Deep Unrolling Network with Active Dictionary Learning for Hyperspectral Anomaly Detection

Deep learning-based anomaly detectors are receiving increasing attention, with the black box represented by neural networks being an ever-present puzzle. Recently, an interpretable deep network (LRR-Net) –the first model-driven deep network in the field of hyperspectral anomaly detection–provided an inspiring endeavor. However, this work is incomplete, as it separates dictionary construction from anomaly detection, leading to a mismatch problem. To address this issue, this paper proposes a deep unrolling network with active dictionary learning, aiming to integrate dictionary construction into the network rather than two separate phases (one is a traditional method, and another is a deep learning method). Specifically, the approach incorporates the concept factorization (CF) theory into the design of an anomaly detection model based on low-rank representation, i.e., the dictionary is regarded as a linear combination of the original data and the coefficient matrix is constrained to be orthogonal. Furthermore, a deep unrolling network is designed to simulate the dictionary learning and variables updating process, achieving dictionary construction and anomaly detection. simultaneously. Experimental results on two datasets demonstrate that the proposed deep unrolling network achieves superior performance than other state-of-the-art methods.