基于改进YOLOv8n的液晶屏Mura缺陷检测  

Mura defect detection of LCD screen based on improved YOLOv8n

作　　者：陈顺龙 廖映华 林峰 舒成业 CHEN Shunlong;LIAO Yinghua;LIN Feng;SHU Chengye(School of Mechanical Engineering,Sichuan University of Science&Engineering,Yinbin 644000,China;Sichuan Jinglong Optoelectronic Technology Co.Ltd.,Yinbin 644000,China)

机构地区：[1]四川轻化工大学机械工程学院,四川宜宾644000 [2]四川京龙光电科技有限公司,四川宜宾644000

出　　处：《液晶与显示》2025年第3期439-447,共9页Chinese Journal of Liquid Crystals and Displays

基　　金：宜宾三江新区“揭榜挂帅”科技项目(No.2022JBGS001);宜宾市引进高层次人才项目(No.2022YG01);四川省中央引导地方科技发展专项(No.2024ZYD0300)。

摘　　要：针对液晶屏Mura缺陷检测中因对比度低和尺度差异多样而导致的检测精度不足的问题,从提升模型对小尺度缺陷和微弱缺陷检测性能的角度,提出了一种基于改进YOLOv8n的液晶屏Mura缺陷检测模型YOLO-D3MNet。首先,通过引入ConvNeXtv2模块重构模型的主干网络和颈部网络,提高模型在复杂纹理背景下的微弱特征提取能力;其次,针对检测头模块特征信息跨通道交流不足的问题,提出了一种结合通道混洗策略和深度可分离卷积的高效解耦头,促进不同特征通道之间的信息流动,降低模型的算力需求;最后,针对基于预测框和真值框的交并比度量对小尺度缺陷的位置偏差敏感的问题,引入归一化高斯Wasserstein距离损失函数,提供更多的正样本候选框,从而提高模型对Mura缺陷的检测性能。改进后的YOLO-D3MNet模型的准确率、召回率和mAP_(50)分别为92.9%、88.8%和94.8%。相较于基础模型YOLOv8n,YOLO-D3MNet模型的准确率、召回率和mAP_(50)分别提高了3.4%、2.7%和3.6%,同时模型的GFLOPs降低了24.7%。与YOLOv5n等主流目标检测模型相比,本文提出的YOLO-D3MNet模型在液晶屏Mura缺陷检测方面具有更好的性能。To address the problem of insufficient accuracy in LCD Mura defect detection due to low contrast and diverse scale differences,from the perspective of improving the model’s performance in detecting small-scale defects and weak defects,an improved YOLOv8n-based LCD Mura defect detection model,YOLO-D3MNet,is proposed.Firstly,the backbone and neck networks of the model are reconstructed through the introduction of the ConvNeXtv2 module,which improves the weak feature extraction capability of the model under the background of complex texture.Secondly,for the problem of insufficient cross-channel communication of feature information in the detection head module,an efficient decoupling head combining the channel shuffle strategy and depth-separable convolution is proposed to promote the information flow between different feature channels and reduce the model computation power requirement.Finally,to address the problem that the intersection and concatenation ratio metric based on prediction box and truth box is sensitive to the positional bias of small-scale defects,the normalized Gaussian Wasserstein distance loss function is introduced to provide more positive sample candidate boxes,which improves the model’s detection performance of Mura defects.The precision,recall and mAP_(50)of the improved YOLO-D3MNet model are 92.9%,88.8%and 94.8%,respectively.Compared to the base model YOLOv8n,the precision,recall and mAP_(50)of the YOLO-D3MNet model are improved by 3.4%,2.7%and 3.6%,respectively,while the GFLOPs of the model are reduced by 24.7%.Compared with mainstream target detection models such as YOLOv5n,the experimental results show that the YOLO-D3MNet model proposed in this paper has better performance in LCD Mura defect detection.

关 键 词：Mura缺陷 液晶屏 目标检测 深度学习 微弱特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]