FEGAN: A Feature Extraction Based Approach for GAN Anomaly Detection and Localization

Abstract

With the continuous advancement of industry 4.0 and intelligent manufacturing, there is a growing need for automation and intelligence in industrial production processes. Anomaly detection of industrial product surface images is a key technology in achieving this goal and has thus become a significant area of research. However, this endeavor still encounters challenges such as scarcity of abnormal samples, complexities in data labeling, and uncertainties stemming from unknown factors and randomness. To this end, we propose the Feature Extraction-based for Generative Adversarial Network (FEGAN), aimed at detecting and precisely localizing surface anomaly in industrial products. FEGAN focuses on the deep features of an image, and it builds a feature extraction network and an improved generative adversarial network based on VGG19, respectively. We also jointly determines the anomaly score through the deep feature space as well as the Euclidean distance in 2D image space to better identify and locate the anomaly. Furthermore, we introduce a novel Multi-scale Self-Enhancement (M-SE) strategy to bolster the model’s generalization capabilities. We conducted training and testing on the MVTec and Bottle-Cap public datasets. A plethora of experimental results indicate that the proposed method outperforms existing methods significantly in terms of anomaly detection. Additionally, through an evaluation of the model’s localization accuracy, we demonstrate that FEGAN exhibits certain competitive advantages.

Keywords

• Feature extraction
• generative adversarial networks
• surface anomaly