Unsupervised Anomaly Detection on Metal Surfaces Based on Frequency Domain Information Fusion

Metal products are widely used in industrial manufacturing, and the quality of metal products is becoming more and more demanding. At present, although there are many methods for detecting defects on metal surfaces, there are still various limitations. The limited number of defect samples, unpredictable defect characteristics, and the interference of metal grain bring great challenges to metal surface defect detection. For this reason, this paper proposes an unsupervised algorithm, FFnet, based on the fusion of frequency domain information, which introduces the frequency domain features into the unsupervised detection. A method of the adaptive enhancement of features in the frequency domain is proposed to make the features on the frequency domain more concerned with anomalies rather than textures. A scale-adaptive feature reconstruction module is used to effectively fuse the spatial and frequency domain features to fully utilize the information from different domains. In addition, a feature selection module is designed to improve the anomaly detection capability and reduce the computational redundancy by selecting the most representative subset of features. The proposed method outperforms other state-of-the-art methods on the connecting rod surface image dataset. In addition, in the generalization experiments of Kolektor Surface-Defect Dataset 2, our method also achieves optimal results and demonstrates strong generalization ability.