirectory() as temp_dir:
+        request_id = str(uuid.uuid4())
+        image_path = os.path.join(temp_dir, f"{request_id}.png")
+        yolo_image_path = os.path.join(temp_dir, f"{request_id}_yolo_updated.png")
+
+        save_base64_image(request.image, image_path)
+        
+        try:
+            image_description = run_wrapper(image_path=image_path, output_dir=temp_dir)
+            analyzed_description = json.loads(image_description)
+        except Exception as e:
+            logging.exception("Image processing failed in analyze_and_get_yolo endpoint.")
+            raise HTTPException(status_code=500, detail=f"Image processing failed: {e}")
+        
+        if not os.path.exists(yolo_image_path):
+            logging.error("YOLO updated image not found.")
+            raise HTTPException(status_code=500, detail="YOLO updated image not generated.")
+        
+        try:
+            with open(yolo_image_path, "rb") as f:
+                yolo_img_bytes = f.read()
+            yolo_b64 = base64.b64encode(yolo_img_bytes).decode("utf-8")
+            yolo_image_encoded = f"data:image/png;base64,{yolo_b64}"
+        except Exception as e:
+            logging.exception("Error reading or encoding YOLO updated image.")
+            raise HTTPException(status_code=500, detail=f"Error handling YOLO updated image: {e}")
+
+    return JSONResponse(content={
+        "description": analyzed_description,
+        "yolo_image": yolo_image_encoded
+    })
+

requirements.txt

@@ -0,0 +1,15 @@
+easyocr==1.7.2
+numpy==1.26.4
+openai==1.55.3
+pillow==11.0.0
+pydantic==2.9.2
+pyspellchecker==0.8.1
+requests==2.32.3
+tensorflow==2.18.0
+torch==2.5.1
+transformers==4.50.3
+ultralytics==8.3.29
+webcolors==24.11.1
+httpx==0.27.2
+gradio==5.23.1
+opencv-python-headless==4.10.0.84

script.py

@@ -0,0 +1,808 @@
+import cv2
+import os
+import subprocess
+from PIL import Image
+import easyocr
+from spellchecker import SpellChecker
+import numpy as np
+import webcolors
+from collections import Counter
+import torch
+from transformers import AutoProcessor, Blip2ForConditionalGeneration
+import tensorflow as tf
+import argparse
+import json
+from concurrent.futures import ThreadPoolExecutor, as_completed
+import time
+from utils.json_helpers import NoIndent, CustomEncoder
+
+
+# constants
+BARRIER = "********\n"
+
+# Check if a model is in the cache
+def is_model_downloaded(model_name, cache_directory):
+    model_path = os.path.join(cache_directory, model_name.replace('/', '_'))
+    return os.path.exists(model_path)
+
+# Convert color to the closest name
+def closest_colour(requested_colour):
+    min_colours = {}
+    css3_names = webcolors.names("css3")
+    for name in css3_names:
+        hex_value = webcolors.name_to_hex(name, spec='css3')
+        r_c, g_c, b_c = webcolors.hex_to_rgb(hex_value)
+        rd = (r_c - requested_colour[0]) ** 2
+        gd = (g_c - requested_colour[1]) ** 2
+        bd = (b_c - requested_colour[2]) ** 2
+        distance = rd + gd + bd
+        min_colours[distance] = name
+    return min_colours[min(min_colours.keys())]
+
+def get_colour_name(requested_colour):
+    """
+    Returns a tuple: (exact_name, closest_name).
+    If an exact match fails, 'exact_name' is None, use the 'closest_name' fallback.
+    """
+    try:
+        actual_name = webcolors.rgb_to_name(requested_colour, spec='css3')
+        closest_name = actual_name
+    except ValueError:
+        closest_name = closest_colour(requested_colour)
+        actual_name = None
+    return actual_name, closest_name
+
+def get_most_frequent_color(pixels, bin_size=10):
+    """
+    Returns the most frequent color among the given pixels,
+    using a binning approach (default bin size=10).
+    """
+    bins = np.arange(0, 257, bin_size)
+    r_bins = np.digitize(pixels[:, 0], bins) - 1
+    g_bins = np.digitize(pixels[:, 1], bins) - 1
+    b_bins = np.digitize(pixels[:, 2], bins) - 1
+    combined_bins = r_bins * 10000 + g_bins * 100 + b_bins
+    bin_counts = Counter(combined_bins)
+    most_common_bin = bin_counts.most_common(1)[0][0]
+
+    r_bin = most_common_bin // 10000
+    g_bin = (most_common_bin % 10000) // 100
+    b_bin = most_common_bin % 100
+    r_value = bins[r_bin] + bin_size // 2
+    g_value = bins[g_bin] + bin_size // 2
+    b_value = bins[b_bin] + bin_size // 2
+
+    return (r_value, g_value, b_value)
+
+def get_most_frequent_alpha(alphas, bin_size=10):
+    bins = np.arange(0, 257, bin_size)
+    alpha_bins = np.digitize(alphas, bins) - 1
+    bin_counts = Counter(alpha_bins)
+    most_common_bin = bin_counts.most_common(1)[0][0]
+    alpha_value = bins[most_common_bin] + bin_size // 2
+    return alpha_value
+
+# downscale images for OCR. TODO change dim to a suitable one
+def downscale_for_ocr(image_cv, max_dim=600):
+    """
+    If either dimension of `image_cv` is bigger than `max_dim`,
+    scale it down proportionally. This speeds up EasyOCR on large images.
+    """
+    h, w = image_cv.shape[:2]
+    if w <= max_dim and h <= max_dim:
+        return image_cv  # No downscale needed
+
+    scale = min(max_dim / float(w), max_dim / float(h))
+    new_w = int(w * scale)
+    new_h = int(h * scale)
+    image_resized = cv2.resize(image_cv, (new_w, new_h), interpolation=cv2.INTER_AREA)
+    return image_resized
+
+# Worker function to process a single bounding box
+def process_single_region(
+    idx, bounding_box, image, sr, reader, spell, icon_model,
+    processor, model, device, no_captioning, output_json, json_mini,
+    cropped_imageview_images_dir, base_name, save_images,
+    model_to_use, log_prefix="",
+    skip_ocr=False,
+    skip_spell=False
+):
+    """
+    Processes one bounding box (region)
+    Returns a dict with:
+      * "region_dict" (for JSON)
+      * "text_log" (file/captions output)
+    """
+    (x_min, y_min, x_max, y_max, class_id) = bounding_box
+    class_names = {0: 'View', 1: 'ImageView', 2: 'Text', 3: 'Line'}
+    class_name = class_names.get(class_id, f'Unknown Class {class_id}')
+    region_idx = idx + 1
+    logs = []
+
+    x_center = (x_min + x_max) // 2
+    y_center = (y_min + y_max) // 2
+    width = x_max - x_min
+    height = y_max - y_min
+
+    def open_and_upscale_image(img_path, cid):
+
+        if cid == 2: # Text 
+            MAX_WIDTH, MAX_HEIGHT = 30, 30
+        else:
+            MAX_WIDTH, MAX_HEIGHT = 10, 10
+
+        def is_small(w, h):
+            return w <= MAX_WIDTH and h <= MAX_HEIGHT
+
+        if cid == 0:  # "View" - use PIL to preserve alpha
+            pil_image = Image.open(img_path).convert("RGBA")
+            w, h = pil_image.size
+            if not is_small(w, h):
+                logs.append(f"{log_prefix}Skipping upscale for large View (size={w}×{h}).")
+                return pil_image
+
+            # If super-resolution is provided, use it
+            if sr:
+                image_cv = cv2.cvtColor(np.array(pil_image), cv2.COLOR_RGBA2BGR)
+                upscaled = sr.upsample(image_cv)
+                return Image.fromarray(cv2.cvtColor(upscaled, cv2.COLOR_BGR2RGBA))
+            else:
+                return pil_image.resize((w * 4, h * 4), resample=Image.BICUBIC)
+        else:
+            # For other classes, load the image with OpenCV (BGR)
+            cv_image = cv2.imread(img_path)
+            if cv_image is None or cv_image.size == 0:
+                logs.append(f"{log_prefix}Empty image at {img_path}, skipping.")
+                return None
+
+            h, w = cv_image.shape[:2]
+            if not is_small(w, h):
+                logs.append(f"{log_prefix}Skipping upscale for large region (size={w}×{h}).")
+                return cv_image
+
+            if sr:
+                return sr.upsample(cv_image)
+            else:
+                return cv2.resize(cv_image, (w * 2, h * 2), interpolation=cv2.INTER_CUBIC)
+
+    if json_mini:
+        simplified_class_name = class_name.lower().replace('imageview', 'image')
+        new_id = f"{simplified_class_name}_{region_idx}"
+        mini_region_dict = {
+            "id": new_id,
+            "bbox": NoIndent([x_center, y_center, width, height])
+        }
+        
+        # only need to process text for the mini format
+        if class_name == 'Text' and not skip_ocr:
+            cropped_image_region = image[y_min:y_max, x_min:x_max]
+            if cropped_image_region.size > 0:
+                # Save the cropped image so open_and_upscale_image can use it
+                cropped_path = os.path.join(cropped_imageview_images_dir, f"region_{region_idx}_class_{class_id}.jpg")
+                cv2.imwrite(cropped_path, cropped_image_region)
+
+                upscaled = open_and_upscale_image(cropped_path, class_id)
+                if upscaled is not None:
+                    if isinstance(upscaled, Image.Image):
+                        upscaled_cv = cv2.cvtColor(np.array(upscaled), cv2.COLOR_RGBA2BGR)
+                    else:
+                        upscaled_cv = upscaled
+                    
+                    gray = cv2.cvtColor(downscale_for_ocr(upscaled_cv), cv2.COLOR_BGR2GRAY)
+                    text = ' '.join(reader.readtext(gray, detail=0, batch_size=8)).strip()
+
+                    if text:
+                        if not skip_spell and spell:
+                            corrected_words  = []
+                            for w in text.split():
+                                corrected_words.append(spell.correction(w) or w)
+                            mini_region_dict["text"] = " ".join(corrected_words)
+                        else:
+                             mini_region_dict["text"] = text
+                
+                # Clean up the temporary cropped image
+                if os.path.exists(cropped_path) and not save_images:
+                    os.remove(cropped_path)
+
+        return {"mini_region_dict": mini_region_dict, "text_log": ""}
+
+    logs.append(f"\n{log_prefix}Region {region_idx} - Class ID: {class_id} ({class_name})")
+    x_center = (x_min + x_max) // 2
+    y_center = (y_min + y_max) // 2
+    logs.append(f"{log_prefix}Coordinates: x_center={x_center}, y_center={y_center}")
+    width = x_max - x_min
+    height = y_max - y_min
+    logs.append(f"{log_prefix}Size: width={width}, height={height}")
+
+    region_dict = {
+        "id": f"region_{region_idx}_class_{class_name}",
+        "x_coordinates_center": x_center,
+        "y_coordinates_center": y_center,
+        "width": width,
+        "height": height
+    }
+
+    # Crop region
+    cropped_image_region = image[y_min:y_max, x_min:x_max]
+    if cropped_image_region.size == 0:
+        logs.append(f"{log_prefix}Empty crop for Region {region_idx}, skipping...")
+        return {"region_dict": region_dict, "text_log": "\n".join(logs)}
+
+    # Save cropped region
+    if class_id == 0:
+        # Save as PNG if it's a View
+        cropped_path = os.path.join(
+            cropped_imageview_images_dir, f"region_{region_idx}_class_{class_id}.png"
+        )
+        cv2.imwrite(cropped_path, cropped_image_region)
+    else:
+        # Save as JPG
+        cropped_path = os.path.join(
+            cropped_imageview_images_dir, f"region_{region_idx}_class_{class_id}.jpg"
+        )
+        cv2.imwrite(cropped_path, cropped_image_region)
+
+    # for LLaMA (ollama)
+    def call_ollama(prompt_text, rid, task_type):
+        model_name = "llama3.2-vision:11b"
+        cmd = ["ollama", "run", model_name, prompt_text]
+        try:
+            result = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+            if result.returncode != 0:
+                logs.append(f"{log_prefix}Error generating {task_type} for Region {rid}: {result.stderr}")
+                return None
+            else:
+                response = result.stdout.strip()
+                logs.append(f"{log_prefix}Generated {task_type.capitalize()} for Region {rid}: {response}")
+                return response
+        except Exception as e:
+            logs.append(f"{log_prefix}An error occurred while generating {task_type} for Region {rid}: {e}")
+            return None
+
+    # for BLIP-2
+    def generate_caption_blip(img_path):
+        pil_image = Image.open(img_path).convert('RGB')
+        inputs = processor(images=pil_image, return_tensors="pt").to(device, torch.float16)
+        gen_ids = model.generate(**inputs)
+        return processor.batch_decode(gen_ids, skip_special_tokens=True)[0].strip()
+
+    # Handle each class type
+    if class_id == 1:  # ImageView
+        if no_captioning:
+            logs.append(f"{log_prefix}(Icon-image detection + captioning disabled by --no-captioning.)")
+            if not output_json:
+                block = (
+                    f"Image: region_{region_idx}_class_{class_id} ({class_name})\n"
+                    f"Coordinates: x_center={(x_min + x_max) // 2}, y_center={(y_min + y_max) // 2}\n"
+                    f"Size: width={width}, height={height}\n"
+                    f"{BARRIER}"
+                )
+                logs.append(block)
+        else:
+            upscaled = open_and_upscale_image(cropped_path, class_id)
+            if upscaled is None:
+                return {"region_dict": region_dict, "text_log": "\n".join(logs)}
+
+            # Icon detection
+            if icon_model:
+                icon_input_size = (224, 224)
+                if isinstance(upscaled, Image.Image):
+                    upscaled_cv = cv2.cvtColor(np.array(upscaled), cv2.COLOR_RGBA2BGR)
+                else:
+                    upscaled_cv = upscaled
+                resized = cv2.resize(upscaled_cv, icon_input_size)
+                rgb_img = cv2.cvtColor(resized, cv2.COLOR_BGR2RGB) / 255.0
+                rgb_img = np.expand_dims(rgb_img, axis=0)
+                pred = icon_model.predict(rgb_img)
+                logs.append(f"{log_prefix}Prediction output for Region {region_idx}: {pred}")
+                if pred.shape == (1, 1):
+                    probability = pred[0][0]
+                    threshold = 0.5
+                    predicted_class = 1 if probability >= threshold else 0
+                    logs.append(f"{log_prefix}Probability of class 1: {probability}")
+                elif pred.shape == (1, 2):
+                    predicted_class = np.argmax(pred[0])
+                    logs.append(f"{log_prefix}Class probabilities: {pred[0]}")
+                else:
+                    logs.append(f"{log_prefix}Unexpected prediction shape: {pred.shape}")
+                    return {"region_dict": region_dict, "text_log": "\n".join(logs)}
+
+                pred_text = "Icon/Mobile UI Element" if predicted_class == 1 else "Normal Image"
+                region_dict["prediction"] = pred_text
+                if predicted_class == 1:
+                    prompt_text = "Describe the mobile UI element on this image. Keep it short."
+                else:
+                    prompt_text = "Describe what is in the image briefly. It's not an icon or typical UI element."
+            else:
+                logs.append(f"{log_prefix}Icon detection model not provided; skipping icon detection.")
+                region_dict["prediction"] = "Icon detection skipped"
+                prompt_text = "Describe what is in this image briefly."
+
+            # Caption
+            temp_image_path = os.path.abspath(
+                os.path.join(cropped_imageview_images_dir, f"imageview_{region_idx}.jpg")
+            )
+            if isinstance(upscaled, Image.Image): # TODO check optimization
+                upscaled.save(temp_image_path)
+            else:
+                cv2.imwrite(temp_image_path, upscaled)
+
+            response = ""
+            if model and processor and model_to_use == 'blip':
+                response = generate_caption_blip(temp_image_path)
+            else: # TODO check optimization
+                resp = call_ollama(prompt_text + " " + temp_image_path, region_idx, "description")
+                response = resp if resp else "Error generating description"
+
+            region_dict["description"] = response
+
+            if not output_json:
+                block = (
+                    f"Image: region_{region_idx}_class_{class_id} ({class_name})\n"
+                    f"Coordinates: x_center={(x_min + x_max) // 2}, y_center={(y_min + y_max) // 2}\n"
+                    f"Size: width={width}, height={height}\n"
+                    f"Prediction: {region_dict['prediction']}\n"
+                    f"{response}\n"
+                    f"{BARRIER}"
+                )
+                logs.append(block)
+
+            if os.path.exists(temp_image_path) and not save_images:
+                os.remove(temp_image_path)
+
+    elif class_id == 2:  # Text
+        if skip_ocr or reader is None:
+            logs.append(f"{log_prefix}OCR skipped for Region {region_idx}.")
+
+            if not output_json:
+                block = (
+                    f"Text: region_{region_idx}_class_{class_id} ({class_name})\n"
+                    f"Coordinates: x_center={(x_min + x_max) // 2}, "
+                    f"y_center={(y_min + y_max) // 2}\n"
+                    f"Size: width={width}, height={height}\n"
+                    f"OCR + spell-check disabled\n"
+                    f"{BARRIER}"
+                )
+                logs.append(block)
+
+            return {"region_dict": region_dict, "text_log": "\n".join(logs)}
+
+        upscaled = open_and_upscale_image(cropped_path, class_id)
+        if upscaled is None:
+            return {"region_dict": region_dict, "text_log": "\n".join(logs)}
+
+        if isinstance(upscaled, Image.Image):
+            upscaled_cv = cv2.cvtColor(np.array(upscaled), cv2.COLOR_RGBA2BGR)
+        else:
+            upscaled_cv = upscaled
+
+
+        # TODO use other lib to improve the performance
+        upscaled_cv = downscale_for_ocr(upscaled_cv, max_dim=600)
+        gray = cv2.cvtColor(upscaled_cv, cv2.COLOR_BGR2GRAY)
+        result_ocr = reader.readtext(gray, detail=0, batch_size=8)
+        text = ' '.join(result_ocr).strip()
+
+        # TODO use other lib to improve performance
+        if skip_spell or spell is None:
+            corrected_text = None
+            logs.append(f"{log_prefix}Spell-check skipped for Region {region_idx}.")
+        else:
+            correction_cache = {}
+            corrected_words  = []
+            for w in text.split():
+                if w not in correction_cache:
+                    correction_cache[w] = spell.correction(w) or w
+                corrected_words.append(correction_cache[w])
+            corrected_text = " ".join(corrected_words)
+
+
+        logs.append(f"{log_prefix}Extracted Text for Region {region_idx}: {text}")
+        if corrected_text is not None:
+            logs.append(f"{log_prefix}Corrected Text for Region {region_idx}: {corrected_text}")
+
+        region_dict["extractedText"] = text
+        if corrected_text is not None:
+            region_dict["correctedText"] = corrected_text
+
+        if not output_json:
+            block = (
+                f"Text: region_{region_idx}_class_{class_id} ({class_name})\n"
+                f"Coordinates: x_center={(x_min + x_max) // 2}, y_center={(y_min + y_max) // 2}\n"
+                f"Size: width={width}, height={height}\n"
+                f"Extracted Text: {text}\n"
+                + (f"Corrected Text: {corrected_text}\n" if corrected_text is not None else "")
+                + f"{BARRIER}"
+            )
+            logs.append(block)
+
+    elif class_id == 0:  # View
+        upscaled = open_and_upscale_image(cropped_path, class_id)
+        if upscaled is None:
+            return {"region_dict": region_dict, "text_log": "\n".join(logs)}
+
+        data = np.array(upscaled)
+        if data.ndim == 2:
+            data = cv2.cvtColor(data, cv2.COLOR_GRAY2BGRA)
+        elif data.shape[-1] == 3:
+            b, g, r = cv2.split(data)
+            a = np.full_like(b, 255)
+            data = cv2.merge((b, g, r, a))
+
+        pixels = data.reshape((-1, 4))
+        opaque_pixels = pixels[pixels[:, 3] > 0]
+
+        if len(opaque_pixels) == 0:
+            logs.append(f"{log_prefix}No opaque pixels found in Region {region_idx}, cannot determine background color.")
+            color_name = "Unknown"
+        else:
+            dom_color = get_most_frequent_color(opaque_pixels[:, :3], bin_size=10)
+            exact_name, closest_name = get_colour_name(dom_color)
+            color_name = exact_name if exact_name else closest_name
+
+        alphas = pixels[:, 3]
+        dominant_alpha = get_most_frequent_alpha(alphas, bin_size=10)
+        transparency = "opaque" if dominant_alpha >= 245 else "transparent"
+
+        response = (
+            f"1. The background color of the container is {color_name}.\n"
+            f"2. The container is {transparency}."
+        )
+        logs.append(f"{log_prefix}{response}")
+        region_dict["view_color"] = f"The background color of the container is {color_name}."
+        region_dict["view_alpha"] = f"The container is {transparency}."
+
+        if not output_json:
+            block = (
+                f"View: region_{region_idx}_class_{class_id} ({class_name})\n"
+                f"Coordinates: x_center={(x_min + x_max) // 2}, y_center={(y_min + y_max) // 2}\n"
+                f"Size: width={width}, height={height}\n"
+                f"{response}\n"
+                f"{BARRIER}"
+            )
+            logs.append(block)
+
+    elif class_id == 3:  # Line
+        logs.append(f"{log_prefix}Processing Line in Region {region_idx}")
+        line_img = cv2.imread(cropped_path, cv2.IMREAD_UNCHANGED)
+        if line_img is None:
+            logs.append(f"{log_prefix}Failed to read image at {cropped_path}")
+            return {"region_dict": region_dict, "text_log": "\n".join(logs)}
+
+        hh, ww = line_img.shape[:2]
+        logs.append(f"{log_prefix}Image dimensions: width={ww}, height={hh}")
+
+        data = np.array(line_img)
+        if data.ndim == 2:
+            data = cv2.cvtColor(data, cv2.COLOR_GRAY2BGRA)
+        elif data.shape[-1] == 3:
+            b, g, r = cv2.split(data)
+            a = np.full_like(b, 255)
+            data = cv2.merge((b, g, r, a))
+
+        pixels = data.reshape((-1, 4))
+        opaque_pixels = pixels[pixels[:, 3] > 0]
+
+        if len(opaque_pixels) == 0:
+            logs.append(f"{log_prefix}No opaque pixels found in Region {region_idx}, cannot determine line color.")
+            color_name = "Unknown"
+        else:
+            dom_color = get_most_frequent_color(opaque_pixels[:, :3], bin_size=10)
+            exact_name, closest_name = get_colour_name(dom_color)
+            color_name = exact_name if exact_name else closest_name
+
+        alphas = pixels[:, 3]
+        dom_alpha = get_most_frequent_alpha(alphas, bin_size=10)
+        transparency = "opaque" if dom_alpha >= 245 else "transparent"
+
+        response = (
+            f"1. The color of the line is {color_name}.\n"
+            f"2. The line is {transparency}."
+        )
+        logs.append(f"{log_prefix}{response}")
+        region_dict["line_color"] = f"The color of the line is {color_name}."
+        region_dict["line_alpha"] = f"The line is {transparency}."
+
+        if not output_json:
+            block = (
+                f"Line: region_{region_idx}_class_{class_id} ({class_name})\n"
+                f"Coordinates: x_center={(x_min + x_max) // 2}, y_center={(y_min + y_max) // 2}\n"
+                f"Size: width={width}, height={height}\n"
+                f"{response}\n"
+                f"{BARRIER}"
+            )
+            logs.append(block)
+
+    else:
+        logs.append(f"{log_prefix}Class ID {class_id} not handled.")
+
+    # Remove intermediate if not saving
+    if os.path.exists(cropped_path) and not save_images:
+        os.remove(cropped_path)
+
+    return {
+        "region_dict": region_dict,
+        "text_log": "\n".join(logs),
+    }
+
+
+# Main function
+def process_image(
+    input_image_path,
+    yolo_output_path,
+    output_dir:str = '.',
+    model_to_use='llama',     
+    save_images=False,
+    icon_model_path=None,
+    cache_directory='./models_cache',
+    huggingface_token='your_token', # for blip2
+    no_captioning=False,
+    output_json=False,
+    json_mini=False,
+    sr=None,
+    reader=None,
+    spell=None,
+    skip_ocr=False,
+    skip_spell=False
+):
+    if json_mini:
+        json_output = {
+            "image_size": None, # Will be populated later
+            "bbox_format": "center_x, center_y, width, height",
+            "elements": []
+        }
+    elif output_json:
+        json_output = {
+            "image": {"path": input_image_path, "size": {"width": None, "height": None}},
+            "elements": []
+        }
+    else:
+        json_output = None
+
+
+    start_time = time.perf_counter()
+    print("super-resolution initialization start (in script.py)")
+    # Super-resolution initialization
+    if sr is None:
+        print("No sr reference passed; performing local init ...")
+        model_path = 'EDSR_x4.pb'
+        if hasattr(cv2, 'dnn_superres'):
+            print("dnn_superres module is available.")
+            import cv2.dnn_superres as dnn_superres
+            try:
+                sr = cv2.dnn_superres.DnnSuperResImpl_create()
+                print("Using DnnSuperResImpl_create()")
+            except AttributeError:
+                sr = cv2.dnn_superres.DnnSuperResImpl()
+                print("Using DnnSuperResImpl()")
+            sr.readModel(model_path)
+            sr.setModel('edsr', 4)
+        else:
+            print("dnn_superres module is NOT available; skipping super-resolution.")
+    else:
+        print("Using pre-initialized sr reference.")
+
+
+    elapsed = time.perf_counter() - start_time
+    print(f"super-resoulution init (in script.py) took {elapsed:.3f} seconds.")
+
+    start_time = time.perf_counter()
+
+    if skip_ocr:
+        print("skip_ocr flag set - skipping EasyOCR and SpellChecker.")
+        reader = None
+        spell  = None
+
+    else:
+        print("OCR initialisation start (in script.py)")
+        if reader is None:
+            print("No EasyOCR reference passed; performing local init")
+            reader = easyocr.Reader(['en'], gpu=True)
+        else:
+            print("Using pre-initialised EasyOCR object.")
+
+        if skip_spell:
+            print("skip_spell flag set - not initialising SpellChecker.")
+            spell = None
+        else:
+            if spell is None:
+                print("No SpellChecker reference passed; performing local init")
+                spell = SpellChecker()
+            else:
+                print("Using pre-initialised SpellChecker object.")
+
+    elapsed = time.perf_counter() - start_time
+    print(f"OCR init (in script.py) took {elapsed:.3f} seconds.")
+
+
+    start_time = time.perf_counter()
+    print("icon-model init start (in script.py)")
+    # Load icon detection model (if provided)
+    if icon_model_path:
+        icon_model = tf.keras.models.load_model(icon_model_path)
+        print(f"Icon detection model loaded: {icon_model_path}")
+    else:
+        icon_model = None
+
+    elapsed = time.perf_counter() - start_time
+    print(f"icon-model init (in script.py) took {elapsed:.3f} seconds.")
+
+    
+    # Load the original image
+    image = cv2.imread(input_image_path, cv2.IMREAD_UNCHANGED)
+    if image is None:
+        print(f"Image at {input_image_path} could not be loaded.")
+        return
+
+    image_height, image_width = image.shape[:2]
+
+
+    # Read YOLO labels
+    with open(yolo_output_path, 'r') as f:
+        lines = f.readlines()
+
+    # Check torch device
+    if torch.backends.mps.is_available():
+        device = torch.device("mps")
+        print("Using MPS")
+    elif torch.cuda.is_available():
+        device = torch.device("cuda")
+        print("Using CUDA")
+    else:
+        device = torch.device("cpu")
+        print("Using CPU")
+
+    # Conditionally load the captioning model
+    processor, model = None, None
+    if not no_captioning:
+        if model_to_use == 'blip':
+            print("Loading BLIP-2 model...")
+            blip_model_name = "Salesforce/blip2-opt-2.7b"
+            if not is_model_downloaded(blip_model_name, cache_directory):
+                print("Model not found in cache. Downloading...")
+            else:
+                print("Model found in cache. Loading...")
+            processor = AutoProcessor.from_pretrained(
+                blip_model_name,
+                use_auth_token=huggingface_token,
+                cache_dir=cache_directory,
+                resume_download=True
+            )
+            model = Blip2ForConditionalGeneration.from_pretrained(
+                blip_model_name,
+                device_map='auto',
+                torch_dtype=torch.float16,
+                use_auth_token=huggingface_token,
+                cache_dir=cache_directory,
+                resume_download=True
+            ).to(device)
+        else:
+            print("Using LLaMA model via external call (ollama).")
+    else:
+        print("--no-captioning flag is set; skipping model loading.")
+
+    # Prepare bounding boxes from YOLO
+    bounding_boxes = []
+    for line in lines:
+        parts = line.strip().split()
+        class_id = int(parts[0])
+        x_center_norm, y_center_norm, width_norm, height_norm = map(float, parts[1:])
+        x_center = x_center_norm * image_width
+        y_center = y_center_norm * image_height
+        box_width = width_norm * image_width
+        box_height = height_norm * image_height
+        x_min = int(x_center - box_width / 2)
+        y_min = int(y_center - box_height / 2)
+        x_max = int(x_center + box_width / 2)
+        y_max = int(y_center + box_height / 2)
+        x_min = max(0, x_min)
+        y_min = max(0, y_min)
+        x_max = min(image_width - 1, x_max)
+        y_max = min(image_height - 1, y_max)
+        bounding_boxes.append((x_min, y_min, x_max, y_max, class_id))
+
+
+    # Create output dirs
+    cropped_dir = os.path.join(output_dir, "cropped_imageview_images")
+    os.makedirs(cropped_dir, exist_ok=True)
+    result_dir = os.path.join(output_dir, "result")
+    os.makedirs(result_dir, exist_ok=True)
+
+    base_name = os.path.splitext(os.path.basename(input_image_path))[0]
+    captions_file_path = None
+    if json_mini:
+        json_output["image_size"] = NoIndent([image_width, image_height])
+    elif output_json:
+        json_output["image"]["size"]["width"] = image_width
+        json_output["image"]["size"]["height"] = image_height
+    else: # Text output
+        captions_filename = f"{base_name}_regions_captions.txt"
+        captions_file_path = os.path.join(result_dir, captions_filename)
+        with open(captions_file_path, 'w', encoding='utf-8') as f:
+            f.write(f"Image path: {input_image_path}\n")
+            f.write(f"Image Size: width={image_width}, height={image_height}\n")
+            f.write(BARRIER)
+
+    # Number of workers can be increased if hardware is suitable for it. But testing is needed
+    start_time = time.perf_counter()
+    print("Process single region start (in script.py)")
+
+    with ThreadPoolExecutor(max_workers=1) as executor:
+        futures = [
+            executor.submit(
+                process_single_region,
+                idx, box, image, sr, reader, spell,
+                icon_model, processor, model, (model and device),
+                no_captioning, output_json, json_mini,
+                cropped_dir, base_name, save_images,
+                model_to_use, log_prefix="",
+                skip_ocr=skip_ocr,
+                skip_spell=skip_spell
+            ) for idx, box in enumerate(bounding_boxes)
+        ]
+
+        for future in as_completed(futures):
+            item = future.result()
+            if json_mini:
+                if item.get("mini_region_dict"):
+                    json_output["elements"].append(item["mini_region_dict"])
+            elif output_json:
+                if item.get("region_dict"):
+                    json_output["elements"].append(item["region_dict"])
+            else: # Text output
+                if item.get("text_log") and captions_file_path:
+                    with open(captions_file_path, 'a', encoding='utf-8') as f:
+                        f.write(item["text_log"])
+
+    elapsed = time.perf_counter() - start_time
+    print(f"Processing regions took {elapsed:.3f} seconds.")
+
+    if json_mini or output_json:
+        json_file = os.path.join(result_dir, f"{base_name}.json")
+        with open(json_file, 'w', encoding='utf-8') as f:
+            json.dump(json_output, f, indent=2, ensure_ascii=False, cls=CustomEncoder)
+        
+        output_type = "mini JSON" if json_mini else "JSON"
+        print(f"{output_type} output written to {json_file}")
+    else:
+        print(f"Text output written to {captions_file_path}")
+    
+
+# CLI entry point
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='Process an image and its YOLO labels.')
+    parser.add_argument('input_image', help='Path to the input YOLO image.')
+    parser.add_argument('input_labels', help='Path to the input YOLO labels file.')
+    parser.add_argument('--output_dir', default='.', 
+                        help='Directory to save output files. Defaults to the current directory.')
+    parser.add_argument('--model_to_use', choices=['llama', 'blip'], default='llama',
+                        help='Model for captioning (llama or blip).')
+    parser.add_argument('--save_images', action='store_true',
+                        help='Flag to save intermediate images.')
+    parser.add_argument('--icon_detection_path', help='Path to icon detection model.')
+    parser.add_argument('--cache_directory', default='./models_cache',
+                        help='Cache directory for Hugging Face models.')
+    parser.add_argument('--huggingface_token', default='your_token',
+                        help='Hugging Face token for model downloads.')
+    parser.add_argument('--no-captioning', action='store_true',
+                        help='Disable any image captioning.')
+    parser.add_argument('--json', dest='output_json', action='store_true',
+                        help='Output the image data in JSON format')
+    parser.add_argument('--json-mini', action='store_true',
+                        help='Output the image data in a condensed JSON format')
+    args = parser.parse_args()
+
+    process_image(
+        input_image_path=args.input_image,
+        yolo_output_path=args.input_labels,
+        output_dir=args.output_dir,
+        model_to_use=args.model_to_use,
+        save_images=args.save_images,
+        icon_model_path=args.icon_detection_path,
+        cache_directory=args.cache_directory,
+        huggingface_token=args.huggingface_token,
+        no_captioning=args.no_captioning,
+        output_json=args.output_json,
+        json_mini=args.json_mini
+    )
+

utils/json_helpers.py

@@ -0,0 +1,18 @@
+import json
+
+class NoIndent:
+    """ Wrapper class to mark lists that should not be indented """
+    def __init__(self, value):
+        self.value = value
+
+class CustomEncoder(json.JSONEncoder):
+    """
+    Custom JSON encoder that handles the NoIndent class to produce
+    a compact string representation of the list
+    """
+    def default(self, obj):
+        if isinstance(obj, NoIndent):
+            # Return the value formatted as a compact string, without newlines
+            return json.dumps(obj.value, separators=(',',':'))
+        return super().default(obj)
+

utils/pills.py

@@ -0,0 +1,44 @@
+import cv2
+import base64
+import numpy as np
+from typing import Tuple
+
+def preprocess_image(image_bytes: bytes, threshold: int = 1000, scale: float = 0.5, fmt: str = "png") -> Tuple[bytes, str]:
+    """
+    Preprocesses the image by checking its dimensions and downscaling it if needed.
+    
+    Parameters:
+      - image_bytes: Raw image bytes.
+      - threshold: Maximum allowed width or height (in pixels). If either dimension exceeds this,
+                   the image will be downscaled.
+      - scale: Scale factor to use for resizing if the image is too large.
+      - fmt: Format for re-encoding the image (e.g., "png" or "jpg").
+    
+    Returns:
+      - A tuple (new_image_bytes, new_base64_str) where:
+          new_image_bytes: The re-encoded image bytes after potential downscaling.
+          new_base64_str: The base64 string (without header) of the new image bytes.
+    """
+    # Convert raw bytes to a NumPy array then decode with OpenCV
+    nparr = np.frombuffer(image_bytes, np.uint8)
+    cv_image = cv2.imdecode(nparr, cv2.IMREAD_UNCHANGED)
+    
+    if cv_image is None:
+        raise ValueError("Failed to decode image with OpenCV.")
+    
+    h, w = cv_image.shape[:2]
+    
+    # If either dimension is greater than threshold, resize the image.
+    if h > threshold or w > threshold:
+        new_w, new_h = int(w * scale), int(h * scale)
+        cv_image = cv2.resize(cv_image, (new_w, new_h), interpolation=cv2.INTER_AREA)
+    
+    ret, buf = cv2.imencode(f'.{fmt}', cv_image)
+    
+    if not ret:
+        raise ValueError("Failed to re-encode image.")
+    
+    new_image_bytes = buf.tobytes()
+    new_base64_str = base64.b64encode(new_image_bytes).decode("utf-8")
+    
+    return new_image_bytes, new_base64_str

wrapper.py

@@ -0,0 +1,140 @@
+import os
+import sys
+import argparse
+from ultralytics import YOLO
+from os.path import basename, splitext
+import time
+
+from yolo_script import process_yolo
+
+from script import process_image
+
+def process_image_description(
+    input_image: str,
+    weights_file: str,
+    output_dir: str,
+    model_to_use: str = 'llama',
+    save_images: bool = False,
+    icon_detection_path: str = None,
+    cache_directory: str = './models_cache',
+    huggingface_token: str = 'your_token',
+    no_captioning: bool = False,
+    output_json: bool = False,
+    json_mini: bool = False,
+    model_obj: YOLO = None,
+    sr=None,
+    reader=None,
+    spell=None,
+    skip_ocr=False,
+    skip_spell=False,
+) -> None:
+    """
+    Processes an image by running YOLO detection (via the imported process_yolo function)
+    and then calling process_image() from script.py to do the image description work.
+    
+    Parameters:
+      - input_image: Path to the input image.
+      - weights_file: Path to the YOLO weights file.
+      - output_dir: Directory for YOLO output
+      - model_to_use: Which model to use for captioning ('llama' or 'blip').
+      - save_images: Whether to save intermediate images.
+      - icon_detection_path: Optional path to an icon detection model.
+      - cache_directory: Cache directory for models.
+      - huggingface_token: Hugging Face token for model downloads.
+      - no_captioning: If True, disable image captioning.
+      - output_json: If True, output the results in JSON format.
+      - json_mini: same as output_json but has more compact json output.
+      - model_obj: YOLO object that was initialized at a startup time (optional)
+      - sr: Super resolution object (optional)
+      - reader: EasyOCR object (optional)
+      - spell: Spell checker object (optional)
+    """
+
+    base_name = splitext(basename(input_image))[0]
+    
+    process_yolo(input_image, weights_file, output_dir, model_obj=model_obj)
+    
+    labels_dir = os.path.join(output_dir, 'labels')
+    label_file = os.path.join(labels_dir, base_name + '.txt')
+
+    if not os.path.isfile(label_file):
+        raise FileNotFoundError(f"Labels file not found at expected path: {label_file}")
+
+    process_image(
+        input_image_path=input_image,
+        yolo_output_path=label_file,
+        output_dir=output_dir,
+        model_to_use=model_to_use,
+        save_images=save_images,
+        icon_model_path=icon_detection_path,
+        cache_directory=cache_directory,
+        huggingface_token=huggingface_token,
+        no_captioning=no_captioning,
+        output_json=output_json,
+        json_mini=json_mini,
+        sr=sr,
+        reader=reader,
+        spell=spell,
+        skip_ocr=skip_ocr,
+        skip_spell=skip_spell,
+    )
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(
+        description='Wrapper script to run YOLO detection and image description in sequence.'
+    )
+    parser.add_argument('--input_image', required=True, help='Path to the input image.')
+    parser.add_argument('--weights_file', required=True, help='Path to the YOLO weights file.')
+    parser.add_argument('--output_dir', default='./output', help='Output directory for YOLO results.')
+    parser.add_argument('--model_to_use', choices=['llama', 'blip'], default='llama',
+                        help='Model for captioning.')
+    parser.add_argument('--save_images', action='store_true',
+                        help='Flag to save intermediate images.')
+    parser.add_argument('--icon_detection_path', help='Path to the icon detection model.')
+    parser.add_argument('--cache_directory', default='./models_cache',
+                        help='Cache directory for models.')
+    parser.add_argument('--huggingface_token', default='your_token',
+                        help='Hugging Face token for model downloads.')
+    parser.add_argument('--no-captioning', action='store_true',
+                        help='Disable any image captioning')
+    parser.add_argument('--json', dest='output_json', action='store_true',
+                        help='Output the image data in JSON format')
+    parser.add_argument('--json-mini', action='store_true',
+                        help='JSON output in a more condensed format')
+    parser.add_argument('--skip-ocr', action='store_true',
+                        help='Disable OCR & spell-checking (faster).')
+    parser.add_argument('--skip-spell', action='store_true', help='Run OCR but skip spell-check')
+    
+    args = parser.parse_args()
+    
+    try:
+        print("Running YOLO detection...")
+        yolo_output_dir = args.output_dir
+        os.makedirs(yolo_output_dir, exist_ok=True)
+        process_yolo(args.input_image, args.weights_file, yolo_output_dir)
+
+        base_name = splitext(basename(args.input_image))[0]
+        labels_dir = os.path.join(yolo_output_dir, 'labels')
+        label_file = os.path.join(labels_dir, base_name + '.txt')
+        if not os.path.isfile(label_file):
+            raise FileNotFoundError(f"Labels file not found: {label_file}")
+
+        print("Running image description...")
+        process_image(
+            input_image_path=args.input_image,
+            yolo_output_path=label_file,
+            model_to_use=args.model_to_use,
+            save_images=args.save_images,
+            icon_model_path=args.icon_detection_path,
+            cache_directory=args.cache_directory,
+            huggingface_token=args.huggingface_token,
+            no_captioning=args.no_captioning,
+            output_json=args.output_json,
+            json_mini=args.json_mini,
+            skip_ocr=args.skip_ocr,
+            skip_spell=args.skip_spell
+        )
+    except Exception as e:
+        print(e)
+        sys.exit(1)
+

yolo_script.py

@@ -0,0 +1,185 @@
+import ultralytics
+import cv2
+from ultralytics import YOLO
+import os
+import glob
+import argparse
+import sys
+import numpy as np
+import uuid
+
+def iou(box1, box2):
+    # Convert normalized coordinates to (x1, y1, x2, y2)
+    x1_1 = box1[0] - box1[2] / 2
+    y1_1 = box1[1] - box1[3] / 2
+    x2_1 = box1[0] + box1[2] / 2
+    y2_1 = box1[1] + box1[3] / 2
+
+    x1_2 = box2[0] - box2[2] / 2
+    y1_2 = box2[1] - box2[3] / 2
+    x2_2 = box2[0] + box2[2] / 2
+    y2_2 = box2[1] + box2[3] / 2
+
+    xi1 = max(x1_1, x1_2)
+    yi1 = max(y1_1, y1_2)
+    xi2 = min(x2_1, x2_2)
+    yi2 = min(y2_1, y2_2)
+    inter_area = max(0, xi2 - xi1) * max(0, yi2 - yi1)
+
+    box1_area = (x2_1 - x1_1) * (y2_1 - y1_1)
+    box2_area = (x2_2 - x1_2) * (y2_2 - y1_2)
+    union_area = box1_area + box2_area - inter_area
+
+    return inter_area / union_area if union_area > 0 else 0
+
+def load_image(input_image, base_name: str = None):
+
+    if isinstance(input_image, str):
+        img = cv2.imread(input_image)
+        if img is None:
+            raise ValueError(f"Unable to load image from path: {input_image}")
+        if base_name is None:
+            base_name = os.path.splitext(os.path.basename(input_image))[0]
+        return img, base_name
+    else:
+        # Assume input_image is raw bytes or a file-like object
+        if isinstance(input_image, bytes):
+            image_bytes = input_image
+        else:
+            image_bytes = input_image.read()
+        nparr = np.frombuffer(image_bytes, np.uint8)
+        img = cv2.imdecode(nparr, cv2.IMREAD_COLOR)
+        if img is None:
+            raise ValueError("Unable to decode image from input bytes or file-like object.")
+        if base_name is None:
+            base_name = str(uuid.uuid4())
+        return img, base_name
+
+def process_yolo(input_image, weights_file: str, output_dir: str = './yolo_run', model_obj: YOLO = None, base_name: str = None) -> str:
+
+    orig_image, inferred_base_name = load_image(input_image, base_name)
+    base_name = inferred_base_name 
+    
+    os.makedirs(output_dir, exist_ok=True)
+    # Determine file extension: if input_image is a file path, use its extension. Otherwise, default to .png
+    if isinstance(input_image, str):
+        ext = os.path.splitext(input_image)[1]
+    else:
+        ext = ".png"
+    output_image_name = f"{base_name}_yolo{ext}"
+    updated_output_image_name = f"{base_name}_yolo_updated{ext}"
+
+    # If input_image is a file path, call YOLO with the path to preserve filename-based labeling.
+    # Otherwise, if processing in-memory, YOLO might default to a generic name.
+    if isinstance(input_image, str):
+        source_input = input_image
+    else:
+        source_input = orig_image
+
+    # Use provided model or load from weights_file.
+    if model_obj is None:
+        model = YOLO(weights_file)
+    else:
+        model = model_obj
+
+    results = model(
+        source=source_input,
+        save_txt=True,
+        project=output_dir,
+        name='.',
+        exist_ok=True,
+    )
+
+    # Save the initial inference image.
+    img_with_boxes = results[0].plot(font_size=2, line_width=1)
+    output_image_path = os.path.join(output_dir, output_image_name)
+    cv2.imwrite(output_image_path, img_with_boxes)
+    print(f"Image saved as '{output_image_path}'")
+
+    labels_dir = os.path.join(output_dir, 'labels')
+    label_file = os.path.join(labels_dir, f"{base_name}.txt")
+    
+    if not os.path.isfile(label_file):
+        raise FileNotFoundError(f"No label files found for the image '{base_name}' at path '{label_file}'.")
+
+    with open(label_file, 'r') as f:
+        lines = f.readlines()
+
+    boxes = []
+    for idx, line in enumerate(lines):
+        tokens = line.strip().split()
+        class_id = int(tokens[0])
+        x_center = float(tokens[1])
+        y_center = float(tokens[2])
+        width = float(tokens[3])
+        height = float(tokens[4])
+        boxes.append({
+            'class_id': class_id,
+            'bbox': [x_center, y_center, width, height],
+            'line': line,
+            'index': idx
+        })
+
+    boxes.sort(key=lambda b: b['bbox'][1] - (b['bbox'][3] / 2))
+
+    # Perform NMS.
+    keep_indices = []
+    suppressed = [False] * len(boxes)
+    num_removed = 0
+    for i in range(len(boxes)):
+        if suppressed[i]:
+            continue
+        keep_indices.append(i)
+        for j in range(i + 1, len(boxes)):
+            if suppressed[j]:
+                continue
+            if boxes[i]['class_id'] == boxes[j]['class_id']:
+                iou_value = iou(boxes[i]['bbox'], boxes[j]['bbox'])
+                if iou_value > 0.7:
+                    suppressed[j] = True
+                    num_removed += 1
+
+    with open(label_file, 'w') as f:
+        for idx in keep_indices:
+            f.write(boxes[idx]['line'])
+
+    print(f"Number of bounding boxes removed: {num_removed}")
+
+    # Draw updated bounding boxes on the original image (loaded in memory).
+    drawn_image = orig_image.copy()
+    h_img, w_img, _ = drawn_image.shape
+
+    for i, idx in enumerate(keep_indices):
+        box = boxes[idx]
+        x_center, y_center, w_norm, h_norm = box['bbox']
+        x_center *= w_img
+        y_center *= h_img
+        w_box = w_norm * w_img
+        h_box = h_norm * h_img
+        x1 = int(x_center - w_box / 2)
+        y1 = int(y_center - h_box / 2)
+        x2 = int(x_center + w_box / 2)
+        y2 = int(y_center + h_box / 2)
+        cv2.rectangle(drawn_image, (x1, y1), (x2, y2), (0, 255, 0), 2)
+        cv2.putText(drawn_image, str(i + 1), (x1, y1 - 5),
+                    cv2.FONT_HERSHEY_SIMPLEX, 0.8, (143, 10, 18), 1)
+
+    updated_output_image_path = os.path.join(output_dir, updated_output_image_name)
+    cv2.imwrite(updated_output_image_path, drawn_image)
+    print(f"Updated image saved as '{updated_output_image_path}'")
+    
+    return updated_output_image_path
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='Process YOLO inference and NMS on an image.')
+    parser.add_argument('input_image', help='Path to the input image or pass raw bytes via a file-like object.')
+    parser.add_argument('weights_file', help='Path to the YOLO weights file.')
+    parser.add_argument('output_dir', nargs='?', default='./yolo_run', help='Output directory (optional).')
+    parser.add_argument('--base_name', help='Optional base name for output files (without extension).')
+    args = parser.parse_args()
+    
+    try:
+        process_yolo(args.input_image, args.weights_file, args.output_dir, base_name=args.base_name)
+    except Exception as e:
+        print(e)
+        sys.exit(1)