Pixel Value Normalization for Grayscale Images in CNNs?

Hello fellow developers👾,

I'm delving into an image classification project using a CNN and encountered a preprocessing dilemma. My pipeline currently transforms RGB images into grayscale and resizes them to 80x80 pixels. However, pixel value normalization (0-1 range) has been introduced only into my predict function.

Here's a snippet from the current implementation from the mainFileTweaks-branch:

def load_and_process_images(path):
    x_data = [] 
    for img in os.listdir(path):
        if img.endswith('png'):
            pic = cv2.imread(os.path.join(path, img))
            pic = cv2.cvtColor(pic, cv2.COLOR_BGR2GRAY)  # Convert to grayscale
            pic = cv2.resize(pic, (80, 80))  # Resize to 80x80
            #pic = pic / 255.0  #Not present in code, should we normalize pixel values to [0, 1], here as well?
            x_data.append([pic])
    return x_data

def predict_image(model, image_path, image_size=(80, 80)):
    img = cv2.imread(image_path)
    img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)  # Convert to grayscale
    img = cv2.resize(img, image_size)  # Resize the image
    img = img.reshape(1, image_size[0], image_size[1], 1)  # Reshape for the model
    img = img / 255.0  # Normalization line, already present
    

does omitting normalization in such cases significantly affect CNN training efficacy, particularly for grayscale images? Are there specific scenarios where normalization isn't critical?

Start with load_and_process_images and predict_image on the mainFileTweaks-branch, then inspect how their outputs are fed to the CNN. Verify that the training and prediction preprocessing paths use a deliberate, matching pixel-value convention, and document or test the chosen behavior.