RT-DETR-MCDAF: Multimodal Fusion of Visible Light and Near-Infrared Images for Citrus Surface Defect Detection in the Compound Domain

The accurate detection of citrus surface defects is essential for automated citrus sorting to enhance the commercialization of the citrus industry. However, previous studies have only focused on single-modal defect detection using visible light images (RGB) or near-infrared light images (NIR), without considering the feature fusion between these two modalities. This study proposed an RGB-NIR multimodal fusion method to extract and integrate key features from both modalities to enhance defect detection performance. First, an RGB-NIR multimodal dataset containing four types of citrus surface defects (cankers, pests, melanoses, and cracks) was constructed. Second, a Multimodal Compound Domain Attention Fusion (MCDAF) module was developed for multimodal channel fusion. Finally, MCDAF was integrated into the feature extraction network of Real-Time DEtection TRansformer (RT-DETR). The experimental results demonstrated that RT-DETR-MCDAF achieved Precision, Recall, mAP@0.5, and mAP@0.5:0.95 values of 0.914, 0.919, 0.90, and 0.937, respectively, with an average detection performance of 0.598. Compared with the model RT-DETR-RGB&NIR, which used simple channel concatenation fusion, RT-DETR-MCDAF improved the performance by 1.3%, 1.7%, 1%, 1.5%, and 1.7%, respectively. Overall, the proposed model outperformed traditional channel fusion methods and state-of-the-art single-modal models, providing innovative insights for commercial citrus sorting.