Image Super-resolution Studies Based On Sparse Representation

Image super-resolution is a technique used to reconstruct high-resolution image from a single or a series of low-resolution images, and is widely used in image compression, high-definition digital TV, remote sensing, medical diagnostics and other fields. In recent years, the method of signal sparse representation is applied to image super-resolution to obtain better results than the traditional methods. This paper researches on the application of sparse representation in super-resolution, focusing on the super-resolution methods of sparse representation based on learning dictionary. The dictionary of such methods is learned through training, so it can make efficient use of the priori information of image, and obtain high-resolution image with rich details.There are two key problems in sparse representation, dictionary training and sparse decomposition. These two problems are studied throughly in the thesis. The previous methods are improved and two super-resolution methods of sparse representation based on learning dictionary are proposed:a method based on improved dictionary training and a method based on improved sparse decomposition. The innovations of our work are as follows:(1) In dictionary training, an image super-resolution reconstruction method based on improved K-SVD training method is proposed. Firstly, the factors that affect the quality of learning dictionary are analysised, two kinds of methods to improve the quality of training samples are proposed:a) the spline interpolation method is used for the interpolation of the low-resolution training images, replacing the commonly used bicubic interpolation method; b) an image recovery process before extracting the training samples from traning images is added, and the iterative back-projection method is used in this process. Secondly, the iterative back-projection mehtod is improved, and the improved method is used in the optimization of sparse reconstructed image, resulting further improvement of the quality of reconstruced images.(2) In sparse decomposition, an image super-resolution reconstruction method based on improved iterative shrinkage shresholding method is proposed. The tradintional iterative shrinkage shresholding method has shortcomings, such as high computational complexity and slow decomposition rate. Be aimed at these issues, the fast iterative shrinkage threshold method with incremental iterative steps is studied, and an improved method of threshold selection is proposed. This modified method is then used in sparse decomposition of image super-resolution reconstruction. The new method effectively reduces the number of iterations and speeds up sparse decomposition rate. Experiemental results show that this method improves the quality of the reconstructed image and at the same time reduces the computational complexity.