Research For Image Super-resolution Reconstruction Algorithm Based On Learning Method

Image super-resolution reconstruction means using the existing digital signal processing methods and technology to improve an image resolution under the condition of a series of low-resolution image without additional hardware devices cost. It can not only improve the visual effect, but also conduce to further image identification and processing. Currently, with the continuous development in the field of machine learning and pattern recognition theory, the learning-based approach is an effective way to solve the problem. Support vector regression and sparse representation are two kinds of good learning methods, which have attracts researchers more and more attention.This paper is based on ideological framework of learning methods, which is also the main line. Especially, this thesis focuses on the support vector regression and sparse representation theory of these two learning methods and introduces the algorithm theory in detail. Besides, it proposes an improved super-resolution image reconstruction algorithm under sparse domain based on support vector regression and a further optimization algorithm combined with steering kernel regression function.The existing support vector regression using super-resolution image reconstruction algorithm can improve the quality of the reconstructed to some extent, which only uses redundant information of image itself and leads to the reconstruction results to be improved. This paper proposes an image super-resolution reconstruction algorithm which combined on a clustering sparse and support vector regression. First, training image data and clustering, building the corresponding sparse representation sub-dicttionary, and then creating the corresponding support vector regression model based on sparse representation between the high-resolution image patch coefficients of and the low-resolution image patch, and finally experimenting on the test images for super-resolution reconstruction under the training model. Simulation experiments validate the improvement quality of the image super-resolution reconstruction algorithm.In kernel regression theory, steering kernel regression function can adjust dynamically the shape of the kernel function, which can also suppress noise as well as preserve the edge information of an image. The steering kernel regression will improve the quality of reconstruction if it is a regularization constraint for super-resolution image reconstruction, especially for an image with noise. Related experiments demonstrate the effectiveness of the optimization algorithm.