A lightweight RGB-D detection network for small Prunus humilis Bunge targets in complex orchard scenes

Detecting ripe Prunus humilis Bunge ( P. humilis ) in natural orchards is challenging because targets are small, densely distributed, occluded, and embedded in reflective or texture-rich backgrounds. We introduce RePRF-Net, a lightweight RGB-D detection framework built around reverse-enhanced progressive residual fusion. The framework takes paired RGB images and depth maps generated offline from the corresponding RGB images by Depth Anything V2. The model uses a dual-path RGB-D backbone to encode appearance and generated depth cues, a P2-P5 progressive residual path to preserve fine boundaries, a reverse residual reinjection branch to return calibrated high-level fusion cues to high-resolution detection layers, and a multi-order gated context aggregation (MGCA) module to suppress background-induced responses. A progressive stability and stride enhancement (PSSE) auxiliary loss regularizes training, and the final deployable model is obtained through structured channel pruning followed by feature and output distillation. On the held-out test set, the final Pruned RePRF-Net achieves an mAP@0.5 of 0.9162 and an mAP@0.5:0.95 of 0.6158. Compared with the lightweight YOLO11n baseline, the final model improves mAP@0.5 by 0.0416. With 2.77 M parameters and 6.88 GFLOPs, Pruned RePRF-Net runs at approximately 32 FPS on an NVIDIA Jetson Orin Nano using an FP16 TensorRT engine. Mechanism-level evaluations further show that the fusion path improves size-stratified and dense/occluded subset performance, whereas MGCA reduces pure-background false positives under complex background conditions. These results indicate that reliable RGB-D fusion and structured compression can support deployable fruit detection in complex orchard scenes.