Improved SVDD-based anomaly detection models

In this study, an SVDD (Support Vector Data Description) network is employed for unsupervised learning on industrial products. The feature extraction network is improved by incorporating frequency-domain information analysis and differential convolution to enhance image quality, thereby boosting the model's feature extraction accuracy. Meanwhile, a contrastive network is introduced to enable efficient utilization of images, addressing the challenge of few-shot learning. Ultimately, the proposed model achieves an accuracy of 87.6% on the industrial MVTec dataset when trained with only 100 normal samples.