| 基于对比学习生成对抗网络的无监督工业品表面异常检测 |
| |
| 引用本文： | 张 玥,陈锡伟,陈梦丹,周新龙,张宏伟.基于对比学习生成对抗网络的无监督工业品表面异常检测[J].电子测量与仪器学报,2023,37(10):193-201. |
| |
| 作者姓名： | 张 玥 陈锡伟 陈梦丹 周新龙 张宏伟 |
| |
| 作者单位： | 1. 西安工程大学电子信息学院 |
| |
| 基金项目： | 国家自然科学基金(61803292)； |
| |
| 摘 要： | 在工业品表面异常检测中，由于异常的未知性和无规则性，人工标注异常样本难度大、成本高，有监督的深度学习算法在工业品表面异常检测任务中存在局限性。针对上述问题，提出一种基于对比学习生成对抗网络(contrastive learning generative adversarial network, CLGAN)的无监督工业品表面异常检测算法。首先，建立基于无监督学习算法的CLGAN模型；其次，采用对比学习加强潜在特征空间正负例样本约束，使得输入与输出图像对应Patch之间的互信息最大化，增强正负样本特征向量区分度，使模型重构异常样本图像能力得到进一步提升；然后，在检测阶段，利用训练好的模型得到待测工业品的无异常重构图像，并计算得到待测样本与其对应重构图像之间的残差图像；最后，结合双阈值分割的后处理方法和数学形态学处理，实现工业品表面异常区域地快速检测和准确定位。通过在公共数据集MVTec AD上进行实验，与其他的无监督深度学习模型算法相比，所提算法具有更好的识别效果和更强的泛化能力。
|
| 关 键 词： | 异常检测 深度学习 无监督学习 对比学习 重构图像 特征空间 |
Unsupervised surface anomaly detection of industrial products based on contrastive learning generative adversarial network |
| |
| Zhang Yue,Chen Xiwei,Chen Mengdan,Zhou Xinlong,Zhang Hongwei.Unsupervised surface anomaly detection of industrial products based on contrastive learning generative adversarial network[J].Journal of Electronic Measurement and Instrument,2023,37(10):193-201. |
| |
| Authors: | Zhang Yue Chen Xiwei Chen Mengdan Zhou Xinlong Zhang Hongwei |
| |
| Abstract: | In the anomaly detection of industrial surfaces, due to the unknown and irregular nature of the abnormalities, it is difficult and
costly to manually label abnormal samples, and the supervised deep learning algorithms have limitations in the task of anomaly detection
on the surface of industrial products. To address the above problems, an unsupervised surface anomaly detection algorithm based on
contrastive learning generative adversarial network (CLGAN) is proposed. Firstly, the CLGAN model based on unsupervised learning
algorithm is established. Secondly, contrastive learning is used to strengthen the positive and negative sample constraints of the potential
feature space, maximizing the mutual information between the corresponding patches of the input and output images, enhancing the
differentiation of positive and negative sample feature vectors, and further improving the ability of the model to reconstruct abnormal
sample images. Then, in the detection stage, the trained model is used to obtain the anomaly-free reconstruction image of the industrial
product to be tested, and the residual image between the sample to be measured and its corresponding reconstructed image is calculated.
Finally, combined with the double threshold segmentation method and mathematical morphology processing, the rapid detection and
accurate location of abnormal areas on the surface of industrial products are realized. Experimental results on the public dataset MVTec
AD demonstrate that the proposed algorithm has a better recognition effect and stronger generalization ability compared with other
unsupervised deep learning model algorithms. |
| |
| Keywords: | anomaly detection deep learning unsupervised learning contrastive learning reconstructed image feature space |
|
| 点击此处可从《电子测量与仪器学报》浏览原始摘要信息 |
|
点击此处可从《电子测量与仪器学报》下载全文 |
|
