Unsupervised Anomaly Detection Based on Block Pyramid Memory Module

Abstract

Unsupervised anomaly detection based on reconstruction is widely used because it does not require additional anomaly samples and pretrained models during training. However, in practical applications, the model can effectively reconstruct anomalies because of the generalization of convolutional neural networks, making it difficult to detect anomalies based on reconstruction errors. Existing methods suppress abnormal reconstruction by using appropriate memory blocks to store normal data and transform abnormal features into normal features. However, different abnormal areas can vary significantly, and improper selection of the memory block size may result in problems of fuzzy and abnormal reconstruction. Considering the advantages of these methods in the reconstruction model, an unsupervised anomaly detection method based on improved memory block storage is proposed. This involved adding a block pyramid memory module to adapt to anomalies of different area sizes. Block memory modules of different scales are employed to obtain multiple feature maps through reading and aggregation and subsequently obtained the output feature map after fusion. This approach can retain the feature information of normal samples to the maximum extent and enhance the storage and expression of feature information to better reconstruct normal data. Simultaneously, to enhance the clarity of the reconstruction and reduce reconstruction anomalies, a skip connection structure is added to the reconstruction network.Finally, the SSIM loss function is introduced to enhance the image reconstruction effect through the three dimensions of brightness, contrast, and structure. It also served as a component of the anomaly determination index, improving the accuracy of anomaly detection. The experimental results indicate that this method achieves an average AUC that is 1.5% higher than that of the original reconstructed model based on memory block storage and reading and has a better detection effect.