Semantic segmentation is a computer vision task where each pixel in an image is classified into a category, such that the entire image is partitioned into semantically meaningful regions. Unlike object detection (which predicts bounding boxes) or image classification (which assigns a single label to the whole image), semantic segmentation produces a dense prediction: a label map assigning a class to every pixel. For example, in a street scene, all pixels corresponding to the road are labeled "road", all car pixels "car", and so on — regardless of individual object instances.

Semantic segmentation is widely used in applications such as autonomous driving (understanding the layout of roads, sidewalks, and vehicles), medical imaging (highlighting tumors or organs), and satellite image analysis. Architectures typically used for semantic segmentation include Fully Convolutional Networks (FCNs), U-Net, DeepLab, and SegFormer, which preserve spatial information and upscale predictions to match input resolution.

A common challenge is class imbalance, where background classes dominate the image. Techniques like class weighting or data augmentation help address this. Evaluation metrics include Intersection over Union (IoU) and Pixel Accuracy, which assess the overlap between predicted and ground-truth masks. In summary, semantic segmentation is essential for fine-grained scene understanding, enabling machines to perceive visual input at the pixel level.