app.py

import gradio as gr
import re, datetime,time, cv2, numpy as np, tensorflow as tf, sys


CHARS = "ABCDEFGHIJKLMNPQRSTUVWXYZ0123456789" # exclude I, O
CHARS_DICT = {char:i for i, char in enumerate(CHARS)}
DECODE_DICT = {i:char for i, char in enumerate(CHARS)}

interpreter = tf.lite.Interpreter(model_path='detection.tflite')
interpreter.allocate_tensors()
recog_interpreter = tf.lite.Interpreter(model_path='recognition.tflite')
recog_interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
recog_input_details = recog_interpreter.get_input_details()
recog_output_details = recog_interpreter.get_output_details()



def execute_text_recognition_tflite( boxes, frame, interpreter, input_details, output_details):
    x1, x2, y1, y2 = boxes[1], boxes[3], boxes[0], boxes[2]
    save_frame = frame[
        max( 0, int(y1*1079) ) : min( 1079, int(y2*1079) ),
        max( 0, int(x1*1920) ) : min( 1920, int(x2*1920) )
    ]

    # Execute text recognition

    test_image = cv2.resize(save_frame,(94,24))/256
    test_image = np.expand_dims(test_image,axis=0)
    test_image = test_image.astype(np.float32)
    interpreter.set_tensor(input_details[0]['index'], test_image)
    interpreter.invoke()
    output_data = interpreter.get_tensor(output_details[0]['index'])
    decoded = tf.keras.backend.ctc_decode(output_data,(24,),greedy=False)
    text = ""
    for i in np.array(decoded[0][0][0]):
        if i >-1:
            text += DECODE_DICT[i]
    # Do nothing if text is empty
    if not len(text): return 
    license_plate = text
    text[:3].replace("0",'O')

    return text

def greet(image):
    resized = cv2.resize(image, (320,320), interpolation=cv2.INTER_AREA)
    demo_frame = cv2.resize(image, (680,480), interpolation=cv2.INTER_AREA)
    input_data = resized.astype(np.float32)          # Set as 3D RGB float array
    input_data /= 255.                               # Normalize
    input_data = np.expand_dims(input_data, axis=0)  # Batch dimension (wrap in 4D)

    # Initialize input tensor
    interpreter.set_tensor(input_details[0]['index'], input_data)
    interpreter.invoke()
    output_data = interpreter.get_tensor(output_details[0]['index'])

    # Bounding boxes
    boxes = interpreter.get_tensor(output_details[1]['index'])

    text = None
    # For index and confidence value of the first class [0]
    for i, confidence in enumerate(output_data[0]):
        if confidence > .3:
            text = execute_text_recognition_tflite(
                boxes[0][i], image,
                recog_interpreter, recog_input_details, recog_output_details,
            )
            return text
image = gr.inputs.Image(shape=(320,320))

iface = gr.Interface(fn=greet, inputs=image, outputs="text")
iface.launch()