Generation of High Resolution Medical Images Using Super Resolution via Sparse Representation

Abstract

Generation of high resolution (HR) of images is very important in image processing and computer vision applications such as: CT scan-chest, MRI, X-rays and Optical Coherence Tomography (OCT). In medical imaging field images are very frequently using for analyzing different type of disease symptoms. HR medical images are required for proper diagnosing but due to the hardware limitation it is very difficult. In this paper we propose a HR image generation technique based on image statistic research for single low resolution (LR) input image-using joint dictionary learning. The view of LR image is like down sampled version of HR image and its patches are considered to have a sparse representation with respect to an over complete dictionary. Sparse representation can be recovered under mild condition with compressed sensing perspective. This study explores the use of super resolution (SR), due to learning there is no need to aligned subpixels of different LR images. For validity of our research, we use this technique on OCT and Lungs images and train three dictionaries (i) using OCT images (ii) using Lungs images and (iii) using multiple different natural (MDN)-HR images. Images are produced with using MDN-HR and for both medical images and also used OCT images dictionary for OCT images and lungs images dictionary for lungs images. We compare our result with previous SR proposed technique-from all aspects dictionary learning approach have superiority. Proposed dictionary learning based technique produced enhanced and upsampled image.