LR-Inst: A lightweight and robust instance segmentation network for apple detection in complex orchard environments

Abstract

Apple instance segmentation is a critical task in the implementation of automated harvesting systems. Despite significant advances in instance segmentation, current methods remain impractical for deployment due to their architectural complexity and slow inference speeds. While lightweight models have been introduced to improve efficiency, their performance degrades in orchard environments under occlusion, fruit overlap, and varying lighting conditions. To address these challenges, we present LR-Inst, a lightweight and robust instance segmentation network. First, we design an innovative cross-level feature fusion architecture that exploits the rich spatial details and semantic information present in intermediate-layer features. Then, a set of efficient modules is designed to further boost feature representation, including the Spatial-Semantic Feature Fusion Module (SSFM), the Dynamic Spatial-Semantic Fusion Module (DSSFM), the Feature Aggregation and Shuffle Module (FASM), and the Channel-Spatial Attention Module (CSAM). Experimental results demonstrate that LR-Inst contains only 3.742 M parameters and requires 8.581 G FLOPs. When evaluated on our self-collected orchard dataset, LR-Inst achieves a detection average precision (AP) of 0.946 and a segmentation AP of 0.944, outperforming several state-of-the-art (SOTA) models.