YOLO Loss Optimization for Detecting Fruit Defects

Japan’s aging population has led to significant labor shortages, particularly in agriculture. An aging workforce and a lack of successors have exacerbated this issue. Additionally, agricultural processes, such as land preparation, seeding, and regular maintenance, require significant manual labor. This further contributes to the labor shortage. To address this issue, process automation through Information Technology (IT) solutions like “factory Automation” have been increasingly adopted in Japan. Automating simple, repetitive tasks can significantly reduce the need for human labor. Among IT solutions, Artificial Intelligence (AI) excels at automating tasks and is widely applied in labor-intensive industries. In this research, we focus on using an Object Detection model, specifically YOLO, to automate the inspection of harvested products for defects. This reduces labor demands in the agricultural sector. This study aims to improve YOLO’s detection accuracy to make it more effective in real-world applications. We propose enhancing the Loss function, a critical component in AI training. Specifically, we improve the L(obj) component, called SBCE, to account for the higher prevalence of small defects compared to larger ones. By scaling the Loss scores for small defects, our model becomes better at detecting small objects. As a result, our improved model demonstrates higher accuracy in object detection. We observe improvements in the mean Average Precision (mAP) of YOLOv7-tiny from 81.51% to 82.13%, as well as increases in Recall (from 68.97% to 72.55%) and mean Intersection over Union (mIoU) (from 52.91% to 53.48%) on the PASCAL VOC dataset.