《计算机应用》唯一官方网站 ›› 2023, Vol. 43 ›› Issue (10): 3260-3266.DOI: 10.11772/j.issn.1001-9081.2022091452

• 多媒体计算与计算机仿真 • 上一篇    

WT-U-Net++:基于小波变换的表面缺陷检测网络

何国欢, 朱江平()   

  1. 四川大学 计算机学院,成都 610065
  • 收稿日期:2022-09-30 修回日期:2022-12-17 接受日期:2022-12-28 发布日期:2023-03-23 出版日期:2023-10-10
  • 通讯作者: 何国欢,朱江平
  • 作者简介:何国欢(1996—),男,陕西安康人,硕士研究生,主要研究方向:计算机视觉、缺陷检测;
  • 基金资助:
    四川省重点研发专项(2022YFG0053)

WT-U-Net++: surface defect detection network based on wavelet transform

Guohuan HE, Jiangping ZHU()   

  1. College of Computer Science,Sichuan University,Chengdu Sichuan 610065,China
  • Received:2022-09-30 Revised:2022-12-17 Accepted:2022-12-28 Online:2023-03-23 Published:2023-10-10
  • Contact: Guohuan HE, Jiangping ZHU
  • About author:ZHU Jiangping, born in 1984, Ph. D., associate professor. His research interests include computer vision, three dimensional reconstruction, defect detection.
  • Supported by:
    Key Research and Development Project of Sichuan Province(2022YFG0053)

摘要:

针对传统机器视觉算法在表面缺陷检测中精度低、无法适应环境变化和噪声影响的问题,提出一种基于小波变换(WT)的改进UNet++——WT-U-Net++。首先,由WT获取缺陷图像的高频与低频分量,再通过多尺度模块MCI(Mix-Conv Inception)提取高、低频分量的细节特征;其次,将MCI模块提取到的细节特征与原始图像融合,并将融合结果作为改进UNet++的输入;再次,在UNet++的下采样阶段引入通道注意力模块,从而使网络在捕获更多上下文语义信息的同时提高跨层特征级联的质量,而在上采样阶段采用反卷积恢复更多的缺陷细节信息;最后,从UNet++的多个输出中选择最佳结果作为检测结果。在铁轨、磁瓦、硅钢油污这3个公开缺陷数据集上的实验结果表明,相较于次优的UNet++,WT-U-Net++的交并比(IoU)分别提高了7.98%、4.63%和8.74%,相似度度量指标(DSC)分别提高了4.26%、2.99%和4.64%。

关键词: UNet++, 表面缺陷检测, 小波变换, 通道注意力, 反卷积

Abstract:

To address the problems of traditional machine vision algorithms such as low detection accuracy, inability to adapt to environmental changes and noise influence in surface defect detection, a improved UNet++ based on Wavelet Transform (WT) — WT-U-Net++ was proposed. Firstly, the high frequency and low frequency components of the defect image were obtained by the WT, and the detailed features of the high and low frequency components were extracted by the multi-scale module MCI (Mix-Conv Inception). Secondly, the detailed features extracted by MCI module were fused with the original image, and the fusion results were used as the input of the improved UNet++. Thirdly, in the downsampling stage of UNet++, channel attention module was introduced to enable the network to capture more contextual semantic information and improve the quality of cross-layer feature cascade at the same time. In the upsampling stage, deconvolution was adopted to recover more defect details. Finally, the best result was selected from the multiple output of UNet++ as the detection result. Experimental results on three public defect datasets of rail, magnetic tile and silicon steel oil stain show that compared with the sub-optimal algorithm UNet++, WT-U-Net ++ has the Intersection over Union (IoU) increased by 7.98%, 4.63%, and 8.74% respectively, and the Dice Similarity Coefficient (DSC) improved by 4.26%, 2.99% and 4.64% respectively.

Key words: UNet++, surface defect detection, Wavelet Transform (WT), channel attention, deconvolution

中图分类号: