摘要
针对领域偏移环境下金属表面缺陷检测精度不高的问题,提出一种基于改进对抗性域分离与自适应(Improved adversarial domain separation and adaptation,IADSA)深度迁移网络的金属表面缺陷检测方法。首先,建立基于分类损失的IADSA模型性能评价机制,感知模型训练状态,利用空间线性插值方法自适应挖掘迁移空间隐藏的样本信息,以提升网络的特征提取能力;然后,将挖掘的新样本的分类结果作为衡量其对网络贡献性能的主要度量指标,并将贡献性能作为权重应用在分类损失上,旨在消除噪声样本对模型造成的影响;在对抗训练过程中,通过添加动态权重优化对抗损失、平滑网络参数,提高模型的判别性能;最后,融合原始样本与新样本的任务分类器、域分离鉴别器以及域适应鉴别器的损失,利用动态训练实现金属表面缺陷检测。试验结果表明,与其他无监督域自适应方法相比,所提方法实现了更高的金属表面缺陷检测精度。
Aiming at the problem of low detection accuracy of metal surface defects in the domain shift environment,a metal surface defect detection method based on improved adversarial domain separation and adaptation(IADSA)deep transfer network was proposed.First,the performance evaluation mechanism of IADSA model based on classification loss is established to perceive the training status of the model.The spatial linear interpolation method is proposed to adaptively mine the sample information hidden in the migration space to improve the feature extraction ability of the network.Then,the classification result of new samples is used as the main measurement index to measure the performance of its contribution to the network,and the contribution performance is applied as a weight to the classification loss,which aims to eliminate the influence of noise samples on the model.Finally,dynamic weights are added to optimize adversarial loss and smooth network parameters in the process of adversarial training to improve the discrimination performance of the model.The experimental results show that the proposed method achieves higher detection accuracy of metal surface defects in an unsupervised environment.
作者
宿磊
王立建
祁阳
张思雨
顾杰斐
李可
SU Lei;WANG Lijian;QI Yang;ZHANG Siyu;GU Jiefei;LI Ke(School of Mechanical Engineering,Jiangnan University,Wuxi 214122;The 58 th Research Institute of China Electronics Technology Group Corporation,Wuxi 214000)
出处
《机械工程学报》
EI
CAS
CSCD
北大核心
2023年第24期46-55,共10页
Journal of Mechanical Engineering
基金
国家自然科学基金资助项目(52175096,11902124)。
关键词
金属表面缺陷检测
无监督域自适应
域对抗迁移
样本挖掘
动态加权
metal surface defect detection
unsupervised domain adaptation
domain adversarial migration
sample mining
dynamic weighting
