MCCS-Det: A Small-Scale Smoke Detection with YOLO26 Enhanced by Multi-Scale Context Compensation and a Self-Collaborative Head

Fire early warning is crucial for reducing disaster losses at the initial stage. As an important visual cue in early fire scenarios, smoke often appears with small scale, semitransparency, and loose morphology, which makes it prone to missed detection under low-contrast conditions. To address this issue, this paper proposes MCCS-Det, an early small-scale smoke detection network based on YOLO26. The proposed method introduces a Multi-Scale Context Compensation Module (MCCM) to enhance weak smoke features through parallel maxpooling, average-pooling, and residual reinjection, and further designs a Self-Collaborative Decoupled Detection Head (SCHead) to improve cross-layer alignment between classification and localization features. Experimental results on the curated FASDD_CV dataset show that MCCS-Det achieves an inference speed of 1143.52 FPS, while the precision and recall for small-scale smoke reach 0.672 and 0.690, respectively, demonstrating superior accuracy and real-time performance over existing mainstream methods.