Research On Spatial Super-Resolution Image Reconstruction Algorithms

Spatial super-resolution(SR) image reconstruction is a process of producing a highresolution image or image sequence from one or more related low resolution images, whichhas a broad application prospect and significant social and economic benefits in almost alldigital image processing application area, including medical image analysis, videosurveillance, high definition TV conversion, remote sensing survey and mapping, astrometryand star centering, etc. Whereas, because SR image reconstruction itself is an ill-posedinverse problem and involves in combination of several related techniques, this technology isstill faced with lots of challenges and becomes one of research hotspots in image processingarea.In this thesis, related theories and methods such as image registration, image restoration,image reconstruction are discussed and processes in spatial SR image reconstruction aresystematically and deeply studied. According to SR reconstruction process and requirement ofpractical application, the research issues of this thesis include:(1)single-frame SR imagereconstruction;(2)image registration and SR reconstruction for image sequence;(3)performance and robustness of SR reconstruction algorithms;(4)applications of SRreconstruction technology. Motivated by these issues, main contents and contributions of thisthesis are summarized as follows:1. Considering texture in character images mainly lies in vertical, horizontal and diagonaldirection, an image smoothness measurement by using a flexible template as convolutionkernel is put forward. Further, applying this measurement as priori information, a Maximum aposteriori(MAP)-based single-frame blind SR reconstruction algorithm is proposed under theBayes framework. Experiments show that the proposed algorithm is robust to noise, whichcan improve correct recognition rate of character area image in license plate recognition anddocument recognition applications and also can be applied to general images.2. Aimed at image sequence captured in short time intervals which can be described bytranslation and rotation model, a fast SR reconstruction algorithm based on Keren registrationis proposed. This method first registers the image sequence by Keren algorithm, then mapspixels in low-resolution(LR) images onto a high-resolution(HR) grid by transformationparameters, and last conducts image fusion and filling to compute pixel values. It is proved tobe robust to registration errors within a certain range. Experimental results show thatcompared with traditional SR reconstruction algorithms, the proposed method has obvious advantages on reconstruction visual effects and execution efficiency.3. In order to handle multi-resolution image sequence shot under various focal length andextent SR algorithms to more general projection transformation model, SR reconstructionmethods based on Scale Invariant Feature Transform (SIFT) features and Harris corners areintroduce. First, SIFT description vectors or Harris corners are extracted as image features.Then vector angle cosine method for SIFT descriptors or two-way neighborhood cross-correlation method for Harris corners is used as the initial feature matching. Further, toimprove registration accuracy, Random Sample Consensus (RANSAC) is applied foreliminating mismatch features. Finally, pixel reliability is defined, and based on this definitiontwo “holes filling” strategies, including template convolution algorithm and pixel reliabilityweighted algorithm, is proposed. Thus SR reconstruction for the image sequence isaccomplished. Experimental results show that SIFT-based SR reconstruction algorithmsproposed in this thesis have preferable performance when dealing with multi-resolutionsequence with relatively large zoom scale. In addition, the proposed method overcomes thedeficiency of performance degradation in the case that LR frames is few and availableinformation is insufficient, which many other SR reconstruction methods suffer from.4. For astronomical image sequences which cannot be directly processed by commonimage registration and reconstruction methods, based on the comparison and analysis of highaccuracy star centering algorithms, a star image SR reconstruction method combined flat fieldcorrection, automatic star search and centering and triangle-match-based registration isproposed and applied to the recognition problem on multi-peak or double-peak structure ofgalaxy and stars. Both simulated and real image sequence experiments verify the effectivenessand superiority of the proposed method.