Adaptive Multi-Scale Fusion in Lightweight YOLO for Agricultural Defect Detection in Potato Processing Lines

To address the challenges of undetected defective potatoes and real-time processing requirements for surface defect detection of peeled potatoes in intelligent pre-cooked food production lines, this study proposes an enhanced YOLOv5 framework incorporating Adaptive Spatial Feature Fusion (ASFF) and Slimneck lightweight module. The ASFF module facilitates adaptive weighted fusion of multi-scale defect features, significantly improving model robustness against small-scale defects including blackening, sprouting, and mechanical damage. Concurrently, the Slimneck module enhances computational efficiency through parameter redundancy reduction while maintaining detection speed. Experimental results demonstrate that the optimized model achieves 87.1% mAP50 with 12.54 M parameters and 22.9G floating-point operations (FLOPs). Specifically, the model achieves 89.3% precision on the critical bad category, surpassing the baseline YOLOv5s by 2.3 percentage points. In practical deployment at Shandong Yinying's facility, the system achieves over 98% overall accuracy in the binary classification setting. This work provides a reliable and efficient visual inspection solution for agricultural product sorting equipment, demonstrating significant industrial application potential.