OAFBC-Net: A Lightweight RGB-Depth Fusion Network for Occlusion-Aware Apple Detection in Complex Orchard Canopies

Accurate perception of occluded apples in tree canopies is important for precision orchard management and robotic harvesting, but RGB-only vision is sensitive to leaf occlusion, fruit overlap, and unstable illumination. This study presents OAFBC-Net, a lightweight RGB-depth network for joint localization and occlusion-level recognition of apples classified as non-occluded (NO), soft-occluded (SO), or hard-occluded (HO). OAFBC-Net integrates three components: a Fourier amplitude-phase enhancement fusion module (FAPEF) for cross-modal frequency-domain fusion, a weighted bidirectional feature pyramid (BiFPN) for multi-scale interaction, and a Triple Context Guided Block (TCG) for local-to-global semantic compensation. Using an orchard RGB-depth dataset with 3000 image pairs and 134,999 annotated instances acquired with an Intel RealSense D455 camera, OAFBC-Net achieves AP50 values of 72.1%, 87.3%, and 57.9% for SO, NO, and HO, respectively. Under the same split, it outperforms the self-implemented RGB-depth baselines. With FAPEF, BiFPN, and TCG enabled, the macro-average mAP50 and mAP50–95 reach 72.4% and 55.2%, respectively. The model has 3.6 M parameters. On a single image, the total latency is 11.3 ms, corresponding to an inference speed of 88.5 FPS. Three random seeds and five-fold cross-validation with a fixed independent test set show consistent gains over the dual-modal baseline. These findings support OAFBC-Net as a promising visual perception method for occluded-fruit detection in orchards, while cross-site and field-robot validation remain necessary for deployment-oriented claims.