Convolutional autoencoder based on latent subspace projection for anomaly detection

Image anomaly detection (AD) is widely researched on many occasions in computer vision tasks. High-dimensional data, such as image data, with noise and complex background is still challenging to detect anomalies under the situation that imbalanced or incomplete data are available. Some deep learning methods can be trained in an unsupervised way and map the original input into low-dimensional manifolds to predict larger differences in anomalies according to normal ones by dimension reduction. However, training a single low-dimension latent space is limited to present the low-dimensional features due to the fact that the noise and irreverent features are mapped into this space, resulting in that the manifolds are not discriminative for detecting anomalies. To address this problem, a new autoencoder framework is proposed in this study with two trainable mutually orthogonal complementary subspaces in the latent space, by latent subspace projection (LSP) mechanism, which is named as LSP-CAE. Specifically, latent subspace projection is used to train the latent image subspace (LIS) and the latent kernel subspace (LKS) in the latent space of the autoencoder-like model respectively, which can enhance learning power of different features from the input instance. The features of normal data are projected into the latent image subspace, while the latent kernel subspace is trained to extract the irrelevant information according to normal features by end-to-end training. To verify the generality and effectiveness of the proposed method, we replace the convolutional network with the fully-connected network contucted in the real-world medical datasets. The anomaly score based on projection norms in two subspace is used to evaluate the anomalies in the testing. Consequently, our proposed method can achieve the best performance according to four public datasets in comparison of the state-of-the-art methods.