Hyperspectral Anomaly Detection Based on Two-Stage Deep Transformation With Latent Features Guidance

Hyperspectral anomaly detection (HAD) has attracted widespread attention in the field of hyperspectral image processing. In recent years, model-driven and data-driven methods have been widely used in HAD. However, existing methods either do not fully utilize the various data characteristics of the background or directly use the various coupled data characteristics of the background in the image domain or a single deep transform domain, where the coupling characteristics in these domains are not obvious and difficult to separate, resulting in poor detection performance. To solve this problem, we innovatively proposed a two-stage deep transformation network with latent features guidance (TDTLFG) for HAD, aiming to fully utilize the potential data characteristics of the background tensor in different transform domains, and train the network under purposeful and directional guidance with that latent features guidance. Specifically, a multiscale feature extraction (MSFE) and a multiscale feature fusion (MSFF) two-stage deep transform blocks are designed to adapt to the anomaly scale-varying scenarios; the tensor matrix factorization is embedded into the network after the first MDFE deep transformation stage to guide the network to exploit the low rankness of the background in 3-D; the total variation (TV) constraint is designed on the output of the second MSFF transformation stage to exploit the piecewise-smoothness of the background in 3-D. After that, the fidelity constraints and TV constraints of the network output are iteratively solved by adopting the alternating direction method of multipliers (ADMM) algorithm. Finally, the detection performance effectiveness of the proposed model is proven on several real hyperspectral datasets.