Improved YOLOv5 with BiFPN on PCB Defect Detection

As an image classification technology, target detection also needs to identify specific locations of predefined categories. Therefore, target detection is not only to solve the problem of recognizing what objects exactly, but also points out locations. The development of YOLO series has greatly improved speed and accuracy in target detection technology. However, it performs not so good as detecting normal objects when targets are quite small. Thus, this paper proposes to fuse BIFPN network and set smaller and denser anchors to improve this, and then public PCB defect detection data set is used to test and verify effect of refined method. Experimental results show that improved method shifts mAP@0.5 from 0.968 to 0.979 and mAP@0.5:0.95 from 0.494 to 0.501, as well as deducts confusion rate. With the improved method, the training cost also decrases significantly.

I. Introduction

Object detection is one of the basic tasks in the field of computer vision, and the academic field has a research history of nearly two decades. With rapid development, algorithms based on deep learning emerged, such as R-CNN, Fast/Faster R-CNN, SSD and YOLO series which have better effect on target detection than traditional manual features way[1]. To improve confidence of detection, OverFeat feature extractor and multiscale sliding window in ConvNet[2] was proposed in 2013. Later, region average pooling was used on Fast R-CNN to better detect regional context[3] and Learning RoI Transformer was to Improve dense target detection without adding anchors [4].