A simple script to perform webcam visual anomaly detection with autoencoders built with Keras

visual_anomaly_detection_demo.py

#!/usr/bin/python3
from __future__ import absolute_import, print_function
import cv2
import numpy as np
from time import sleep, time
from collections import deque

from tensorflow.keras.models import Model
from tensorflow.keras.optimizers import RMSprop
from tensorflow.keras.layers import Input, Conv2D, MaxPooling2D, UpSampling2D, ZeroPadding2D
from tensorflow.keras.callbacks import History

def build_model(img_size=32):
    input_img = Input(shape=(img_size, img_size, 1))
    x = Conv2D(32, (5, 5), activation='relu', padding='same')(input_img)
    x = MaxPooling2D((2, 2), padding='same')(x)
    x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
    encoded = MaxPooling2D((2, 2), padding='same', name='encoder')(x)

    x = Conv2D(8, (3, 3), activation='relu', padding='same')(encoded)
    x = UpSampling2D((2, 2))(x)
    x = Conv2D(32, (3, 3), activation='relu', padding='same')(x)
    x = UpSampling2D((2, 2))(x)
    decoded = Conv2D(1, (5, 5), activation='sigmoid', padding='same')(x)

    autoencoder = Model(input_img, decoded)
    autoencoder.compile(optimizer=RMSprop(lr=lr), loss='binary_crossentropy')

    return autoencoder

def load_run_render(frame):
    im = cv2.resize(frame, (img_size, img_size))
    im = cv2.cvtColor(im, cv2.COLOR_RGB2GRAY)
    im = np.expand_dims(np.expand_dims(im, axis=0).astype(np.float32), axis=-1) / 255
    history = History()
    start_time = time()
    autoencoder.fit(im, im,
                    epochs=1,
                    batch_size=1,
                    shuffle=True,
                    callbacks=[history])
    end_time = time()
    return history.history['loss'][-1]

def main():
    # Begin Video Capture
    cam = cv2.VideoCapture(0)
    idx = 0
    while True:
        try:
            ret, frame = cam.read()
            ts = int(time() * 1000)
            print(idx)
            if ret:
                loss = load_run_render(frame)
                q.append(loss)
                if len(q) > 50:
                    prev_hist = list(q)[:-1]
                    threshold = 5 * np.std(prev_hist)
                    if np.abs(loss - np.mean(prev_hist)) > threshold:
                        print(loss, 'anomaly')
                    else:
                        print(loss, 'normal')
                else:
                    pass
        except Exception as e:
            print(e)
            pass
        sleep(DELAY)
        idx += 1
        
if __name__ == '__main__':
    # Declare Parameters
    lr = 1e-4
    DELAY = 0.2
    img_size=128
    q = deque(maxlen=300)

    # Build the Autoencoder
    autoencoder = build_model(img_size=img_size)

    # Run the Training/Inference Loop
    main()