BiGCFormer: A Unified Multimodal Framework for Robust Raspberry Detection in Complex Agricultural Scenes

High-precision fruit recognition and localization in complex agricultural settings is a key element for intelligent orchard management and precise harvesting. However, traditional unimodal detection is constrained in meeting robustness and accuracy needs due to natural interferences like uneven illumination and foliage occlusion. Therefore, BiGCFormer, a Bi-stream Graph Cross-attention Fusion Detection Network, is proposed. It is a fusion-driven multimodal raspberry detection system integrating multi-source visual perception and deep learning, thereby, enhancing detection stability in complex scenes. Building upon the YOLOv11 architecture, it incorporates dual-stream input backbone networks for the respective processing of RGB and depth modal features, and there are three key technological innovations: (1) spatial-semantic structural alignment capabilities between RGB and depth features are strengthened by a modal-aware positional encoding module; (2) complementary enhancement pathways between RGB-D(depth) modalities are established by a bidirectional cross-attention mechanism to extract multimodal information interaction features; and (3) dynamic adjustment of fusion intensity is implemented by a dual-gated residual fusion mechanism through learnable gating functions, with robustness against modal heterogeneity and noise interference enhanced. Experimental evaluations on raspberry fruit detection datasets demonstrate that the proposed BiGCFormer model achieves a mean Average Precision at IoU threshold 0.5 (mAP@.5) of 70.9%, representing a 2.1 percentage point improvement over the Fusion-Mamba (FMB) methodology, with comprehensive enhancements across key metrics such as precision and recall. Compared with existing multimodal detection models, the proposed BiGCFormer exhibits superior detection stability and generalization capabilities under varying fruit sizes, poses, and occlusion conditions.