基于SE-RetinaNet的面向玻璃面板的小尺寸低显著性缺陷检测  

Defect Detection of Small Size Glass Panel Based on SE-RetinaNet

作　　者：王为 赵涛[1] 钟羽中 佃松宜[1] WANG Wei;ZHAO Tao;ZHONG Yuzhong;DIAN Songyi(College of Electrical Engineering,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学电气工程学院,成都610065

出　　处：《组合机床与自动化加工技术》2024年第7期123-127,131,共6页Modular Machine Tool & Automatic Manufacturing Technique

摘　　要：玻璃面板中的缺陷具有低显著、尺寸小、形态多样、数量少等特点,现有先进目标检测算法难以胜任玻璃面板的质检任务。基于此,提出了SE-RetinaNet—一种面向玻璃面板的小尺寸低显著性的缺陷检测算法。该算法在特征金字塔的顶层和底层引入了SE注意力机制和自注意力机制,增强网络对底层小尺寸特征的提取能力并强化顶层网络捕捉特征的长距离依赖关系的能力,同时在网络末端引入定位子网络SE-Regression,通过结合残差块和Inception模块的优点加强了定位的准确度同时防止网络退化。实验结果表明,所提算法能有效检测玻璃面板中各种尺寸的低显著性缺陷,其检测性能优于现有经典目标检测的算法,能够在玻璃面板缺陷检测问题上发挥较好的性能。The defects in glass panels have the characteristics of low salience,small size,diverse forms,and few numbers.The existing advanced target detection algorithms are difficult to be competent for the quality inspection task of glass panels.Based on this,this paper proposes SE-RetinaNet,a small-size and low-saliency defect detection algorithm for glass panels.The algorithm introduces attention mechanism and self-attention mechanism to the Feature pyramid at both the top and the bottom,strengthen the network′s ability to extract low-level features small size and the top-level network′s ability to capture the characteristics of long-distance dependencies,and at the same time the introduction of the positioning in the end of the network subnet SE-Regression,by combining the advantages of residual block and Inception module,the accuracy of positioning is strengthened and the network degradation is prevented.The experimental results show that the proposed algorithm can effectively detect the low saliency defects of various sizes in the glass panel,and its detection role is better than the existing classical target detection algorithms,which can play a good performance in the problem of glass panel defect detection.

关 键 词：小目标检测 玻璃面板缺陷检测 Focal loss SE注意力机制 自注意力机制 

分 类 号：TH16[机械工程—机械制造及自动化] TG659[金属学及工艺—金属切削加工及机床]