DSC-Det: A Detail–Scale–Context Detection Network for Forest-Fire-Oriented Early Fire and Smoke Detection in UAV-View and Complex-Background Imagery

Abstract

Early and reliable fire and smoke detection is essential for forest-fire warning and emergency response, especially in UAV-view and complex-background imagery, where small fire spots and diffuse smoke are easily affected by illumination variations and visually similar high-brightness or cloud- and fog-like backgrounds. To address these challenges, this paper formulates early fire and smoke recognition as a bounding-box detection task and proposes a Detail–Scale–Context Detection Network, named DSC-Det. DSC-Det is designed as a lightweight one-stage detection network and introduces three task-oriented components: a Detail–Context Downsampling Module (DCDM) for reducing information loss during early feature compression, a Dynamic Dual-Branch Fusion Module (DDFM) for adaptive multi-scale feature interaction under complex backgrounds, and a Shared-Regression Asymmetric Classification Head (SACH) for improving classification adaptation across feature layers while maintaining shared regression. Experiments on a constructed forest-fire-oriented fire and smoke dataset for UAV-view and complex-background monitoring scenes show that DSC-Det achieves 90.1% mAP@0.5 and 66.9% mAP@0.5:0.95, outperforming the lightweight reference detector by 2.3% and 4.4%, respectively. The results demonstrate that DSC-Det improves early forest-fire and smoke detection with controlled model complexity.