A reliability-guided RGB-IR object detection network with complementary information decoupling for autonomous driving
Year: 2026
Type: article
Abstract: Autonomous driving systems require robust object detection under complex open-road conditions, where RGB images are sensitive to illumination changes and background interference, while infrared images may suffer from weak textures and unreliable thermal responses. To improve RGB-IR detection performance, this paper proposes a Feature-level Scale-aware Multispectral Fusion Network, termed FSMF, for autonomous driving scenarios. The network consists of three key components. First, a Reliability-Guided Modulation module estimates RGB and IR feature reliability from channel, spatial, and cross-modal perspectives, and adaptively adjusts their contributions before fusion. Second, a Complementary Information Decoupling module separates shared semantics, complementary differences, and redundant interference, thereby enhancing target-related cross-modal cues while suppressing ineffective responses. Third, a Scale-Aware Fusion Pyramid Network performs differentiated enhancement for shallow, middle, and deep features to improve multi-scale feature representation. Experiments on M3FD and FLIR demonstrate the effectiveness of FSMF. The proposed FSMF-s achieves 85.9% mAP@50 on M3FD, and 72.2% mAP@50 and 38.3% mAP@50:95 on FLIR, outperforming mainstream detectors and existing RGB-IR methods. Ablation results further verify the effectiveness of the proposed modules.
Source: Scientific Reports
Authors: Miaomiao Yang, Fan Guo, Lin Cheng, Ping Fang
Institution: Chuzhou Polytechnic
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7220416628