It is not the individual pixel values, but their relationships, right? The algorithm can learn the geometric relationships, like edges and curves and then putting those together to recognize higher level features: e.g. two edges that meet at a certain angle are probably a cat’s ear.
Another way to see the point I’m making above is to consider what would happen if you took the images and just subtracted 5 from all the pixel values. The shapes would still be the same, right? A cat still looks like a cat if you slightly change all the colors in the image. In fact the idea of randomly perturbing the pixel values is a legitimate technique for “data augmentation” to use when you have limited training data. Speaking of which …
The sample datasets here are incredibly small for the complexity of this task. It turns out the datasets are very carefully curated so that it even works as well as it does. But you’ll notice that if you try the model on your own imported images, the performance is not that great. Here’s a thread that shows some experiments with rearranging the training and test sets.