ImageDataGenerator gives different result than cv2

Question

I have been trying to create a neural network architecture that recognizes traffic signs. I am using the German Traffic Signs dataset and is composed of 43 classes. First of