Prototype-based scatter learning for smoke segmentation

Abstract

Smoke segmentation enables precise localization of smoke for firefighting and gas leak detection. While current approaches focus on multi-scale feature aggregation, challenges remain in handling blurred edges and diverse smoke patterns. We propose prototype scatter learning (PSL), which improves smoke segmentation through unified optimization of feature extractors and prototypes. PSL introduces two key innovations: bottomK eigenvalues scatter (BES) loss for enhanced discrimination, and prototype uncorrelated optimization (PUO) for maintaining prototype diversity. Experimental results show that PSL achieves state-of-the-art performance on mainstream smoke segmentation datasets (SmokeSeg and SMOKE5K). The code can be found at https://github.com/LujianYao/psl.