Dual feature extraction network for hyperspectral image analysis

Highlights：

• The proposed DFEN is the first method to jointly train two AEs in the HAD task. One AE focuses on mining the latent features in the original spectral data, and the other AE is used to learn latent features in the background spectral data.

• We present an end-to-end discriminative learning loss between dual networks to make background uniform and anomaly prominent. Especially, adversarial learning and Gaussian constraint learning are imposed on the deep latent space to extract more discriminative features.

• The orthogonal projection divergence (OPD) spectral distance between the two latent spaces is combined with the pixel-level differences, i.e., mean squared error (MSE), to obtain the comprehensive detection results. The experiments on eight real HSIs illustrate that our DFEN-based HAD is capable of offering competitive detection results, particularly in reducing the false alarm rate.