Physical Electronics Research On Image Super-Resolution Algorithm Based On Sparse Representation

Nowadays,the people's needs for high-resolution images with high quality are becoming higher and higher.With the development of the science and technology,the usage of the software to upscale images has a very high research value.The image super-resolution technique uses a number of low-resolution images to reconstruct high-resolution images with high quality as much as possible.Currently,this technique has been widely implemented into remote sensing,medical imaging,video surveillance,video transmission areas and achieve good results.It has become a hot research topic in image processing areas.The super-resolution algorithms in the literature can be divided into three categories,such as methods based on the interpolation,the method based on reconstruction and method based on the example learning.In these methods,the interpolation methods are the fastest methods and earliest methods in history.The advantage of them is that they are fast and simple,and the disadvantage of them is that the quality is low and cannot meet people's needs.The reconstruction methods are not difficult to implement.However,they cannot achieve a very good quality,especially when the up-scale factor is large.The learning-based methods have the best quality among the three kinds of methods.They usually use the learned low-resolution dictionary and high-resolution dictionary,and the relationship in the dictionaries to reconstruct a high-resolution image.The advantage of these methods is that they are usually slow in the running time.They can be treated as the methods to use the time to exchange the space,currently a lot of research efforts are in these methods.In the learning-based method,currently there are tow subcategory methods,one uses single image and the other uses multiple input images.In the thesis,the research is carried out for single input image,and it achieves the following three improvements:(1)The method based on the relative total variation(RTV)decomposition model and the homotopy method.It decomposes an image into texture and structure components.In the pre-processing,the RTV method is used to decompose an image.It has good results and is beneficial to the following reconstruction process.In the training process,a pair of low and high resolution dictionaries is jointly trained for the texture component.In the reconstruction phase,the Lanczos3 interpolation method is used for the structure component to improve the bicubic interpolation method.For the texture component,the homotopy method is used to improve the traditional orthogonal matching pursuit(OMP)algorithm.The results show that the proposed method effectively improves the reconstruction effects.(2)The super-resolution methods based on RTV decomposition model and LRTV(low rank total variation model).The LRTV model was firstly used in the medical image areas.This thesis revises the original TV structure model and uses the method discussed in(1)previously.After the RTV decomposition,the LRTV model is revised and applied to the structure component processing.Finally,the processed structure component and the texture component are weighted combined.Results show that good effects have been achieved for this method.(3)In the traditional sparse representation learning,only on pair of low-resolution Dl and high-resolution Dh is used.This thesis uses the sharpness measure(SM)for feature classification and classifies the training set into two classes.Two pairs of dictionaries are trained.The high-frequency part of the images are reconstructed by using the L1-homotopy method,and the Lanczos2 method is used for the low frequency part to boost the performance.