A optimized YOLO method for object detection

With the continuous breakthrough of object detection, there are many achievements in the combination field of computer science and mechanical automation. Especially in manufacturing industry, the object detection has played a great role in the weld inspection link during the production of vehicle wheels. The algorithm is used to align the weld on the wheel hub with the air sensor by identifying welds and rotate wheel hub in order to achieve the purpose of automatic inspection. The application of this technology can greatly Improve detection efficiency and accuracy, getting rid of the inconvenience caused by manual detection. There are few applications of object detection for wheel weld quality detection at present. The aim of this paper is to use improved YOLOv3 detector for vehicle wheel weld detecting applications to meet the needs of the industry. YOLOv3 as a one-stage deep learning-based approach, solves the shortcomings of traditional machine learning-based approach such as high time complexity and many redundant windows, which seriously affect the speed and performance of subsequent feature extraction and classification, since YOLO is one of the ends to end object detection algorithms. In this paper, we modify the original yolov3 model according to the actual situation of vehicle wheel weld. On the basis of the original model structure and layers, the paper uses Distance-IoU(DIoU) loss to improve the loss function of yolov3 and Non-maximum suppression using distance-IoU(DIoU-NMS) to eliminate the redundant candidate bounding boxes, which further accelerates the convergence speed of loss function and improves the accuracy of object detection(6.91%-point higher than base model in AP50 and 3.61%-point higher in AP75).