YOLO-Spectra-Net: Reliability-Aware RGB-Infrared Fusion for Robust Object Detection Under Adverse Conditions.

Robust object detection in practical sensing systems remains challenging under low visibility, rapidly varying illumination, occlusion, and cluttered backgrounds. Visible images provide rich structural and texture information but are vulnerable to darkness, haze, and glare, whereas infrared images preserve thermal saliency but may suffer from weak spatial detail and thermal crossover. Conventional RGB-infrared fusion strategies often assign fixed or symmetric importance to the two modalities, allowing degraded features to propagate through the detector and impairing perception stability. To address this issue, we propose YOLO-Spectra-Net (SNYOLO), a reliability-aware RGB-infrared fusion framework for adverse-condition object detection. SNYOLO estimates modality reliability and integrates it into adaptive feature fusion, multi-scale perception enhancement, and conflict-aware spatial suppression. The proposed design adaptively emphasizes reliable modality cues, attenuates locally inconsistent responses, and stabilizes multi-scale feature representations while maintaining a relatively lightweight multimodal inference structure. Experiments on a self-collected railway monitoring dataset and two public RGB-infrared benchmarks, M3FD and FLIR, demonstrate improved robustness and stable target-focused responses under challenging sensing conditions. SNYOLO achieves 82.7% mAP@0.5 on M3FD and 79.1% mAP@0.5 on FLIR, with competitive localization accuracy and lower computational cost than most compared multimodal baselines.