Images have structure. Nearby pixels relate to each other, and patterns repeat across the frame. Convolutional networks exploit this by sliding small filters across the input, sharing weights and detecting local features wherever they appear.

A typical architecture stacks convolution layers, activation functions, and pooling. Early layers detect edges and textures. Deeper layers combine them into parts and objects. A final classifier maps the learned features to labels.

Key building blocks

• Convolution filters with learnable weights
• Pooling to reduce spatial dimensions
• Stride and padding to control output size
• Fully connected layers for classification

CNNs dominated vision for a decade, starting with AlexNet in 2012. Vision transformers now compete on many benchmarks, but CNNs remain efficient and widely deployed in embedded and mobile systems.