Research On Image Super-resolution Algorithm Based On Mapping Matrix

In the field of computer vision,image super-resolution reconstruction technology is one of the research hotspots.It can restore clearer images without changing the existing hardware imaging equipment.Therefore,image super-resolution reconstruction can be widely used in many key areas,such as: public safety,remote sensing images,data compression,high-definition video,and so on.Image super-resolution reconstruction uses a series of algorithms to process low-resolution images to obtain an output high-resolution image.The image reconstructed by the traditional super-resolution reconstruction algorithm has blur problems and artifacts.In order to improve the sharpness of image reconstruction,this thesis mainly conducts related research based on the method of anchor point regression.The main work and innovations are:(1)In order not to be restricted by the external image sample set,this thesis builds a sample set based on the image itself.The image is analyzed from the perspective of energy.The energy of the image is mainly concentrated on the low-frequency component,and the energy occupied by the high-frequency component is only a very small part of it.Therefore,in order to better recover the high-frequency information of the image,the horizontal gradient operator and vertical gradient operator are used to extract the high-frequency components of the image,and then the high-frequency and low-frequency components are separately reconstructed.Finally,iterative algorithms are used to jointly reconstruct low-frequency information and high-frequency information.Compared with other super-resolution algorithms,simulation experiments prove that the image reconstructed by the method proposed in this thesis is sharper on the edges of the image.(2)Aiming at the problem that the edges reconstructed by the existing anchor point regression method have fuzzy edges,this thesis first introduces a cartoon texture decomposition algorithm to decompose the images into cartoon images and texture images to obtain two types of edge images.Different from the traditional anchor point regression method,the MI-KSVD method is used to train the low-resolution dictionary in this thesis,instead of the original K-SVD method,and the independence of the dictionary atoms obtained is stronger.In addition,in the process of regression learning,introducing locality can obtain a better mapping matrix.Experiments show that the method proposed in this thesis is superior in performance to the traditional anchor point regression method,and the restoration effect of image edge information is more realistic.(3)Considering that adding a priori information as a regularization term can make the reconstructed image better,this article adds a priori information item of the image to constrain the reconstructed image based on the anchor regression method to make the reconstructed image The effect of the image is more natural.In order to extract the features of images in multiple directions,this thesis uses four Sobel direction operators to extract low-resolution image features.Since mapping from low-resolution space to high-resolution space is a non-linear problem,this thesis does two linear mappings to improve the accuracy of the reconstructed image.After a series of experiments,it is shown that the image reconstructed by the method proposed in this thesis is richer in detail information.