Convolutional networks (CNN) for images

In short: A convolutional network (CNN) applies small learnable filters across an image and thereby recognises edges, shapes and finally objects. The same filters apply at every position, which saves parameters.

In more detail: Images have structure: nearby pixels belong together and an object is the same everywhere in the image. Convolution exploits that.

In detail

• Convolution: a filter (e.g. 3×3) slides over the image and forms weighted sums. Each filter looks for a feature, such as a vertical edge.
• Feature maps: each layer produces many maps, early layers detect edges, middle ones textures, late ones parts and objects.
• Pooling: shrinks the maps (maximum of the region), makes the network robust against small shifts.
• Stride and padding: step width and border padding control the output size.
• Milestones: LeNet (digits, 1998), AlexNet (2012, breakthrough on ImageNet), VGG, ResNet, EfficientNet.
• Use: image classification, object detection (YOLO), segmentation, medical images, face recognition, also audio spectrograms.
• Today: vision transformers (Vision transformer and CLIP) take over much, CNNs stay efficient on small devices.

See also: Neurons, layers and activation functions, Image generation with diffusion models