Light-YOLOv5: A Lightweight Algorithm for Improved YOLOv5 in PCB Defect Detection
Meng Ye, Hao Wang, Hang Xiao
Abstract
When the object detection algorithm faces the chanllenge of PCB defect detection, we need to consider the speed and accuracy of detection under the condition of limited computing resources. This paper provides an improved object detection algorithm to enhance the detection efficiency of small objects with defects and achieves efficient model lightweighting with little loss of detection accuracy. First, GSConv and VoV-GSCSP networks are introduced to replace the original Conv and C3 network modules in the feature extraction part and add an attention mechanism module to enhance the dimension of information and reduce redundant information candidate boxes. The attention module of the Shuffle backbone network is used to extract information effectively without losing too much information. The number of invalid extracted features is reduced by improving the feature extraction capability of the backbone network and the effectiveness of feature fusion. The final experimental results show that the improved lightweight model in this paper reduces the weight of the yolov5s6 model size by 51% on the PCB defect detection dataset published by Peking University. Our proposed method can improve the model's mAP.5 by 2-3%. For defect detection, the small object recognition effect is significantly improved, and the lightweight effect is remarkable.