Dense Prediction as an Extension: Segmentation Heads and Encoder–Decoder Intuition

Image classification collapses an entire chart, floor plan, or satellite tile into one label — often too coarse for risk mapping. Semantic segmentation assigns a class to every pixel: which roof is damaged, which lane is road, which region of a volatility surface is in a stress regime. This note introduces dense prediction through encoder–decoder intuition: a CNN backbone compresses spatial context, then a decoding path restores resolution for per-pixel decisions.