Pre-processing images for fine-tuning

Question

When you are fine-tuning a CNN like ResNet, VGG, EfficientNet, etc and you want to train the model with your own images, or even when you want to do a inference with any image of your dataset, do you have to do the same preprocessing steps to your images like do the same mean subtraction with the same values as the ones that was done to originally train that model? Or is it not necessary to get good results?

Thank you.

Answer 1

You will get the best results if the new images are preprocessed in the exact same way the training images were preprocessed.

The biggest issue is resizing the images to be the same size as the input layer expects.