计量学报2025，Vol.46Issue(9)：1377-1384,8.DOI:10.3969/j.issn.1000-1158.2025.09.19
基于改进DCGAN网络的钢材表面缺陷图像数据增强方法
Steel Surface Defect Image Data Enhancement Method Based on Improved DCGAN Network
摘要
Abstract
In order to solve the problems of poor quality and unstable network training of traditional deep convolutional generative adversal network(DCGAN)in generating steel surface defect images,an improved DCGAN steel surface defect image data enhancement model is proposed.Firstly,the residual module and Self-attention mechanism are added to DCGAN network to improve the feature extraction capability of defect images.Secondly,spectral normalization and Wasserstein distance loss function with gradient penalty term are introduced to improve the stability of network training.Finally,the proposed model is tested on the NEU-DET dataset.The test results show that the proposed model can stably generate high-quality steel surface defect images.Compared with the original DCGAN network,the recognition accuracy of images generated by the proposed model on AlexNet network is increased by 6.8%,the FID value is decreased by 61.3%,and the IS value is increased by 20.8%.The image quality is better than GAN,DCGAN,CGAN and WGAN,which can be used as an effective strategy to solve the problem of small samples of steel surface defect images.关键词
几何量计量/钢材表面缺陷/生成对抗网络/图像生成/残差模块/自注意力机制/谱归一化Key words
geometric measurement/steel surface defect/generate adversarial network/image generation/residual module/self-attention mechanism/spectral normalization分类
通用工业技术引用本文复制引用
时培明,申世春,叶蕾,许学方,阚俊明,韩东颖..基于改进DCGAN网络的钢材表面缺陷图像数据增强方法[J].计量学报,2025,46(9):1377-1384,8.基金项目
河北省自然科学基金(F2024203035,E2022203093) （F2024203035,E2022203093）
复杂零部件智能检测与识别湖北省工程研究中心开放课题(IDICP-KF-2024-10) （IDICP-KF-2024-10）
