Research On Image Super-Resolution Based On Self-Similarity

With the development of multimedia devices and digital circuit technology, people want to get higher-quality and high-resolution images through an existing image, and extract more high frequency detail information. However, the resolutions of some old images have been unable to meet human needs. In the field of medical observation, space exploration, film production, we demand a higher resolution image to assist judgment and analysis. We study image super-resolution algorithm to obtain high-resolution images to display clearly and easy to process. Furthermore, we can have practical applications in life and production. This paper studies the super-resolution algorithm which is based on the image of self-similarity. To the given low-resolution image, we give an algorithm model and propose the solutions and implementation details. When got a high resolution image, we have the comparative analysis of the effect of the new algorithm and existing common image super-resolution algorithms.The paper first introduces several traditional interpolation algorithms:nearest neighbor interpolation algorithm, bilinear interpolation, autoregressive interpolation and the method based on a combination of adaptive image interpolation. The nearest neighbor interpolation is simple and convenient, but there exist jagged and artifacts. The bilinear interpolation in the x and y directions separately interpolated once, and the effect is better than the nearest neighbor interpolation. But the enlarged image edge is not well maintained with a serious loss of details. The autoregressive interpolation algorithm is an adaptive interpolation method, which has a batch of adaptive processing pixels of low-resolution images. It can better retain the texture features of the images, but will still lose some details. The method of cubic surface fitting to image by combination researches on the inverse sampling process image to fit the curved surface approximation which matches the original patch reversely. This algorithm produces the good results, but it will cause blurred edges and introducing artifacts. For the image super-resolution algorithm judge, we use two methods of subjective and objective method of criteria. One is from the visual effects and the other is the average value of the gradient algorithm the merits of a comprehensive evaluation and analysis. The super-resolution algorithms are researched based on 3x3 window,4x4 window, and 5×5 window. For the image to be amplified, we construct quadric surfaces polynomial approximation of different sized squares in the field to find the correspondence between the pixels to be amplified and amplified between the pixels. Then, we could obtain a mathematical expression. When given the mathematical model, we separately look for the relationship between the pixel and its surrounding pixels based on the image of self-similarity. And then we use the second-order differential, increase the interpolation conditions, introduce the Lagrange multipliers, and give the weight functions. We solve the pixel values of high-resolution images with the known the low-resolution image pixel values. Lastly, we could reconstruct a high-resolution image and complete the algorithm.In the paper, we give the image super-resolution algorithm to achieve a high resolution image from a low-resolution image which based on the self-similarity. The enlarged image could save better edge information and texture features, with good visual effects and high accuracy. With comparison of several commonly used image super-resolution algorithms, we find that the new algorithms are more effective with better fitting effect.