A Novel Super-Resolution Reconstruction from Multiple Frames via Sparse Representation

Abstract

In this paper, we have put forward a novel technique of multiple frame super-resolution (SR) reconstruction based on sparse representation. The available SR reconstruction techniques are used to generate a high-resolution (HR) image either from single low-resolution (LR) frame or multiple LR frames of the same scene. But these techniques require some modification to provide desired SR frame when LR inputs are contaminated by the high amount of noise. The proposed multi-frame based reconstruction technique overrates the SR part as well as noise removal simultaneously to achieve better outputs. Multiple frames of a noisy low-resolution (LR) image of the same scene with a sub-pixel shift or rotation are used as input of the proposed algorithm. The registration technique of these images results in a single HR frame having more information than any one frame from the set of multiple frames. Later noise removal, as well as edge reservation, is done by applying a median filter to the HR frame. Then sparse representation technique is used to reconstruct the super-resolution frame of the filtered HR frame. To ensure the qualitative goodness, some well-known quality metrics like PSNR, SSIM, and BLUR are measured for different image inputs and the results are compared with other techniques to confirm the claims of the authors.