False Alarm Rectification for Early Smoke Segmentation

Early smoke segmentation plays a critical role in forest protection and industrial safety. With the increasing deployment of fixed cameras and drones, vision-based smoke detection has become widely adopted. However, in open environments, smoke is easily confused with visually similar phenomena such as clouds, fog, and water vapour, leading to frequent false positives. To address this challenge, we propose SmokeCPC, a framework composed of a Confidence-Aware Prior Correction Module (CPCM) and a Cross-scale Fusion Module (CFM) that explicitly targets false alarm suppression in pixel-level smoke segmentation while preserving overall detection performance. The core idea of our method is to leverage the confidence of an image-level smoke classifier as a prior to guide both training and inference of the segmentation model. Building on this idea, CPCM strengthens supervision for high-confidence samples to enhance discriminative capability, while down-weighting low-confidence samples to mitigate noise propagation. In parallel, CFM integrates texture and semantic cues across feature hierarchies, improving robustness to thin smoke plumes and complex backgrounds. We further incorporate a supervised contrastive loss to encourage intra-class compactness and inter-class separability in the feature space. Overall, our method reduces the false positive rate without sacrificing segmentation quality. Experiments on the SmokeSeg dataset demonstrate the effectiveness of our approach, achieving an IoU of 61.83% and an FPR of only 0.28%.