A low-light RGB-thermal infrared target detection algorithm based on adaptive cross-modal fusion and multiscale feature enhancement

Abstract

Small-object detection under low illumination remains a persistent challenge in aerial surveillance, nighttime patrol, and safety-critical vision tasks, where single-modality sensors—RGB or infrared alone—provide degraded or incomplete target information. This paper proposes an RGB-T object detection network that combines adaptive cross-modal fusion with multi-scale feature enhancement. Two core modules are introduced. First, an adaptive modal fusion module (MSFusion) learns modality importance coefficients through a weight branch and applies spatial attention to dynamically fuse RGB and infrared features according to scene conditions. Second, a multi-scale feature extraction module (FES) employs dilated convolution with coordinate attention and a gating mechanism to amplify small-object feature responses while remaining parameter-efficient. On the LLVIP low-light benchmark, the proposed network achieves 65.6% mAP50- 95, a 2.2 percentage-point improvement over the YOLOv11-RGBT baseline, with 3.62 M parameters and 9.14 GFlops— fewer than the baseline.