Learning to Diversify Patch-Based Priors for Remote Sensing Image Restoration

The restoration of clean remote sensing images from the degraded images is usually an ill-posed problem. A feasible solution for the problem is incorporating various priors into restoration procedure as constrained conditions. Recently, the patch-based Gaussian mixture model (GMM) priors became popular. However, the learning of patch-based GMM priors usually assumes the components of the GMM are independent and identically distributed (i.i.d.). The assumption means that two components could model the same training samples, and thus a possible decrease in model's description ability. This work presents a new learning method to diversify the patch-based GMM priors through modeling the GMM's parameters as a determinantal point process (DPP). The introduction of DPP allows us to specify a preference for diversity in patch-based GMM using a positive definite kernel function and thus to increase the model's description ability for a given model complexity. Furthermore, under the expected patch log-likelihood (EPLL) restoration framework, we further introduce the diversified patch-based GMM priors into remote sensing literature and develop new remote sensing image restoration algorithms. The extensive experiments for denoising and deblurring of remote sensing images demonstrated that the proposed algorithms produced better restoration results than other methods that already proven to perform well in the literature.