Super-resolution of remote sensing images via sparse structural manifold embedding

Exploring signal processing technologies to enhance the resolution of remote sensing images has received increasing interests in the last decade. In order to well preserve the structural details such as edges, contours and textures in the recovered high-resolution images, we advance a new Sparse Structural Manifold Embedding (SSME) approach in this paper. By incorporating the geometric regularities of images along singularity of edges or contours into neighbors׳ selection, SSME can well recover structural information of images. Moreover, considering that outliers are often included into embedding to generate inaccurate structures, a robust and sparse embedding is used to exclude outliers in synthesizing high-resolution images, where normalized weights are employed to acquire more accurate neighbors and coding coefficients. Experiments are taken on realizing a 3× amplification of remote sensing images, and the results indicated that SSME has an improvement of about 0.1-0.3 dB over the state-of-the-art results in peak signal to noise ratio.