Study And Implementation Of Fast Super-resolution Reconstruction Methods

Image super-resolution reconstruction is a digital signal processing technology, whichreconstructs the high-resolution images from one or more low-resolution images. Compared withthe method to get high-resolution images by hardwares, super-resolution reconstruction techniquescost less. Super-resolution reconstruction techniques are widely applied in many fields, such aspublic security detection, medical imaging, high definition video, military remote sensing detectionand so on. It is of great significance to use super-resolution techniques to obtain high-resolutionimages.However, in these practical applications, state-of-the-art super-resolution reconstructionalgorithms have the characteristics of complicated computation, high time-consuming, and can notsatisfy the requirement of real-time reconstruction. Considering the above deficiencies, fast imagesuper-resolution reconstruction algorithms are studied, and research results are achieved as follows:1. Studying the super-resolution reconstruction algorithm based on the sparse representationmodel proposed by Yang, we propose a fast super-resolution reconstruction method based on sparserepresentation and atom-clustering. We mainly use K-means algorithm to cluster the dictionaries,and similar atoms are partitioned into the same cluster. An atom-cluster is selected adaptively foreach given image patch. Experimental results show that the proposed method not only reduces thetime to reconstruct image, but also improves the quality.2. Studying the PADDLE(Proximal Algorithm for Dual Dictionaries Learning), we introduce thePADDLE into the super-resolution image reconstruction based on sparse representation. Theproposed method approximates the sparse decomposition of the low-resolution patch asmultiplication of the dual dictionary and the low-resolution patch matrix. Experiments show that themethod is able to achieve desirable super-resolution images with significant computationaladvantages.3. Studying the Graphic Processing Unit (GPU) architecture and the Compute Unified DeviceArchitecture(CUDA) programming-model, we analyze the time-consuming of the fastsuper-resolution reconstruction method based on the PADDLE, and implement the fastsuper-resolution reconstruction using CUDA programming-model on the GPU.