app.py

from transformers import DetrImageProcessor, DetrForObjectDetection, TrOCRProcessor, VisionEncoderDecoderModel
import cv2
from PIL import Image, ImageDraw
import torch
import streamlit as st

# Load Hugging Face Models
detr_processor = DetrImageProcessor.from_pretrained("facebook/detr-resnet-50")
detr_model = DetrForObjectDetection.from_pretrained("facebook/detr-resnet-50")
trocr_processor = TrOCRProcessor.from_pretrained("microsoft/trocr-base-stage1")
trocr_model = VisionEncoderDecoderModel.from_pretrained("microsoft/trocr-base-stage1")

# Detect license plates
def detect_license_plate(frame):
    pil_image = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
    inputs = detr_processor(images=pil_image, return_tensors="pt")
    outputs = detr_model(**inputs)

    target_sizes = torch.tensor([pil_image.size[::-1]])
    results = detr_processor.post_process_object_detection(outputs, target_sizes=target_sizes, threshold=0.9)

    return results[0]["boxes"], pil_image

# Recognize text
def recognize_text_from_plate(cropped_plate):
    inputs = trocr_processor(images=cropped_plate, return_tensors="pt")
    outputs = trocr_model.generate(**inputs)
    return trocr_processor.batch_decode(outputs, skip_special_tokens=True)[0]




# Streamlit configuration
st.title("Real-Time Car Number Plate Recognition")
st.text("This application uses Hugging Face Transformers to detect and recognize car plates.")

# Authorized car database
authorized_cars = {"KA01AB1234", "MH12XY5678", "DL8CAF9090"}

# Verification function
def verify_plate(plate_text):
    if plate_text in authorized_cars:
        return f"✅ Access Granted: {plate_text}"
    else:
        return f"❌ Access Denied: {plate_text}"


# Live video feed and processing
def live_feed():
    cap = cv2.VideoCapture(0)  # Open the webcam
    stframe = st.empty()  # Streamlit frame for displaying video

    while cap.isOpened():
        ret, frame = cap.read()
        if not ret:
            break

        # Detect license plates
        boxes, pil_image = detect_license_plate(frame)
        draw = ImageDraw.Draw(pil_image)

        recognized_plates = []
        for box in boxes:
            # Crop the detected plate
            cropped_plate = pil_image.crop((box[0], box[1], box[2], box[3]))

            # Recognize text
            plate_text = recognize_text_from_plate(cropped_plate)
            recognized_plates.append(plate_text)

            # Draw bounding box and text
            draw.rectangle(box.tolist(), outline="red", width=2)
            draw.text((box[0], box[1]), plate_text, fill="red")

        # Convert PIL image back to OpenCV format
        processed_frame = cv2.cvtColor(np.array(pil_image), cv2.COLOR_RGB2BGR)

        # Stream the video to Streamlit
        stframe.image(processed_frame, channels="BGR")

        # Show results
        for plate_text in recognized_plates:
            st.write(verify_plate(plate_text))

    cap.release()
    cv2.destroyAllWindows()

if st.button("Start Camera"):
    live_feed()