Defect detection of light guide plate based on improved YOLOv5 networks

Abstract:

Light guide plate (LGP) is a kind of material used in the backlight module. How to improve the quality control of LGP has become the focus of research in the industry. To address issues such as low gray contrast and a high proportion of small target defects in LGP images, an improved you only look once version 5 (YOLOv5) neural network based on multi-scale dilation convolution and a novel loss function is proposed. First, the LGP image is preprocessed, and then the context amplification module (CAM) is integrated into the feature fusion part of the detection algorithm to fuse multi-scale expansion convolution features to obtain rich context information. The extended intersection over union (XIoU) is selected as the location regression loss function. The results show that this method can effectively improve the detection accuracy and positioning accuracy. Compared with YOLOv5, the proposed method achieves an average accuracy increase of 4.7% and a recall rate increase of 2.7%. It can achieve accurate detection of defects, such as white/bright spots, black spots, line scratches, and surface foreign objects in LGP.