视觉注意力网络在工件表面缺陷检测中的应用

工件表面缺陷检测是现代化工业生产中不可缺少的环节,利用卷积神经网络实现工件表面缺陷检测能有效地提升检测效果.当工件表面出现微小缺陷时,缺陷部分的特征容易被其他区域的特征所掩盖,影响检测的准确率.针对这一问题,提出了每级由3个卷积模块和一个视觉注意力模块构成的3级视觉注意力网络.通过注意力模块生成软注意力模板,为卷积模块构成的主干网络的特征图加权,增强缺陷区域特征并抑制背景区域特征,提升缺陷检测的准确率.实验采用具有明显缺陷和微小缺陷的5类工件图像进行对比测试,结果表明,软注意力模板在容易出现缺陷的区域具有更高的权值;加入视觉注意力模块能将缺陷检测的准确率从90.9%提升至98.1%.

Application of Visual Attention Network in Workpiece Surface Defect Detection

Workpiece surface defect detection is an indispensable part of modern industrial production and the effect of surface defect detection can be improved effectively by convolutional neural network.When the surface has tiny defects,the features of the defective part are easily covered by the features of other areas,thus affecting the accuracy of detection.In this paper,a three-level attention network is proposed,in which each level network consists of three convolution modules and one visual attention module.The soft attention template from the attention module is used as the weight of the feature map of the backbone network with convolution modules to enhance features of defect regions and suppresses background region features,and improves the accuracy of defect detection.Five kinds of workpiece images with obvious defects and minor defects are tested in the experiment.The experimental results show that the soft attention templates have higher weights in areas prone to defects and the accuracy of defect detection is improved from 90.9%to 98.1%by adding visual attention module.