改进RT-DETR的液晶面板喷墨打印表面缺陷检测  

Improved RT-DETR for surface defect detection of LCD panels inkjet printing

作　　者：李昂[1] 刘竹丽[1] 宋伟 王立新[1] LI Ang;LIU Zhuli;SONG Wei;WANG Lixin(School of Mechanical and Power Engineering,Zhengzhou University,Zhengzhou 450001,China;Suzhou UB Precision Co.,Ltd.,Suzhou 215021,China)

机构地区：[1]郑州大学机械与动力工程学院,郑州450001 [2]苏州优备精密智能装备股份有限公司,江苏苏州215021

出　　处：《重庆理工大学学报（自然科学）》2024年第11期147-154,共8页Journal of Chongqing University of Technology：Natural Science

基　　金：科技部创新方法工作专项(2019IM010400-03)。

摘　　要：液晶面板喷墨打印表面缺陷检测中存在目标小、样本少、纹理背景干扰等问题,应用传统图像处理算法检测精度低、泛化性差,针对以上问题提出了一种改进RT-DETR(real-time detection transformer)的目标检测算法。改进RT-DETR算法通过将主干网络ResNet模型替换为特征提取性能更优的ConvNeXt模型,提高算法整体检测精度。设计了基于通道注意力的增强通道压缩模块,使算法更有效地消除背景干扰专注于定位缺陷目标,加快算法收敛,提高小目标检测精度。在构建的喷墨打印缺陷数据集训练实验上,改进RT-DETR算法检测平均精度mAP(mean average precision)为80.58%,较原始RT-DETR算法提升了2.89%,较原始DETR算法提升了15.88%,检测速度达到20 FPS(frames per second),改进RT-DETR算法的综合检测性能更优。改进RT-DETR算法在小目标检测数据集VisDrone训练实验上表现出良好的通用性,为其他工业场景下的表面小目标缺陷检测提供了参考价值。There are problems such as small objects,limited samples,and texture background interference in the detection of surface defects in inkjet printing of LCD panels.The existing image processing algorithms achieve low detection precision and poor generalization.To address these problems,we propose an improved object detection algorithm based on RT-DETR for real-time detection of surface defects in inkjet printing.By replacing the backbone network ResNet model with ConvNeXt model,which extracts features more efficiently,the improved algorithm acquires more effective features and improves the overall detection precision.An Enhanced Channel Squeeze Module(ECSM)based on channel attention is designed to make the algorithm more effective in eliminating the background interference and focusing on the localization of the defective objects,accelerating the convergence of the algorithm,and improving the precision of the detection of small objects.On the training experiments of the constructed inkjet printing defect dataset,the detection mean average precision(mAP)of the improved RT-DETR algorithm is 80.58%,which is 2.89%higher than that of the original RT-DETR algorithm and 15.88%higher than that of the original DETR algorithm.Meanwhile,the detection speed reaches 20 FPS(Frames Per Second).The overall detection performance of our RT-DETR algorithm is superior.On the small object detection dataset VisDrone,it exhibits extraordinary versatility and thus has great potentials for the defect detection of small objects in other industrial scenarios.

关 键 词：表面缺陷检测 目标检测 RT-DETR算法 ConvNeXt模型 通道注意力 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]