LRD-Inst: a lightweight and robust dual-branch framework for instance segmentation of mixed bagged and unbagged apples.

Introduction: Amid global labor shortages, automated harvesting robots are essential for enhancing agricultural productivity, with robust instance segmentation serving as the core vision task. However, existing methods fail to balance high-fidelity boundary delineation and real-time efficiency under severe visual degradations caused by protective fruit bagging and dense canopy occlusions. Methods: To resolve these limitations, LRD-Inst, a lightweight and robust dual-branch instance segmentation framework, is introduced for unstructured orchards and resource-constrained edge platforms. The architecture explicitly decouples feature extraction: a spatial pathway utilizes Parallel Hierarchical Enhancement Blocks (PHEB) and Frequency-Decoupled Spatial Pyramids (FDSP) to safeguard high-frequency boundary cues, while a contextual branch embeds a High-frequency Detour State Space Model (HDSSM) to capture long-range global dependencies for obscured targets. A Spatially-Refined Adaptive Fusion (SRAF) module bridges these pathways, optimized via an Area-Stratified Dice (AS-Dice) loss to reinforce small-target geometric fidelity. Results: Extensive experiments on a mixed-apple dataset demonstrate that LRD-Inst achieves a primary Average Precision (AP) of 0.568 with only 3.43 M parameters and 9.12 GFLOPs, outperforming contemporary baselines including the YOLOv8-YOLOv26 families and RTMDet. The model operates at 45.4 FPS on an NVIDIA RTX 3060 GPU. Discussion: LRD-Inst establishes an optimal equilibrium between accuracy and efficiency, providing a highly deployable solution for autonomous agricultural robotics. The source code is available at https://github.com/ly27253/LRD-Inst.