| Image super-resolution technology is originated from the field of image restoration. With the development of computer and network, images become more and more important in people’s daily life. Due to the hardware technology and the cost constraints, images often have low resolution, such as medical images, remote sensing images and monitoring images, which usually difficult to clearly distinguish clearly the objects of images. At the same time, development of hardware has been limited. The arrangement and size of CCD or CMOS sensor are difficult to get further optimization and become smaller in short term, which gave birth to the improving of image super-resolution reconstruction techniques in the method of software algorithms. The advantages of image super-resolution reconstruction techniques are relatively low cost and do not need replace the expensive hardware to get the cutoff frequency information of image. It has a high value of research at present in many fields, such as medical imaging, satellite remote sensing, public security monitoring, military reconnaissance, multimedia video. So it has broad prospects for development and application.Keeping the geometric features of the images is the premise and target of the super-resolution reconstruction technology. But the input of learning-based image super-resolution technology has only one low resolution image, which called single image, causes certain image geometry prediction distortion phenomenon. For example, the micro translation restoration of the face image contour or positions of facial features, the micro change of the ground target shape, and micro deformation in the contour of bones or organs in medical examination images, these are all problems and key issues of learning-based image super-resolution technology. As far as possible to reconstruct the original image information, and to reduce the geometric distortion are the objectives of this technology. This dissertation first introduces the development history of super-resolution reconstruction technology which focused on learning-based super-resolution technology and based on Bayesian theory. Then using a priori knowledge and starting from the geometric characteristic of image, this dissertation proposed several methods to improve the reconstruction quality. The classification of image training library, extraction and matching of high frequency and low frequency image features, anti aliasing and anti joint between neighbor image blocks, and the adaptive iterative selection method of the neighbors, are all focused points of this research. The algorithms of this dissertation cover data reduction process, sparse representation, Pyramid transform, NSCT, MRF and LLE manifold embedding super-resolution reconstruction methods. The main innovations of this dissertation are as follows:(1) The dimension reduction method of manifold learning and LLE algorithm is analyzed, so the weight matrix sparse coding algorithm SSME is proposed, which can ensure the accuracy and efficiency of the neighborhood selection. The method through the optimized LLE algorithm of neighborhood selection strategy, by setting the threshold iterative, adaptive to select appropriate neighborhood number and return to neighborhood weights. The neighborhood image dimension reduction is achieved by sparse encoding the weight matrix. Experiment results show that the SSME algorithm can effectively eliminate some outliers in the neighbors, make the reconstructed image has an improvement of about 0.1-0.3dB over the state-of-the art results in peak signal to noise ratio.(2) The problem and the advantages and disadvantages of the NSCT algorithm are analyzed, combining with the characteristics of MRF, the image super resolution reconstruction algorithm based on NSCT and MRF method is proposed. By combining NSCT feature blocks and MRF characteristic block selection, this method can extract high and low frequency characteristic coefficient block of images for super resolution reconstruction of image. The results show the advantages of the method, which include effective elimination of the aliasing in the reconstructed image, easing the storage redundancy, eliminate the "joint" of the super-resolution image reconstruction. At the same time, the geometric features of the image can be obtained, reduce distortion and the image super resolution reconstruction effect is also improved.(3)The questions of sparse dictionary of image super-resolution reconstruction method based on it are mentioned, so the method of image super-resolution reconstruction based on sparse atomic cluster dictionary learning is proposed, which make the reconstruction process in more training images and less sparse dictionary. This method not only makes reduction of the actual learning dictionary size, but also builds in a more accurate collection of image block, which improves the learning speed and quality, and adding a priori knowledge (more training images are joined). Result of it can indicate the obvious improvement of the peak signal to noise ratio and visual effects of image super-resolution reconstruction.The dissertation is based on image sparse representation theory, the reconstruction method, from the keeping geometric features of the image as perspective, and Research on image super resolution reconstruction algorithm, which is for the purpose of improve the image reconstruction of host and guest view evaluation. It focused on the three different methods which aim on the searching image block neighborhood methods, and it includes the manifold learning method, MRF, and classification sparse dictionary learning. The experimental results are better. There is very high practical value of the learning-based image super-resolution method. The future aims are the decrement of image amplification distortion and improvement of speed of the algorithm which by multi feature, multi scale and dimension reduction, machine learning and other related technologies. And achievement of hardware programming can be better service for routine applications.
|
PDF Full Text Request