On the importance of normalizing images

I’ve been using pre-trained 152-resnet’s for some time now for transfer learning fine-tuning the top few layers and noticed that whether or not I normalize the images during preprocessing doesn’t really help. Now I believe the pre-trained resnets were trained on normalized images, so I would assume that the lower layers at the very least would be tuned to normalized images, but it seems like it doesnt matter. Is there a reason for this? Am I messing something up potentially?

Thanks