Research On Learning-based Image Super-resolution Reconstruction Algorithms

Image super-resolution reconstruction technology is to estimate a high-resolution imagewith better quality from one or a sequence of low-resolution images, with the help of signalprocessing technology. The core idea is to integrate useful information with strongcorrelations and complementarities from multiple images as desired. As super-resolutionreconstruction itself is an ill-posed inverse problem, this subject is still facing manychallenges.At present, image super-resolution reconstruction technology can be divided into threecategories: interpolation-based, multi-frame reconstruction-based and learning-based. Thelearning-based super-resolution reconstruction technology is studied in this thesis, amongwhich the focal points include the manifold learning based method and the sparserepresentation based method. Main contents and contributions include:1. A novel single image super-resolution reconstruction algorithm is proposed, whichintegrates the improved manifold learning based super-resolution and gradient constraintbased regularized reconstruction. At first, a new feature extraction method is put forward formanifold learning based super-resolution reconstruction. The new method combines twofeature vectors: norm luminance and detail sub-band coefficients of stationary wavelettransform, to improve the reconstruction performance. Then the gradient constraint basedregularized reconstruction is implemented, with the learned high-resolution image as theinitial estimate and its gradient as the target gradient field. Experiments show that theproposed algorithm obtains better reconstruction performance both in visual effect and inobjective evaluation.2. A constrained stepwise magnification strategy is put forward for locally linearembedding based image super-resolution reconstruction algorithm, to increase neighborhoodpreserving rate and improve reconstruction effect. Iterative back-projection constraint is usedto modify the magnified image in each step, which decreases errors that may occur during thelearning procedures and ensures the solution of each step to evolve towards correct direction.In addition, in order to take full advantage of the information of the test image, the idea ofbuilding a joint training set is proposed to further improve the performance of the algorithm.Experiments show that, compared with some existing algorithms, the proposed algorithmobtains better reconstruction results.3. A novel image super-resolution algorithm based on the double sparsity dictionary is put forward for the sparse representation method. The double sparsity dictionary combines theadvantage of the analytic dictionary and the learning-based dictionary, and is both adaptiveand efficient. Experiments show that the algorithm significantly improves the reconstructionspeed when at the same time ensures the reconstruction quality, and is suitable forapplications with a high demand on time performance.4. A novel super-resolution reconstruction algorithm based on sparse representation ofwavelet coefficients is proposed. The wavelet coefficients of the training images are separatedinto three parts: low-frequency (LF), median-frequency (MF) and high-frequency (HF)coefficients. The dictionary pairs are trained over the MF-HF wavelet coefficients. The HFdetails lost in the observed low-resolution image are inferred from the learned dictionary pairs,and the initial high-resolution estimate is obtained by inverse wavelet transform. Finally,iterative back-projection technology is used to enforce global reconstruction constraint, whichis simple but efficient. Experiments show that the proposed algorithm obtains betterreconstruction performance both in visual effect and in objective evaluation.5. Two supervised image patch classification methods are put forward forsuper-resolution reconstruction with multi-class dictionaries. These two methods use thephase congruency information and the gradient information respectively as priors to guide theclustering of image patches. Phase congruency provides a measure of the significance of alocal structure, and the gradient is also important information to characterize the imagefeatures. Image patches are classified into smooth patches and non-smooth patches withdifferent orientations, and patches within the same category have similar patterns.6. The two image patch classification methods mentioned above are applied to thesuper-resolution reconstruction with multi-class dictionaries, and satisfactory effects areobtained in both reconstruction quality and running time. And the algorithms are robust tonoise to some degree. Since the image patch has similar pattern with the selectedsub-dictionary, it can be better sparse represented and reconstructed. Finally, non-localself-similarity constraint is used to regularize the problem of super-resolution reconstruction,and to further improve the reconstruction performance. Comparative experiments with someexisting state-of-art algorithms verify the effectiveness and superiority of the proposedalgorithm.