RGB-IR YOLO combining Modality-Specific Reconstruction and Information Integration

Utilizing cross-modality information between infrared and visible light presents a novel approach to enhancing detection performance. However, existing methods for leveraging visible-infrared information are relatively simplistic, failing to thoroughly explore the relationship between infrared and visible light features. Therefore, this paper introduces RI-YOLO, a new framework for infrared-visible images object detection. Firstly, to retain the inherent advantages of each modality image, we propose G&M-DI, a technique for modality-specific reconstruction and cross-modality information fusion. Secondly, by introducing AWI, we employ an adaptive weight allocation strategy to mitigate cross-modality interference, effectively leveraging the relationship between different modalities. Experimental results on datasets M3FD and LLVIP demonstrate that RI-YOLO surpasses existing state-of-the-art (SOTA) models.