R2F-YOLO: Multiscale Cross-Modal Residual Fusion for RGB–Thermal Wildfire Flame Detection
TL;DR
R2F-YOLO is proposed, a dual-modality detection network that fuses RGB and thermal infrared imagery for wildfire flame detection and achieves the best overall result among the methods evaluated under the same RGBT-3 M protocol while maintaining real-time inference capability.
Abstract
Uncrewed aerial vehicles provide effective platforms for early wildfire detection owing to their mobility and broad coverage. However, conventional single-modality RGB detection methods exhibit degraded robustness and localization accuracy in complex scenarios involving low-light conditions, smoke occlusion, high-brightness backgrounds, and small-scale flame targets. This article proposes residual-to-residual fusion YOLO (R2F-YOLO), a dual-modality detection network that fuses RGB and thermal infrared imagery for wildfire flame detection. The architecture employs a dual-branch backbone to extract multiscale features from both modalities. A cross-scale residual fusion module continuously injects shallow fine-grained features into deeper semantic representations, enhancing small target discriminability. A reverse residual enhancement (RRE) mechanism propagates fused evidence back to high-resolution lateral pathways, strengthening small-scale flame representation. Coordinate attention is specifically embedded at cross-modal fusion nodes for position-sensitive feature recalibration, suppressing background interference and pseudofire responses. Content-Aware ReAssembly of FEatures content-aware upsampling synergizes with RRE to form a complete small-target enhancement pathway from evidence injection to high-fidelity reconstruction. Experiments on RGBT-3 M demonstrate that R2F-YOLO achieves 93.1% Precision, 93.9% Recall, 96.6% mAP@0.5, and 63.0% mAP@0.5:0.95, achieving the best overall result among the methods evaluated under the same RGBT-3 M protocol while maintaining real-time inference capability. Comprehensive ablation studies and cross-scale feature visualization validate the contribution and synergistic effects of each component. Zero-shot transfer evaluation on FireMan-UAV-RGBT and UAV-StrawFire datasets, including failure case analysis, demonstrates cross-domain generalization capability and identifies transferability boundaries.