Research On Image Super-resolution Reconstruction Method Based On Multi-task Learning

Limited by the inherent properties of optical imaging system and the influences of the environment outside, the resolution of the acquired images may not reach the application standard. Higher resolution images are desired. Image Super-Resolution Reconstruction has been one of the most active subjects in the image process field for these years. The basic idea is to get a high resolution image from one low resolution image or some low resolution images. It can restore the information beyond the cut-off frequency, so the acquired images have more detailed information. This technique is not only widely used in many areas in our daily life, but has important meanings for theoretical research.This thesis is about the image super-resolution reconstruction based on machine learning. The main work completed is as follows:The classical algorithms for Super-Resolution Reconstruction in both spatial and frequency domain and their characters have been studied in detail. including the classical interpolation algorithm, IBP algorithm, POCS algorithm, probability statistics based method and the reconstruction algorithm based on single-task learning.To make up for the lower precision and worse ability of noise resistance caused by the using of single task learning during the solution of complex learning problem, a multi-task learning algorithm is deeply studied. The framework is analyzed in theory, and under the condition that factors, which can describe the relationship among tasks, are introduced, the computation complexity is reduced. According to the framework, the object optimal function and the dual problem for multi-task regression learning are deduced through combining with support vector regression.This thesis proposes an image super-resolution reconstruction algorithm based on multi-task learning. In order to improve the accuracy of learning, the pixels of the image is firstly classified using Gaussian mixture model and the training data for each pixel class are chose according to the maximum membership principle. Then the map model for each pixel class is established using multi-task learning for regression algorithm. The reconstruction is the inner product of the map model and the membership grade. Experimental results demonstrate that the proposed method get a better result.