Super-resolution: Sparse dictionary design method using quantitative comparison

Abstract

Single image super resolution (SISR) is the process that obtains a high resolution image from a single low resolution (LR) image by increasing the high frequency information and removing the degradation of the noise. Sparse representation of signal assumes linear combinations of a few atoms from a pre -specified dictionary. Sparse representation has been used successfully as a prior in signal reconstruction. Dictionary design is crucial for the success of reconstruction high resolution images. This paper evaluates the performance of dictionary design models in both mathematical and learning based models, it also compares the wavelet method, Haar method, DCT method, MOD method and K-SVD method. Various experiments are conducted using a real SPOT-4 satellite image. Experimental results demonstrate that the learning based approaches are very effective in increasing resolution and compare favorably to mathematical based approaches.