PR-YOLOv9: An improve defect detection network for hot-pressed light guide plates

Abstract

As one of the key components of liquid crystal display, the quality of the hot-pressed light guide plate (LGP) directly affects the display performance. To address the challenges posed by complex background textures, diverse types of defects, large variations in defect resolutions, and low contrast, this paper proposes a surface defect detection method for hot-pressed LGPs based on the PR-YOLOv9. The poly kernel inception network (PKINet) module is integrated by replacing the second convolution module of the YOLOv9 backbone network, effectively reducing interference from invalid targets such as complex textured backgrounds, thereby enhancing the network's ability to detect multi-scale defects and decreasing the network's parameters. Additionally, the receptive-field attention convolutional operation (RFAConv) module is incorporated, replacing the first and last layers of the YOLOv9 backbone network with this module. RFAConv module provides attention weights for large convolution kernels, effectively improving the network's ability to extract spatial feature information. Experimental results show that the proposed PR-YOLOv9 network achieves a mean average precision (mAP) of 98.40% and F1-Score of 97.14% on a self-constructed hot-pressed LGP defect dataset, with a reduction of 6.19 M in network parameters compared with YOLOv9, representing a decrease of 10.18%, making it suitable for real-time detection in industrial settings.