YOLO-FG: YOLO-Based Visual Inspection for Fruits Grading

Artificial intelligence (AI) is attracting attention because of its versatility due to research development. Progress of Advances in AI gives us the expectations to make our lives more comfortable. AI is beginning to be focused on in our personal and working lives. On the other hand, the labor shortage has recently become a serious problem in the Japanese industry. Automation of industries through AI and other IT technology is the most effective solution for this problem. Among the various industrial tasks, we focused on the inspection task because of its simplicity and the required time. However, AI has not penetrated into agriculture yet. One of the reasons is AI's model size is not appropriate to use in working. The higher detection score needs the bigger size of the model. Our research improves small AI to get higher detection score. Eventually, we aim to alleviate labor shortage problems by AI-based automatic product inspections. In this research, we develop an efficient object detection model, which is improved from YOLOX for the Visual Inspection of products. Our YOLO is optimized by focusing on the shape and size of a fruits grading. Due to our new proposal methods of convolution block(pELAN) and new IOU(eIOU), our model achieves a higher detection score than existing models. pELAN can supress to get excessive range information. It means that our model can get small defects information smoothly. eloU evaluates the center of the IoU by cutting off the circle from IoU. The above improvements improved the detection accuracy from 62.12% to 79.66% in comparison with the existing model.