Mixture of experts with convolutional and variational autoencoders for anomaly detection

This study focused on the problem of anomaly detection (AD) by means of mixture-of-experts network. Most of the existing AD methods solely based on the reconstruction errors or latent representation using a single low-dimensional manifold are often not ideal for the image objects with complex background. However, modeling the data as a mixture of low-dimensional nonlinear manifolds is natural and promising for the classification of anomalies. In this study to realize the promise of multi-manifold latent information for AD, we propose a mixture of experts ensemble with two convolutional variational autoencoders (CVAEs) and convolution network (MEx-CVAEC) which explicitly learns manifold relationships of data that make use of multiple encoded detections. Additionally, we integrate a linear-based CAE as a gating network which optimizes the expert structures for efficient data characterization based on the manifold of the latent space. In the expert structure the data is re-encoded after each decoder to enhance the latent detection performance and the VAE is used as a core element in the encoder-decoder-encode (EDE) pipeline. To the best of our knowledge, this is the first study suggesting a mixture of CVAEs-based models for AD. The performance of the MEx-CVAE with EDE pipeline which we names as (MEx-CVAEC) compared over two basic MEx-CVAE model with ED pipeline based on logistic regression (MEx-L) and based on CAE (MEx-C) structures. In addition, the performance of the proposed model on three different datasets show the highest average AUC value than that of the state-of-the-art for image anomalies detection task.

Automated summary

This study proposes a method for anomaly detection using a mixture of CVAEs models, which learns the multi-manifold relationships of data through an ensemble of experts. The model shows superior performance on multiple datasets compared to existing methods for image anomaly detection tasks.