A Feature Defusion Network for Surface Anomaly Detection of Industrial Products
Bing Hu, Jianhui Wang · 2023
In this paper, we propose a feature defusion network for industrial product surface anomaly detection. This method addresses two prominent issues in industrial product surface anomaly detection tasks: the presence of complex textures on the surface of the product to be inspected and the scarcity of anomaly samples. In our network, two variational autoencoders are used as the main branch network, and the similarity loss is applied to separate the texture features and structural features of the sample data in the latent space. Then, an external anomaly classifier is used to detect the structural features. We conducted a set of comparative experiments to validate the effectiveness of this method in anomaly detection tasks of this kind, and the results show that this method can meet the detection requirements of industrial production very well.