Research On Super-resolution Reconstruction Of Image Sequences

With the rapid developments of scientific research and applications, the demandof high-resolution video and image sequences grows fast in recent years. However,images are always affected by various degraded factors in actual acquisition process,and it is difficult to meet the need for practical applications. Super-resolutionreconstruction technology has been proved to be one of the efficient techniques tosolve the above problem. Super-resolution reconstruction fuses a sequence oflow-resolution frames to produce one or a set of high-resolution images at a relativelylow cost. So, it has become one of the hot topics of digital image processing.Firstly, this dissertation introduced the research background and developmenttrends of super-resolution reconstruction. Secondly, the systematic analysis waspresented including the theoretical basis, system architecture and the main types ofsuper-resolution method. Subsequently, the dissertation focused on several key issuesof image sequence super-resolution reconstruction, such as motion estimation,keeping locally structural characteristics, locally adaptive regularization, adaptiveparameter setting, and color image reconstruction. On this basis, this dissertationproposed a number of algorithms, and achieved certain research results. The maincontributions and innovation points of the dissertation are as follows:1. To improve the accuracy and speed of image registration, a novel method wasproposed to register consecutive frames based on wavelet transformation andimproved multi-restriction criterion. At first, wavelet image pyramids of the referenceframe and the sensed frame were generated to narrow the search space. The featurepoints were found using Harris detector. Followed by the use of improvedmulti-restriction criterion, matching feature points were extracted from the highestlevel of image pyramids. Least squares technique was employed to calculate theregistered parameters. Then the coarse-to-fine hierarchical strategy was applied. Theestimates of the mapping function parameters were gradually improve by thefollowing levels of the pyramids. Finally, artificial images and actual images wereused to test. Experimental results demonstrated the presented method can quicklyobtain the registration parameters with high accuracy.2. Based on the total-variation regularization model, a novel super-resolution reconstruction method was proposed. The method combined the ideas of low-ordertotal-variation model and beyond digital total-variation model. A new regularizationnorm was presented, termed as locally adaptive digital total-variation, to keep edgesand more details. The experimental results were introduced to illustrate theeffectiveness of the proposed algorithm. Performance analysis shows that our methodis superior to similar existing methods.3. Based on learning and sparse representation, a high-resolution iterative initialimage generation method was proposed. Firstly, non-uniform interpolation was usedto fuse the information of low-resolution images to generate a high-resolution initialinterpolated image. Secondly, a high resolution priori image was calculated by aover-complete sparse dictionary. Then, two images were fused as an high-resolutioninitial image which could be used to reconstruct. The method was made full use of thelow-resolution image-sequence information and the high-frequency priori informationto improve the quality of a high-resolution initial reconstruction image. It shouldcertainly help to enhance the reconstructed image. Simulation results confirmed theeffectiveness of this method.4. An adaptive super-resolution reconstruction method based on trilateralregularization was proposed. To reduce the complexity of regularization parametersadjustment, the parameters were computed through low-resolution image sequencesautomatically. And then, the trilateral regularization function was adopted to keepslope and roof edges. At the same time, the learning-based approach was utilized togenerate a high-resolution iterative initial image. In addition, in order to eliminate theimpact of possible registration errors, the iterative algorithm was used tosimultaneously estimate registration parameters and reconstructed image. Syntheticimage sequences and real image sequences of experiments showed that the methodhas better performance.5. A method of color super-resolution based on a MAP estimation technique byminimizing a multi-term cost function was proposed. The method integrated theimage information of RGB and YCbCr. The RGB information was used to define thedata fidelity penalty term. The components of YCbCr were employed to generateluminance regularization and chrominance regularization items. Then, thelearning-based method was used to improve the quality of the initial reconstructionimage. A series of numerical experiments were performed to show the effectiveness ofthe proposed approach, both in the visual effect and PSNR.