Structural-Correlated Self-Examples Based Superresolution of Single Remote Sensing Image

Image superresolution methods are of great importance to image analysis and interpretation and have been intensively studied and widely applied. The main research works on single-image superresolution are how to construct the training image database and how to learn the mapping relationship between low- and high-resolution images. Considering only a single image, a novel super-resolution method for self-examples learning without depending on any external training images is proposed in this paper. The training self-examples are extracted from the gradually degraded versions of the testing image and their corresponding interpolated counterparts to build internal high- and low-resolution training databases. Inspired by the concept of “coarse-to-fine,” the upscaling process is performed gradually as well. The algorithm includes two steps during each upscaling procedure. For each low-resolution patch, the first step is to find structural-correlated patches by sparse representation throughout the training database to learn global linear mapping function between low- and high-resolution image patches without any assumption on the data, and the second step takes the advantage of sparse representation as a local constraint on super-resolution result. At each upscaling procedure, iterative back projection is applied to guarantee the consistency of the estimated image. Moreover, the internal training database will be updated according to the newly generated upscaled image. Experiments show that the proposed algorithm can achieve good performance on peak signal-to-noise ratio and structural similarity index and produce excellent visual effects compared with other super-resolution methods.