Abstract
The defect detection of printed circuit boards (PCBs) is a crucial step in the production process. Defects usually appear in small sizes, thus image data needs to be obtained using high-resolution camera equipment, which leads to the model requirement to process high-resolution images. In addition, there is an imbalance issue in defect categories. In order to have a more efficient and accurate PCB defect detection method in the production process,we propose the Residual Large Convolutional Network (ResLCNet) as the backbone of YOLOv5. This article draws inspiration from the application of Transformers in the field of vision and finds that using large convolution kernels instead of a bunch of traditional 3x3 small convolution kernels is a more powerful standard. In traditional concepts, large convolutions require a lot of computation and consume a lot of resources. However, through efficient convolutional network structure optimization methods, we can make large kernel convolutions more powerful, and at the same time, large kernel networks integrate some advantages of kernel methods and can adaptively learn feature representations of data, thereby shortening the distance between CNN and Transformers. After training and testing on the PCB defect dataset, our method achieved an average mAP value of 95.6% after multiple experiments, which is 1.2% higher than the original YOLOv5. The large kernel network proposed in this article reduces background interference to a certain extent and has high accuracy, providing ideas for using large convolutional kernels as the backbone to improve the accuracy of industrial defect detection in the future.Code & models at https://github.com/hjllovecv/Yolov5-ResCL