Why do CNNs work so well with images?

Simply, due to kernels .

Kernel is a window that scans the image. kernels could be different, one kernel could be responsible for detecting arcs within the image features, another is responsible for diagonal lines, etc.

Kernel weights arrangement determine the feature it will detect.

Convolution means, incorporating multiple points to judge a single point. This process is done by kernels/filters.

Kernel’s advantage is, incorporating the surrounding pixels for judging the centre pixel feature type, unlike regular neural networks which depends only on the pixel itself, regardless of its surroundings.

Image is a coherent structure, each pixel is a part of a feature. Pixels depend on each other to form a feature, so CNN is the best choice for feature extraction and working well with images.