YOLOv8-Orah: An Improved Model for Postharvest Orah Mandarin (Citrus reticulata cv. Orah) Surface Defect Detection

Orah mandarin (Citrus reticulata cv. Orah) lacks systematic grading treatment after harvesting, resulting in a high fresh fruit loss rate and affecting the economic value. There are many drawbacks to traditional manual and mechanical sorting. Therefore, intelligent, rapid, non-destructive surface defect detection technology is significant. In addition to the fruit size, surface defects (e.g., canker, sunburn) are another important criterion for grading fruit. To overcome the challenges in detecting surface defects of orah mandarin, like multi-scale features, significant size differences, and slow convergence speed, we propose the YOLOv8-Orah detection model based on YOLOv8n. Path Aggregation Network (PANet) is replaced by a Focusing Diffusion Pyramid Network (FDPN), and the Diffusion and Spatial Interaction (DASI) module is introduced to effectively fuse and enhance features of different scales and improve detection accuracy. The Bottleneck in the C2f module is replaced by the Hybrid Dilated Residual Attention Block (HDRAB) module to reduce missed detections and false detections. We also introduce the NWD-CIoU joint bounding box loss to accelerate the convergence speed and improve the detection accuracy of small defects. The experimental results show that the improved YOLOv8-Orah model performs well in terms of precision, recall, and average precision, reaching 81.9%, 78.8%, and 84.2%, respectively. Compared with the original YOLOv8n, the improved model increased by 4.0%, 1.7%, and 3.0%, respectively. Meanwhile, the parameter count decreased by 7.76%. Compared with other mainstream models, YOLOv8-Orah achieves a good balance between detection accuracy and computational efficiency. The results technically support defect detection in postharvest orah mandarin and real-time grading of their quality. Meanwhile, it can promote the intelligent development of the bergamot industry.