Resizing to a fixed size is a common first step in deep learning approaches, but that fixed size is generally large enough to keep a lot of image data. They don't usually pixellate down to 8x8 or whatever it is that the mosaic filter gives you.