Detection method of wheel hub weld defects based on the improved YOLOv3

This study proposes a method for intelligent detection of wheel weld defects, against manual detection, by improving an existing deep learning target detection algorithm, called "You only look once" version 3 (YOLOv3). The improved algorithm is called YOLOv3-MC. First, an industrial camera was used to capture the images of the wheel hub weld defects, which were then annotated and developed into a data set. The data set was then expanded using a data enhancement method. Second, the detection accuracy of the algorithm was improved using the Mish activation function instead of the Leaky ReLU activation function in the YOLOv3 backbone network. Furthermore, the loss function of the algorithm was modified, and the positioning accuracy of the detection algorithm was improved using the method of complete intersection over union (CIoU). Finally, the batter model was trained with a training set. The detection experiment was implemented using a validation set and test set. The experimental results yielded a mean average precision (mAP) of 98.94% for the validation set in the optimal model of YOLOv3-MC. The F1 score value of the model was 0.99; the average Intersection over Union (AvgIoU) of the model was 80.92%; the detection speed was 76.59 frames per second (fps); the model size was 234 MB; and the detection accuracy rate of the optimal model on the test set reached 99.29%. Compared to the traditional machine vision detection method, this method offers an improved detection accuracy and meets the real-time online detection needs of the welding seam during wheel manufacturing. © 2021, Science Press. All right reserved.