No abstract

A crucial component of quality control during printed circuit board (PCB) production is defect detection. The PCB should be inspected during the manufacturing process in order to minimize defects such as printing errors, incorrect component selections, and incorrect soldering. Convolutional neural networks (CNNs) have become widely used due to their high recognition power. One of the factors affecting the accuracy of CNNs is loss function. Intersection over union (IoU) based loss function and its variants such as IoU, generalized IoU (GIoU), distance IoU (DIoU), complete IoU (CIoU), and improved CIoU (ICIoU) are common metrics for bounding box regression. The IoU-based loss functions, such as ICIOU, achieve remarkable success but still have some main drawbacks such as inaccurate regression. The ICIoU performance is degraded under conditions between the ratio of the corresponding height or width of the predicated bounding boxes and the ground truth bounding box. In this paper, an improved IoU called VIoU is introduced to solve this problem. By incorporating VIoU loss into state-of-the-art YOLOv4 object detection algorithm, an average accuracy of 98.63% on the PCB defect detection is achieved which is an improvement compared to existing IoU models.