可在TFT-LCD面板中实现多背景视觉细微缺陷检测的YOLO-DSM方法  

YOLO-DSM method for detecting multi-background visual micro-defects in TFT-LCD panels

作　　者：孔祥飞 王森[1] 赵林 陈明方[1] KONG Xiangfei;WANG Sen;ZHAO Lin;CHEN Mingfang(Faculty of Mechanical and Electrical Engineering,Kunming University of Science and Technology,Kunming 650500,China;Henan Tobacco Industry Company Limited,Zhengzhou 450000,China)

机构地区：[1]昆明理工大学机电工程学院,云南昆明650500 [2]河南中烟工业有限责任公司,河南郑州450000

出　　处：《中山大学学报(自然科学版)(中英文)》2025年第2期129-137,共9页Acta Scientiarum Naturalium Universitatis Sunyatseni

基　　金：国家自然科学基金(52065035)。

摘　　要：提出了一种基于YOLO-DSM深度学习图像检测模型。首先,在每个Dark模块后引入HMU模块,以提高TFT-LCD面板上目标缺陷的检测精度。将原始SPP替换为SSMA,使得网络更加关注背景低对比度目标。其次,引入DSM模块以帮助网络增强有用特征且抑制无用特征,增强语义信息的集成。最后,用ODConv模块替换原始网络的下采样卷积,细化局部特征映射,实现局部缺陷特征的充分提取。在自制的TFT-LCD缺陷数据集中,与当前较为先进的算法进行对比。结果表明,YOLO-DSM网络在mAP精度方面达到了97.40%,且FPS达到了77.42帧,可满足TFT-LCD缺陷任务检测要求。A deep learning image detection model based on You Only Look Once-Double Spatial-Squeeze Module(YOLO-DSM)is proposed.First,the Hierarchical Mixed-scale Unit(HMU)module is introduced after each Dark module to improve the detection accuracy of target defects on TFT-LCD panels.The original Spatial Pyramid Pooling(SPP)is replaced with Simple Spatial Mlp Attention(SSMA)to enable the network to focus more on targets with low contrast against the background.Second,the Double Spatial-Squeeze Module(DSM)is introduced to help the network enhance useful features and suppress useless ones,thereby enhancing the integration of semantic information.Finally,the Omni-dimensional Dynamic Convolution(ODConv)module replaces the down-sampling convolution of the original network to refine local feature mapping and achieve full extraction of local defect features.In comparative experiments on a self-made TFT-LCD defect dataset,the YOLO-DSM network achieved an mAP accuracy of 97.40%and an FPS of 77.42 frames.This meets the requirements of TFT-LCD defect detection tasks.

关 键 词：视觉细微缺陷 YOLO-DSM 全维动态卷积 SCSE注意力机制 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]