Title: Skip.Self attention GAN for anomaly detection

Authors: Hanlin Sun; Yefeng Liu; Yihang Ma; ChongZheng Na; Qichun Zhang

Addresses: School of Automation and Electrical Engineering, Shenyang Ligong University, Shenyang 110159, China ' School of Automation and Electrical Engineering, Shenyang Ligong University, Shenyang 110159, China; Liaoning Key Laboratory of CNC Machine Tool Cyber-Physical Fusion and Intelligent Manufacturing, Shenyang Institute of Technology, Shenfu Demonstration Zone, Liaoning 113122, China ' Liaoning Key Laboratory of CNC Machine Tool Cyber-Physical Fusion and Intelligent Manufacturing, Shenyang Institute of Technology, Shenfu Demonstration Zone, Liaoning 113122, China ' Liaoning Key Laboratory of CNC Machine Tool Cyber-Physical Fusion and Intelligent Manufacturing, Shenyang Institute of Technology, Shenfu Demonstration Zone, Liaoning 113122, China ' School of Creative and Digital Industries, Buckinghamshire New University, High Wycombe, HP11 2JZ, UK

Abstract: Generative adversarial network (GAN) has low discrimination ability for abnormal samples or noise data in anomaly detection tasks, which affects the detection effect. This paper proposes an improved model based on Skip-GANomaly network, which adds a self-attention module to the encoder of the generator to improve the model's ability to capture long-distance dependence of input data, so as to better understand the overall structure and local details of the image. Moreover, it dynamically focuses on the most important parts of the anomaly detection task, such as the shape, texture and colour of the object, so that the data representation ability and anomaly detection accuracy of the model are significantly improved. MvTec public dataset and KolektorSDD dataset were used for experimental verification. The results show that compared with AnoGAN, GANomaly, and Skip-GANomaly, it has better performance in terms of receiver operating characteristic (ROC) area under the curve (AUC), which proves the effectiveness of this model and has a certain application prospect.

Keywords: generative adversarial network; GAN; anomaly detection; Skip-GANomaly; self-attention.

DOI: 10.1504/IJSCC.2025.144544

International Journal of Systems, Control and Communications, 2025 Vol.16 No.1, pp.17 - 32

Received: 02 Oct 2024
Accepted: 18 Nov 2024

Published online: 18 Feb 2025 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article