点击次数：

DOI码：10.1049/cvi2.12297

发表刊物：IET Computer Vision

关键字：computer vision, feature extraction, image recognition, image reconstruction, vision defects

摘要：Anomaly Detection, also known as outlier detection, is critical in domains such asnetwork security, intrusion detection, and fraud detection. One popular approach toanomaly detection is using autoencoders, which are trained to reconstruct input byminimising reconstruction error with the neural network. However, these methodsusually suffer from the trade‐off between normal reconstruction fidelity and abnormalreconstruction distinguishability, which damages the performance. The authors find thatthe above trade‐off can be better mitigated by imposing constraints on the latent space ofimages. To this end, the authors propose a new Dual Adversarial Network (DualAD) thatconsists of a Feature Constraint (FC) module and a reconstruction module. The methodincorporates the FC module during the reconstruction training process to impose con-straints on the latent space of images, thereby yielding feature representations moreconducive to anomaly detection. Additionally, the authors employ dual adversariallearning to model the distribution of normal data. On the one hand, adversarial learningwas implemented during the reconstruction process to obtain higher‐quality recon-struction samples, thereby preventing the effects of blurred image reconstructions onmodel performance. On the other hand, the authors utilise adversarial training of the FCmodule and the reconstruction module to achieve superior feature representation, makinganomalies more distinguishable at the feature level. During the inference phase, the au-thors perform anomaly detection simultaneously in the pixel and latent spaces to identifyabnormal patterns more comprehensively. Experiments on three data sets CIFAR10,MNIST, and FashionMNIST demonstrate the validity of the authors’ work. Results showthat constraints on the latent space and adversarial learning can improve detectionperformance.

论文类型：期刊论文

学科门类：工学

文献类型：J

卷号：Volume18,

期号：Issue8

页面范围：1138 - 1148

字数：6000

ISSN号：1751-9632

是否译文：否

发表时间：2024-06-20

收录刊物：SCI

通讯作者：冯爱民