One-Class Novelty Detection via Sparse Representation with Contrastive Deep Features

We propose a novel one-class novelty detection method by using sparse representation with contrastive deep features. To learn representative deep features from target one- class training data, we generate in-class and pseudo out-of-class training samples at first. Then, shared-weight networks with the contrastive loss are used to learn the contrastive deep features based on the generated training samples. We model the distribution of the learned features of the target class by using sparse dictionary learning. Testing samples are identified based on sparse reconstruction results by using the learned dictionary. The experimental results show the performance of the proposed method is comparable with state-of-the-art methods in the MNIST dataset and outperforms those methods in the CIFAR- 10 dataset.