opencv.py

import cv2
import numpy as np
from keras.models import load_model
from skimage.transform import resize, pyramid_reduce

# Load your pre-trained model
model = load_model('model.h5')

def get_square(image, square_size):
    
    height, width = image.shape    
    if height > width:
        differ = height
    else:
        differ = width
    differ += 4

    mask = np.zeros((differ, differ), dtype="uint8")
    x_pos = int((differ - width) / 2)
    y_pos = int((differ - height) / 2)
    mask[y_pos: y_pos + height, x_pos: x_pos + width] = image[0: height, 0: width]

    if differ / square_size > 1:
        mask = pyramid_reduce(mask, differ / square_size)
    else:
        mask = cv2.resize(mask, (square_size, square_size), interpolation=cv2.INTER_AREA)
    return mask

def keras_predict(model, image):
    data = np.asarray(image, dtype="int32")
    pred_probab = model.predict(data)[0]
    pred_class = list(pred_probab).index(max(pred_probab))
    return max(pred_probab), pred_class

def keras_process_image(img):
    image_x = 28
    image_y = 28
    img = get_square(img, 28)
    img = np.reshape(img, (image_x, image_y))
    return img

def crop_image(image, x, y, width, height):
    return image[y:y + height, x:x + width]

cam_capture = cv2.VideoCapture(0)

while True:  
    _, image_frame = cam_capture.read()
    # Select ROI - replace these with actual bbox coordinates
    im2 = crop_image(image_frame, 300, 300, 300, 300)
    image_grayscale = cv2.cvtColor(im2, cv2.COLOR_BGR2GRAY)
    image_grayscale_blurred = cv2.GaussianBlur(image_grayscale, (15,15), 0)

    resized_img = keras_process_image(image_grayscale_blurred)
    ar = resized_img.reshape(1, 784)

    pred_probab, pred_class = keras_predict(model, ar)
    print(pred_class, pred_probab)
     
    # Display cropped image
    cv2.imshow("Image2", im2)
    cv2.imshow("Image4", resized_img)
    cv2.imshow("Image3", image_grayscale_blurred)

    if cv2.waitKey(25) & 0xFF == ord('q'):
        break

cam_capture.release()
cv2.destroyAllWindows()