HM-YOLOv5: A fast and accurate network for defect detection of hot-pressed light guide plates

Due to the complex texture background, uneven brightness, various defect sizes and types of hot-pressed light guide plate (LGP) images, the HM-YOLOv5 network for fast and accurate defect detection of hot-pressed LGPs is developed and evaluated in this paper. First, a hybrid attention module (HAM) is constructed by combining the spatial attention of the convolutional block attention module(CBAM) with the channel attention of the efficient channel attention network(ECA-Net), and is introduced behind each C3 module of the backbone network. HAM does not require dimensionality reduction processing, and can better fuse channel and spatial information to focus on defects targets. Therefore, the HAM can enhance the defect feature extraction ability of the network. Second, a multi-expansion convolution module (MCM) is constructed based on the pyramid structure of YOLOv5. MCM can improve the target receptive field, enrich contextual information, reduce loss of information during downsampling and improve the defect detection ability Finally, a self-built dataset is constructed by using the images of hot-pressed LGPs collected from industrial sites, and many experiments are performed using this database. Experimental results show that the mean average precision (mAP) of the network is 98.9%, especially for white point and dark line defects improved by 2.7% and 2.0% respectively, and that the detection speed can reach 417 frames per second(Fps). Compared with the mainstream surface defect detection algorithm, the accuracy and detection time are significantly improved. Moreover, the accuracy and real-time performance meet the industrial detection requirements.

Automated summary

In this paper, the HM-YOLOv5 network is developed and evaluated for fast and accurate defect detection of hot-pressed LGPs. The network utilizes a hybrid attention module and a multi-expansion convolution module to enhance the defect feature extraction ability and improve the detection performance. Experimental results show significant improvements in accuracy and detection speed.