Design of Dictionary Learning Based Algorithm for Single Image Super Resolution via Sparse Representation

Abstract

Super-Resolution (SR) is increasingly becoming a vital factor in image processing due to its widespread applications: medical imaging, remote sensing, security surveillance, video standard conversions. An image super-resolution estimates a high resolution (HR) image through provided single or multiple low resolution (LR) images. The term resolution refers spatial resolution of an image which indicates the smallest region that a sensor can resolve. A higher spatial resolution enables image processing tasks, such as recognition, classification, or analysis of the small object in an image effectively. Traditionally, to observe such information, image zooming has been performed. However, zooming in an image does not improve spatial resolution but stretching the pixel. SR methods solve this problem by estimating missing pixel values intelligently such that detailed information is visible. The last few decades have witnessed huge growth in SR methods, either single or multi-frame-based. These are two broad classes of SR methods based on single or multiple input LR images. Multi-frame SR methods utilize complementary information from multiple images of the same scene. A common issue in this approach is it requires subpixel misalignment among the images of the same scene. Further, multiple images of the same scene may not be available for reconstruction. In such cases, single image SR (SISR) methods are helpful. These methods reconstruct an HR image from a given single LR image. As a result, SISR becomes a challenging ill-posed problem since many HR images map with the same LR image. The simplest techniques to achieve SISR include interpolation-based approaches. Although these approaches are inherently simpler, it fails to recover high-frequency textures and produces overly smooth images. Due to this, learning-based methods are popular for SISR. Learning-based methods include an exhaustive dataset of images to train a pre-defined model. The model architecture, model hyperparameters, loss functions, and training dataset a