The Single Image Super-Resolution Reconstruction Based On Machine Learning

In practical applications, the demand of the images is becoming higher and higher along with the rapid development of economy and society. However, because of the limitation of hardware equipment and imaging technology, it is very difficult for obtained images to meet the requirements of the people. The needed cost value would be high for improving the imaging quality in hardware equipment. With the rapid development of software technology, people want to achieve the improvement of image resolution through the software technology instead of the hardware device, which is called image super-resolution reconstruction technology. In this thesis, focusing on the single super-resolution reconstruction we have improved reconstruction algorithms from different aspects, and the main research contents and innovation points are listed as follows:(1)Image super-resolution reconstruction is an ill-posed inverse problem without unique solution, so we define the image super-resolution reconstruction as an optimization problem and propose a super-resolution reconstruction algorithm based on genetic algorithm and regularization prior model approaches. The regularization prior model obtains the optimal solution by the iterative search in the local area and is easy to get the local optimal solution. In order to overcome the local convergence, we introduce the genetic algorithm to the iterative process, so that it can search solution in the global scope. This algorithm is divided into two steps. Firstly, we introduce the nonlocal mean filter and the total variation to the sparse representation based on adaptive sparse representation to build the mathematical model. Secondly, the mathematical model is solved with interactive shrinkage-thresholding algorithm. When the estimation result is close to the local optimal solution, the genetic algorithm is introduced to jump out of local search. The contrast experiment shows that the proposed algorithm can obtain better reconstruction results in both theoretical and visual.(2)We propose an image super-resolution reconstruction algorithm based on clustering on the sparse representation coefficient. Sparse representation puts the energy of the image on a very small number of atoms so that the image can be represented in another form, which is more suitable for clustering. In this paper, first we use principal components analysis to obtain sparse dictionary and sparse representation coefficients. Then, we use K-means to divide the feature space of high resolution image and low resolution image intoseveral subspaces. Compared with the previous algorithms, our method can recover the high frequency information and is better than other algorithms.(3)We design a novel algorithm as the improvement of the clustering on the sparse representation coefficient. The results show that previous method can reconstruct the image with more abundant edge and texture. Nevertheless, the reconstruct results are a bit worse for the image with less information. Because the images have complex structures and information, we proposed an image super-resolution reconstruction algorithm based on selection processing of the image patches. It uses different methods to deal with different image features.