Why convolutions
A convolution slides a small learned kernel across an image, sharing weights across positions. That gives translation equivariance and cuts parameters by orders of magnitude compared with a dense layer over pixels.
The architectures that mattered
AlexNet (2012) proved GPU-trained deep nets on ImageNet; VGG showed depth with small kernels; ResNet introduced residual blocks and reached hundreds of layers; EfficientNet balanced depth, width and resolution.
Beyond classification
Detection (YOLO, Faster R-CNN) predicts boxes; segmentation (U-Net, Mask R-CNN) predicts per-pixel labels. Vision transformers now split an image into patches and apply attention, matching or beating CNNs when pretraining data is plentiful.