A scale adaptation autoencoder model for anomaly detection

The emergence of generative adversarial networks provides a new approach for anomaly detection, which alleviates the shortage of anomaly sample data in anomaly detection scenarios to a certain extent. However, in some scenarios, the distribution of data samples is diverse, complex, and the features are not clear enough. For auto-encoder networks, sample reconstruction is still difficult. In response to the problem of insufficient semantic expression ability of autoencoders, this paper proposes an anomaly detection model based on scale adaptive autoencoder. Based on deep feature decoding at two scales, scale adaptive autoencoders can automatically adapt neural network structures through backpropagation. This model promotes the feature representation of neural networks, avoids overfitting to a certain extent, and achieves accurate recognition of abnormal data. This article conducted comparative experiments on the CIFAR10 anomaly detection dataset, demonstrating the effectiveness of this method in unsupervised anomaly detection tasks.