Abstract
A novel framework for pansharpening based on Latent Low-Rank Representation theory, called Detail injection using Latent Low-Rank decomposition based Pansharpening approach (DiLRP), is proposed. Our proposal comprises two fusion stages. In the first step, a primary joint fusion scheme is defined as a combination of low-rank and saliency images, which is used further for extracting spatial details. In this model, the histogram-matched PAN and up-sampled images are decomposed into low-rank and saliency components, and the corresponding fusion strategies are designed according to their characteristics. Indeed, to preserve more global structural information, low-rank components are combined by an optimized weighted average fusion strategy to generate a low-rank image. Furthermore, the saliency image is obtained by a simple average fusion rule in order get high-frequencies details (i.e., edges) that are highly relevant to MS image. The second stage consists in injecting (globally or locally) the high-frequency details, extracted from the reconstructed primary joint fused image using a multi-scale approach, into the up-sampled MS image. The performances of the proposed method have been studied both at reduced resolution and at full resolution. Three different datasets, acquired by the QuickBird, Pleaides-1A and WorldView-2 sensors, are used for validation. Compared with several well-known algorithms, experimental results reveal the validity and the advantages of the proposed DiLRP method.