loading page

  • Xuan Huang
Xuan Huang
Chengdu College of University of Electronic Science and Technology of China

Corresponding Author:[email protected]

Author Profile


In the realm of computer vision, image anomaly detection has always been an important research direction. Despite various anomaly detection approaches based on deep neural networks have been developed successively, the progress of end-to-end unsupervised image anomaly detection is still limited. This paper proposes a novel unsupervised anomaly detection automatic adaptive dynamic clustering model named UAD-ADC that can performs end-to-end outlier detection on images without available labels. Specifically, UAD-ADC employs a deep auto-encoder to learn the feature representation of images and combine with data augmentation to improve the quality of learned features. To detect abnormal images, we propose an effective surrogate supervision method and construct a discriminative model based on a deep clustering method. In particular, an iterative optimization method combined with self-paced learning is adopted to select reliability samples to train the discriminant model. Finally, an efficient anomaly scoring strategy is designed to evaluate images to improve anomaly detection performance. Extensive experiments under different outlier ratios on image benchmarks demonstrate that UAD-ADC significantly improves anomaly detection performance compared to the state-of-the-art unsupervised anomaly detection methods.