Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -43,11 +43,6 @@ depth_model.eval()
|
|
43 |
# Define the Segmentation-Based Blur Effect
|
44 |
# -----------------------------
|
45 |
def segmentation_blur_effect(input_image: Image.Image):
|
46 |
-
"""
|
47 |
-
Creates a segmentation mask using RMBG-2.0 and applies a Gaussian blur (sigma=15)
|
48 |
-
to the background while keeping the foreground sharp.
|
49 |
-
"""
|
50 |
-
# Resize input image for segmentation processing
|
51 |
imageResized = input_image.resize(seg_image_size)
|
52 |
input_tensor = seg_transform(imageResized).unsqueeze(0).to(device)
|
53 |
|
@@ -55,54 +50,28 @@ def segmentation_blur_effect(input_image: Image.Image):
|
|
55 |
preds = seg_model(input_tensor)[-1].sigmoid().cpu()
|
56 |
pred = preds[0].squeeze()
|
57 |
|
58 |
-
# Convert predicted mask to a PIL image and ensure it matches imageResized's size
|
59 |
pred_pil = transforms.ToPILImage()(pred)
|
60 |
mask = pred_pil.resize(imageResized.size)
|
61 |
|
62 |
-
# Convert mask to grayscale and threshold to create a binary mask
|
63 |
mask_np = np.array(mask.convert("L"))
|
64 |
_, maskBinary = cv2.threshold(mask_np, 127, 255, cv2.THRESH_BINARY)
|
65 |
|
66 |
-
# Convert the resized image to an OpenCV BGR array
|
67 |
img = cv2.cvtColor(np.array(imageResized), cv2.COLOR_RGB2BGR)
|
68 |
-
# Apply Gaussian blur (sigmaX=15, sigmaY=15)
|
69 |
blurredBg = cv2.GaussianBlur(np.array(imageResized), (0, 0), sigmaX=15, sigmaY=15)
|
70 |
|
71 |
-
# Create the inverse mask and convert it to 3 channels
|
72 |
maskInv = cv2.bitwise_not(maskBinary)
|
73 |
maskInv3 = cv2.cvtColor(maskInv, cv2.COLOR_GRAY2BGR)
|
74 |
|
75 |
-
# Extract the foreground and background using the mask
|
76 |
foreground = cv2.bitwise_and(img, cv2.bitwise_not(maskInv3))
|
77 |
background = cv2.bitwise_and(blurredBg, maskInv3)
|
78 |
|
79 |
-
# Combine foreground and background; convert back to RGB for display
|
80 |
finalImg = cv2.add(cv2.cvtColor(foreground, cv2.COLOR_BGR2RGB), background)
|
81 |
finalImg_pil = Image.fromarray(finalImg)
|
82 |
|
83 |
return finalImg_pil, mask
|
84 |
|
85 |
-
# -----------------------------
|
86 |
-
# Define the Depth-Based Lens Blur Effect with Slider-Controlled Thresholds
|
87 |
-
# -----------------------------
|
88 |
def lens_blur_effect(input_image: Image.Image, fg_threshold: float = 85, mg_threshold: float = 170):
|
89 |
-
|
90 |
-
Uses DepthPro to estimate a depth map and applies a dynamic lens blur effect
|
91 |
-
by blending three versions of the image with increasing blur levels.
|
92 |
-
|
93 |
-
Parameters:
|
94 |
-
input_image: The original PIL image.
|
95 |
-
fg_threshold: Foreground threshold (0-255). Pixels with depth below this are considered foreground.
|
96 |
-
mg_threshold: Middleground threshold (0-255). Pixels with depth between fg_threshold and mg_threshold are middleground.
|
97 |
-
|
98 |
-
Returns:
|
99 |
-
depthImg: The computed depth map (PIL Image).
|
100 |
-
lensBlurImage: The final lens-blurred image (PIL Image).
|
101 |
-
mask_fg_img: Foreground depth mask.
|
102 |
-
mask_mg_img: Middleground depth mask.
|
103 |
-
mask_bg_img: Background depth mask.
|
104 |
-
"""
|
105 |
-
# Process the image with the depth estimation model
|
106 |
inputs = depth_processor(images=input_image, return_tensors="pt").to(device)
|
107 |
with torch.no_grad():
|
108 |
outputs = depth_model(**inputs)
|
@@ -111,39 +80,32 @@ def lens_blur_effect(input_image: Image.Image, fg_threshold: float = 85, mg_thre
|
|
111 |
)
|
112 |
depth = post_processed_output[0]["predicted_depth"]
|
113 |
|
114 |
-
# Normalize depth to [0, 255]
|
115 |
depth = (depth - depth.min()) / (depth.max() - depth.min())
|
116 |
depth = depth * 255.
|
117 |
depth = depth.detach().cpu().numpy()
|
118 |
depth_map = depth.astype(np.uint8)
|
119 |
depthImg = Image.fromarray(depth_map)
|
120 |
|
121 |
-
# Convert input image to OpenCV BGR format
|
122 |
img = cv2.cvtColor(np.array(input_image), cv2.COLOR_RGB2BGR)
|
123 |
|
124 |
-
# Precompute blurred versions for different depth regions
|
125 |
img_foreground = img.copy() # No blur for foreground
|
126 |
img_middleground = cv2.GaussianBlur(img, (0, 0), sigmaX=7, sigmaY=7)
|
127 |
img_background = cv2.GaussianBlur(img, (0, 0), sigmaX=15, sigmaY=15)
|
128 |
|
129 |
print(depth_map)
|
130 |
-
depth_map
|
131 |
|
132 |
-
|
133 |
-
|
134 |
-
threshold2 = mg_threshold # e.g., default 170
|
135 |
|
136 |
-
# Create masks for foreground, middleground, and background based on depth
|
137 |
mask_fg = (depth_map < threshold1).astype(np.float32)
|
138 |
mask_mg = ((depth_map >= threshold1) & (depth_map < threshold2)).astype(np.float32)
|
139 |
mask_bg = (depth_map >= threshold2).astype(np.float32)
|
140 |
|
141 |
-
# Expand masks to 3 channels
|
142 |
mask_fg_3 = np.stack([mask_fg]*3, axis=-1)
|
143 |
mask_mg_3 = np.stack([mask_mg]*3, axis=-1)
|
144 |
mask_bg_3 = np.stack([mask_bg]*3, axis=-1)
|
145 |
|
146 |
-
# Blend the images using the masks
|
147 |
final_img = (img_foreground * mask_fg_3 +
|
148 |
img_middleground * mask_mg_3 +
|
149 |
img_background * mask_bg_3).astype(np.uint8)
|
@@ -151,27 +113,14 @@ def lens_blur_effect(input_image: Image.Image, fg_threshold: float = 85, mg_thre
|
|
151 |
final_img_rgb = cv2.cvtColor(final_img, cv2.COLOR_BGR2RGB)
|
152 |
lensBlurImage = Image.fromarray(final_img_rgb)
|
153 |
|
154 |
-
# Create mask images for display (scaled to 0-255)
|
155 |
mask_fg_img = Image.fromarray((mask_fg * 255).astype(np.uint8))
|
156 |
mask_mg_img = Image.fromarray((mask_mg * 255).astype(np.uint8))
|
157 |
mask_bg_img = Image.fromarray((mask_bg * 255).astype(np.uint8))
|
158 |
|
159 |
return depthImg, lensBlurImage, mask_fg_img, mask_mg_img, mask_bg_img
|
160 |
|
161 |
-
# -----------------------------
|
162 |
-
# Gradio App: Process Image and Display Multiple Effects
|
163 |
-
# -----------------------------
|
164 |
def process_image(input_image: Image.Image, fg_threshold: float, mg_threshold: float):
|
165 |
-
|
166 |
-
Processes the uploaded image to generate:
|
167 |
-
1. Segmentation-based Gaussian blur effect.
|
168 |
-
2. Segmentation mask.
|
169 |
-
3. Depth map.
|
170 |
-
4. Depth-based lens blur effect.
|
171 |
-
5. Depth masks for foreground, middleground, and background.
|
172 |
-
|
173 |
-
The depth thresholds for foreground and middleground regions are adjustable via sliders.
|
174 |
-
"""
|
175 |
seg_blur, seg_mask = segmentation_blur_effect(input_image)
|
176 |
depth_map_img, lens_blur_img, mask_fg_img, mask_mg_img, mask_bg_img = lens_blur_effect(
|
177 |
input_image, fg_threshold, mg_threshold
|
|
|
43 |
# Define the Segmentation-Based Blur Effect
|
44 |
# -----------------------------
|
45 |
def segmentation_blur_effect(input_image: Image.Image):
|
|
|
|
|
|
|
|
|
|
|
46 |
imageResized = input_image.resize(seg_image_size)
|
47 |
input_tensor = seg_transform(imageResized).unsqueeze(0).to(device)
|
48 |
|
|
|
50 |
preds = seg_model(input_tensor)[-1].sigmoid().cpu()
|
51 |
pred = preds[0].squeeze()
|
52 |
|
|
|
53 |
pred_pil = transforms.ToPILImage()(pred)
|
54 |
mask = pred_pil.resize(imageResized.size)
|
55 |
|
|
|
56 |
mask_np = np.array(mask.convert("L"))
|
57 |
_, maskBinary = cv2.threshold(mask_np, 127, 255, cv2.THRESH_BINARY)
|
58 |
|
|
|
59 |
img = cv2.cvtColor(np.array(imageResized), cv2.COLOR_RGB2BGR)
|
|
|
60 |
blurredBg = cv2.GaussianBlur(np.array(imageResized), (0, 0), sigmaX=15, sigmaY=15)
|
61 |
|
|
|
62 |
maskInv = cv2.bitwise_not(maskBinary)
|
63 |
maskInv3 = cv2.cvtColor(maskInv, cv2.COLOR_GRAY2BGR)
|
64 |
|
|
|
65 |
foreground = cv2.bitwise_and(img, cv2.bitwise_not(maskInv3))
|
66 |
background = cv2.bitwise_and(blurredBg, maskInv3)
|
67 |
|
|
|
68 |
finalImg = cv2.add(cv2.cvtColor(foreground, cv2.COLOR_BGR2RGB), background)
|
69 |
finalImg_pil = Image.fromarray(finalImg)
|
70 |
|
71 |
return finalImg_pil, mask
|
72 |
|
|
|
|
|
|
|
73 |
def lens_blur_effect(input_image: Image.Image, fg_threshold: float = 85, mg_threshold: float = 170):
|
74 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
75 |
inputs = depth_processor(images=input_image, return_tensors="pt").to(device)
|
76 |
with torch.no_grad():
|
77 |
outputs = depth_model(**inputs)
|
|
|
80 |
)
|
81 |
depth = post_processed_output[0]["predicted_depth"]
|
82 |
|
|
|
83 |
depth = (depth - depth.min()) / (depth.max() - depth.min())
|
84 |
depth = depth * 255.
|
85 |
depth = depth.detach().cpu().numpy()
|
86 |
depth_map = depth.astype(np.uint8)
|
87 |
depthImg = Image.fromarray(depth_map)
|
88 |
|
|
|
89 |
img = cv2.cvtColor(np.array(input_image), cv2.COLOR_RGB2BGR)
|
90 |
|
|
|
91 |
img_foreground = img.copy() # No blur for foreground
|
92 |
img_middleground = cv2.GaussianBlur(img, (0, 0), sigmaX=7, sigmaY=7)
|
93 |
img_background = cv2.GaussianBlur(img, (0, 0), sigmaX=15, sigmaY=15)
|
94 |
|
95 |
print(depth_map)
|
96 |
+
depth_map = depth_map.astype(np.float32) / depth_map.max()
|
97 |
|
98 |
+
threshold1 = fg_threshold
|
99 |
+
threshold2 = mg_threshold
|
|
|
100 |
|
|
|
101 |
mask_fg = (depth_map < threshold1).astype(np.float32)
|
102 |
mask_mg = ((depth_map >= threshold1) & (depth_map < threshold2)).astype(np.float32)
|
103 |
mask_bg = (depth_map >= threshold2).astype(np.float32)
|
104 |
|
|
|
105 |
mask_fg_3 = np.stack([mask_fg]*3, axis=-1)
|
106 |
mask_mg_3 = np.stack([mask_mg]*3, axis=-1)
|
107 |
mask_bg_3 = np.stack([mask_bg]*3, axis=-1)
|
108 |
|
|
|
109 |
final_img = (img_foreground * mask_fg_3 +
|
110 |
img_middleground * mask_mg_3 +
|
111 |
img_background * mask_bg_3).astype(np.uint8)
|
|
|
113 |
final_img_rgb = cv2.cvtColor(final_img, cv2.COLOR_BGR2RGB)
|
114 |
lensBlurImage = Image.fromarray(final_img_rgb)
|
115 |
|
|
|
116 |
mask_fg_img = Image.fromarray((mask_fg * 255).astype(np.uint8))
|
117 |
mask_mg_img = Image.fromarray((mask_mg * 255).astype(np.uint8))
|
118 |
mask_bg_img = Image.fromarray((mask_bg * 255).astype(np.uint8))
|
119 |
|
120 |
return depthImg, lensBlurImage, mask_fg_img, mask_mg_img, mask_bg_img
|
121 |
|
|
|
|
|
|
|
122 |
def process_image(input_image: Image.Image, fg_threshold: float, mg_threshold: float):
|
123 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
124 |
seg_blur, seg_mask = segmentation_blur_effect(input_image)
|
125 |
depth_map_img, lens_blur_img, mask_fg_img, mask_mg_img, mask_bg_img = lens_blur_effect(
|
126 |
input_image, fg_threshold, mg_threshold
|