Discriminative Reconstruction Constrained Generative Adversarial Network for Hyperspectral Anomaly Detection

Paper link: https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8972475

The rich and distinguishable spectral information in hyperspectral images (HSIs) makes it possible to capture anomalous samples [i.e., anomaly detection (AD)] that deviate from background samples. However, hyperspectral anomaly detection (HAD) faces various challenges due to high dimensionality, redundant information, and unlabeled and limited samples.

HYPERSPECTRAL data are regarded as a 3-D cube data, where the two dimensions provide spatial information and another dimension describes abundant spectral information in hundreds of approximately continuous spectral bands.

Fortunately, anomalies have two important characteristics that could be used to distinguish from the background, i.e., compared to their surroundings: 1) the spectral signatures of anomalies are quite distinguishable in the spectral domain, and 2) anomalies usually appear with a relatively small area. Anomalies can be detected by referring to a background model or suppressing the background.

Proposed Method: HADGAN method. First, aiming at generating a reconstruction of an HSI’s variability with anomaly targets suppressed, an AE network is constructed on Y, yielding an encoder E and a decoder De. Then, E and De are trained in an adversarial way and some powerful constraints are imposed to learn a discriminative reconstruction background Y˜ on the AE model. Meanwhile, the input Y and the reconstructed image Y˜ are simultaneously fed into the adversarial image discriminator to correct the generated Y˜ , the degree of deviation Y is used as input for subsequent detection. Finally, a joint spatial and spectral method is employed on Y to achieve final AD.