Study On Image Super-Resolution Based On Human Visual Attention And Controllable Image Quality

Super-Resolution(SR) reconstruction is a method recovering a high- resolution(HR) imagefrom one or more low-resolution(LR) images. Compared with the method to method HR image bythe hardware devices, the cost of SR reconstruction method is cheaper. In recent years SR is widelyused in video monitoring, medical image analysis, high-definition video and others. So, it is of greatsignificance to use the existing devices to get HR images based on SR method. SR reconstructionvia sparse representation has been one of important methods because of its characteristics ofaccurate reconstruction, strong robustness and easiness in choosing parameters.This dissertation focused on visual attention and adaptive quality based SR reconstruction. First,SR reconstruction algorithms based on image patches structure, and based on residual error wereproposed to improve the quality of HR image, and a fast HR algorithm was proposed to reduce thehigh complexity of HR algorithm. And, a non-reference image quality assessment method forapplication of SR reconstruction was proposed and used in implementing adaptive quality SRmethod. Then, combined with visual attention, an SR reconstruction algorithms based on visualattention and adaptive quality was proposed. The main contents of this dissertation can be listed asfollows.1. Basic knowledge about sparse representation of images, over-complete dictionary training,sparse decomposition algorithms were introduced, and K-SVD dictionary training algorithm wasdetermined to build HR and LR dictionaries and obtain sparse representation coefficients vector .2. SR reconstruction model based on learning and sparse representation was introduced. Basedon the fault of Yang method, ignoring the structure character of image patches, the SRreconstruction algorithm based on the structure of image patches was proposed. Firstly, the methodmade statistics and analyzed the structure characters of the image patches, and used different SRmethods to the different patches. Experiments show that the proposed can obtain better results andless run-time than the Yang method.3. Based on the fault of Yang method, non-equivalent of HR patches mean in the training andtesting stages, the SR reconstruction algorithm based on residual error was proposed. This methoduse K-mean algorithms for the sample clustering, and use multiple dictionaries and theirreconstructed HR patches that were weighted to build the SR reconstructed patches. The proposedimproves the quality of super-resolution reconstruction results.4. Based on the fault of high computation complexity current SR method, a fast SR method isproposed. First, an over-complete dictionary ’Guide Dictionary’ was proposed to reduce the computation complexity, and used the Parallel Computing Toolbox in Matlab for speedup Theexperiments show that the proposed can greatly decreased the run-time and keep the quality ofimage unaffected obviously.5. In order to improve the practice of SR reconstruction, we proposed the SR reconstructionalgorithms based on visual attention and adaptive quality. Firstly, we proposed a non-referenceimage quality assessment for application of SR reconstruction. Then, combined with the visualcharacter of human eyes, adaptive quality SR reconstruction was proposed and simulated, and itprovides a new idea for the SR reconstruction of big images.