基于改进U-Net的电子换向器表面缺陷分割方法  

Surface Defect Segmentation of Electronic Commutator Based on Improved U-Net

作　　者：黄紫彤 刘尧猛[1] 李瑞[1] HUANG Zitong;LIU Yaomeng;LI Rui(College of Artificial Intelligence,Tianjin University of Science&Technology,Tianjin 300457,China)

机构地区：[1]天津科技大学人工智能学院,天津300457

出　　处：《天津科技大学学报》2023年第4期41-47,共7页Journal of Tianjin University of Science & Technology

摘　　要：电子换向器的表面缺陷形状各异、缺陷与背景差异较小,同时还存在表面杂质干扰缺陷分割结果等问题,导致电子换向器缺陷难以精细分割。本文提出一种基于多尺度融合和残差分离卷积的改进U-Net缺陷分割方法。将不同尺度的图像输入编码模块便于网络模型提取多尺度下缺陷特征信息,并构建残差分离卷积模块,在增大感受野的同时保留细节特征。将多尺度的输出图像放大到相同尺度并融合作为最终输出,实现特征信息语义和位置的信息互补,从而提高网络的分割精度。在公开的KolektorSDD数据集上的实验结果表明,该方法的相似性系数和精确率分别达到97.3%与97.8%,缺陷分割效果相比于SegNet、FCN-8S等经典分割网络更加优秀,能够更加准确地识别细小缺陷。The surface defects of the electronic commutator are diverse in shape,with small differences between defects and backgrounds,and there are also such problems as surface impurities interfering with the detection results,which makes it difficult to finely segment commutator defects.To solve these problems,an improved U-Net defect segmentation method based on multi-scale fusion and residual separation convolution is proposed in this article.In the improved method,different scales of images are input into the coding module,which makes it convenient for the network model to extract defect feature information at multiple scales,and the residual separable convolution module is constructed to increase the receptive field while retaining detailed features.The output images of multiple scales are enlarged to the same scale and fused,thus realizing the complementary of feature information semantics and position,so as to improve the segmentation accuracy of the network.Experimental results on the published KolektorSDD dataset showed that the similarity coefficient and precision of the proposed method reached 97.3% and 97.8%,respectively,and the defect detection effect was better than that of classic segmentation networks such as SegNet and FCN-8S,and could identify small defects more accurately.

关 键 词：电子换向器 缺陷检测 多尺度融合 残差分离卷积 U-Net 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]