摘要
针对液晶屏显示缺陷中微弱特征缺陷经多次卷积与背景纹理同化导致的检测精度低的问题,提出了一种基于YOLOv5的液晶屏微弱特征缺陷检测改进模型YOLO-Mura。首先,在主干网络中引入Involution算子扩大感受野,增强在空间范围内的微弱特征缺陷信息,并降低模型的浮点运算次数。其次,采用CARAFE上采样算子优化上采样方式,加强对微弱特征缺陷的关注能力。然后,在颈部网络,通过嵌入BiFormer注意力模块,提升网络在强背景干扰下的特征提取能力。最后,采用BiFPN加权双向金字塔结构,提高不同层级的特征融合利用率。在自制液晶屏Mura缺陷数据集上的实验结果表明,YOLO-Mura模型的精确率、召回率、mAP@0.5分别提高了2.2%、6.6%、2.7%,模型计算量降低了66.5%。通过与主流目标检测算法进行比较,结果表明本文最终改进模型对于液晶屏微弱特征的Mura缺陷有较好的检测性能。
To address the problem of low detection accuracy of weak feature defects in LCD display defects caused by multiple convolution and background texture assimilation,an improved model YOLO-Mura for LCD weak feature defect detection based on YOLOv5 is proposed.Firstly,Involution operator is introduced in the backbone network to expand perceptual field,enhance the information of weak feature defects in spatial range,and reduce model FLOPs.Secondly,the CARAFE upsampling operator is used to optimize the upsampling method and enhance the ability to focus on weak feature defects.Then,in the neck network,the feature extraction ability of the network under strong background interference is enhanced by embedding the BiFormer attention module.Finally,the BiFPN weighted bidirectional pyramid structure is adopted to improve the feature fusion utilization at different levels.Experimental results on the homemade LCD Mura defect dataset show that the accuracy,recall,and mAP@0.5 of YOLO-Mura model are improved by 2.2%,6.6%,and 2.7%,respectively,and the model computation is reduced by 66.5%.In comparison with the mainstream target detection algorithms,the results show that the final improved model in this paper has better detection performance for Mura defects with weak features of LCDs.
作者
林峰
石艳
陈顺龙
廖映华
赵练
赵黎
周泽民
LIN Feng;SHI Yan;CHEN Shunlong;LIAO Yinghua;ZHAO Lian;ZHAO Li;ZHOU Zemin(School of Mechanical Engineering,Sichuan University of Science&Engineering,Yibin 644000,China)
出处
《液晶与显示》
CAS
CSCD
北大核心
2024年第6期790-800,共11页
Chinese Journal of Liquid Crystals and Displays
基金
宜宾市科技厅重点研发项目(No.2021GY0011)
宜宾市三江新区揭榜挂帅项目(No.2022JBGS001)
四川轻化工大学研究生创新基金(No.Y2022056)。
