双注意力驱动的微小缺陷识别方法研究  

Microdefect recognition method driven by dual attention mechanism

作　　者：邹林丰 邓耀华[1] 陈冠浩 张紫琳 ZOU Linfeng;DENG Yaohua;CHEN Guanhao;ZHANG Zilin(School of Electromechanical Engineering,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学机电工程学院,广东广州510006

出　　处：《中国测试》2025年第3期162-169,共8页China Measurement & Test

基　　金：国家自然科学基金(52175457);广东省基础与应用基础研究基金(2022B1515120053)。

摘　　要：针对深度卷积提取过程中微小缺陷特征消失问题,该文提出融合双注意力机制和跃进残差结构的微小缺陷识别深度卷积网络模型,该模型在训练过程中分别在通道维度和空间维度将权重更多地偏向目标特征,更多地关注到微小缺陷特征,抑制冗余缺陷特征;同时为了进一步缓解深度卷积中微小缺陷特征消失的问题,设计跃进残差结构通过少量的支路连接将微小缺陷特征传递到深层网络,既减少微小缺陷特征漏检,同时提高支路卷积计算速度。以实际采集的布匹缺陷数据集开展模型测试实验。该文提出的模型相比于ResNet50、ResNet101,微小缺陷的识别率分别提高6.79%和6.88%,证明该文模型在微小缺陷识别任务中的有效性。To solve the problem of the disappearance of micro defect features during depth convolution extraction,this paper proposes a micro defect recognition deep convolution network model that integrates a dual attention mechanism and a jump forward residual structure.During training,the model focuses more on target features in both the channel and spatial dimensions,paying greater attention to micro defect features while suppressing redundant ones.Meanwhile,to further alleviate the disappearance of micro defect features in deep convolution,the jump forward residual structure transfers these features to the deep network through a small number of branch connections.This reduces the omission of micro defect features while improving the speed of branch convolution calculations.The model was tested using an actual cloth defect dataset.Compared with ResNet50 and ResNet101,the proposed model improves the recognition rate by 6.79%and 6.88%,respectively,demonstrating its effectiveness in micro defect recognition tasks.

关 键 词：微小缺陷识别 双注意力机制 残差网络 深度卷积神经网络 

分 类 号：TB9[一般工业技术—计量学] U4772.9[机械工程—测试计量技术及仪器]