Classification labels a whole image. Segmentation labels every pixel. The task divides an image into regions, assigning each pixel to a class such as road, pedestrian, sky, or tumour. It is essential for medical imaging, autonomous driving, and satellite analysis.

Three variants matter. Semantic segmentation labels each pixel by category. Instance segmentation distinguishes individual objects within a category. Panoptic segmentation combines both, labelling every pixel and separating instances.

Common segmentation architectures

• U-Net for biomedical images
• Mask R-CNN for instance segmentation
• DeepLab for semantic segmentation
• Segment Anything for general-purpose masks

Annotation is costly. Labelling every pixel in a dataset takes far longer than drawing bounding boxes. Weak supervision and interactive tools help, but high-quality segmentation data remains a bottleneck for new domains.