PlumNet: a dual-path attention-optimized framework for high-precision green plum surface detection

Abstract Green plum, a fruit with both edible and medicinal value, has a long history of cultivation in China. Due to its sour taste, it is often processed into preserved fruit and fruit wine. Currently, defect screening before processing relies on manual inspection, which is inefficient and costly. To address this, this study constructs PlumNet, an intelligent detection network for green plum surface defects based on the YOLOv8n model. The system detects and locates four defects: decay, scar, crack, and rain spot. By adding the Dual-Path Squeeze-Excitation Residual Block to YOLOv8n’s backbone, spatial location information is better captured, enhancing feature expression. Additionally, the P2 small-target detection layer improves detection of small defects, such as rain spots, increasing detection accuracy. The final PlumNet model achieves a detection accuracy of 0.964, with a recall rate of 0.959 and mAP50 of 0.985. Compared to YOLOv8n, YOLOv8n-SimAM, YOLOv8n-DPSE, and YOLOv8n-S2, PlumNet achieves the highest F1-Measure (0.961) and delivers the best detection performance. This model enables multi-defect detection for green plums and provides valuable insights for deep-processing enterprises.