Dictionary Learning Based Super-Resolution Image Reconstruction

Super-resolution image reconstruction (SRIR) is cast as the inverse problem of recovering the original high-resolution (HR) image from one or more low-resolution (LR) images. Recently SRIR has been used in many practical field including medical imaging, satellite imaging and video applications and so on. The model-based and the learning-based approach are two most popular SRIR methods that developed in recent years. The model-based approach is of high efficiency; however, the relationships between images are too complex to be expressed under one model. Moreover, when the magnify factor gets bigger, the reconstructed image degraded quickly. The learning-base approach builds two set of training samples that consists of HR and LR images respectively. The test image is coded under the LR images and the coefficients are taken to recover the HR image using the relationship between HR and LR training samples. This paper is about the learning-based super-resolution image reconstruction. The main works are as follows:(1) A dictionary learning based SRIR method is proposed. Two dictionaries are learned from the low and high resolution images respectively using K-SVD algorithm. The proposed algorithm can reconstruct the HR image by making avail of the relationship between the two dictionaries, and it not only have more accurate coding but also significantly reduces the coding complexity.(2) A multitask dictionary learning based SRIR method is proposed. Considering the differences of image blocks, we cluster the training images into several classes from which multiple dictionaries are trained. Single task is defined as recovering the HR image from a dictionary, and the multitask recovery is adopted which share the information among different tasks. For considering the difference of the training samples, the proposed method has an improvement on the PSNR of the reconstructed images over the single task counterpart.(3) A local constraint and multi-task dictionary learning based SRIR method is proposed. Local constraint about the structural self-similarity of patches is added into the cost function. Therefore, it can balance between the maintaining of the local details and the global approximation. Some experiments on natural images results show that it can improve the quality and visual effects of the image.(4) A residual compensation and multi-task dictionary learning based SRIR method is proposed to optimize the HR image. By adding the residual compensation to the reconstructed image, the proposed method can further refine the details of the edge of the reconstructed image and finally improve the quality and visual effects of the image.This paper was supported by National Science Foundation of China under Grant no.61072108,60601029,60971112 and the Basic Science Research Fund in Xidian University under Grant no.JY10000902041.