Balanced Fidelity Network for Blind Hyperspectral Unmixing

Hyperspectral image (HSI) unmixing is a popular field in HSI processing. Its main task is to decompose mixed pixels into pure spectral signatures called endmembers and their corresponding fractions called abundances . Some unmixing algorithms introduce subspaces and use regularizations to characterize the properties of subspaces to improve the estimated abundance accuracy. While we shift our focus to the balance of the original space and subspace information. We propose a balanced fidelity regularization to realize it. The subspace denoted by a projected matrix is learned through the proposed network without extra trial of designing specialized regularization. The proposed network consists of two parts. One is an autoencoder, estimating abundances, endmembers, and reconstructions, and the other is a convolutional layer projecting the reconstructions and the input HSI into a subspace. Specifically, the balanced fidelity term ensuring the consistence in both original space and the projected subspace to exploit and balance the information in two spaces. In addition, an adapted contraction regularization and a sparse regularization are incorporated to the loss function to further improve the unmixing performance. We conduct simulated and real‐data experiments to show that the proposed network greatly improves the estimation accuracy of endmembers and abundances compared with some state‐of‐the‐art algorithms.