Key points are not available for this paper at this time.
• Building on module-level enhancements, we propose YOLOv8-PM—a lightweight, high-accuracy mold-detection model capable of precisely detecting subtle mold features on pine nuts. • Layer-adaptive magnitude-based pruning (LAMP) is adopted to remove redundant parameters, achieving a lighter model while further boosting detection accuracy. • Deployment on a Jetson Nano was completed, with inference latency kept to 31. 4 ms to meet real-time detection requirements. To mitigate pine-nut yield loss and quality deterioration caused by Aspergillus flavus rot, this study proposes YOLOv8-PMP, a detector for pine-nut rot. Based on YOLOv8n, the backbone integrates a large separable kernel attention (LSKA) module with spatial pyramid pooling fast (SPPF) to strengthen multi-scale feature extraction. In the neck, deformable ConvNets v4 (DCNv4) and a dynamic sampling (DySample) operator improve adaptability to irregular lesion regions while reducing computation, and a ResBlockGAM module strengthens multi-scale feature fusion. We further compress the network using layer-adaptive magnitude-based pruning (LAMP). Deployed on an NVIDIA Jetson Nano, YOLOv8-PMP achieves precision, recall, mAP50, and mAP50–95 of 93. 5%, 86. 9%, 93. 3%, and 65. 6%, outperforming the baseline by up to 11. 7 percentage points, with an average latency of 31. 4 ms per image. These results demonstrate the potential of YOLOv8-PMP for real-time, high-precision rot detection in industrial applications.
Bao et al. (Sun,) studied this question.