Image Super-Resolution Reconstruction Based On Multiscale And Dictionary Learning

Image super-resolution reconstruction refers to the process of obtaining high-resolution images from a single or multiple images with relatively low resolution using digital image processing.This technology can break through the inherent limitations of existing imaging components and external environmental interference,achieve image resolution improvement.It has important value in satellite map,weather forecast and clinical medicine.The learning-based super-resolution reconstruction algorithm is the current research hotspot,by learning in a rich data set,a set of overcomplete dictionaries is used to achieve image recons truction aiid improve the quality of the reconstructed image.In this paper,learning-based super-resolution reconstruction algorithm as the main line of research,we use the multi-scale self-similarity of the image to achieve dictionary learning,regularization constraints to achieve sparse coding and other research work,the major contents are summarized as follows:To overcome the problems of conventional learning-based super-resolution reconstruction algorithm,which is too dependent on external data sets,only using the relationship between the same scale image blocks as a regularization constraint,we researched a super resolution reconstruction method based on multi-scale self similar learning.In the dictionary training stage,we downward sampling of the observed images at different scales to formed a multi-scale image of pyramid,then using nearest neighbor embedding method to find similar blocks of different scales in pyramid,and the obtained similar blocks are assembled as the datasets of dictionary training,thus the self similarity of the image is introduced into the training sample set to reduce the dependence of the algorithm on the external data set.In the image reconstruction phase,we used the difference between the actual value and the estimated value of the target block as a regularization constraint,then used regularization constraints between layers in pyramid,excavated the self similarity information contained in the multi-scale structure of the image,and then we can established the relationship between the high and low resolution images,and obtained the priori knowledge of reconstructed images,used the priori knowledge to guide image reconstruction and improve the quality of image reconstruction.In order to reduce the time complexity of the traditional dictionary learning reconstruction algorithms,we have studied and improved the dictionary learning and reconstruction method based on K-SVD.Firstly,in the sparse decomposition stage,we use the sparse decomposition method to calculate the sparse coefficient,overcome the deficiency of one atom in each iteration of the traditional method,improved as added multiple atoms satisfying certain conditions in each iteration,and we can obtain the better convergence speed and higher precision of the signal estimation.Secondly,during the dictionary update phase,we model atomic models with fewer times of use,achieve the goal of redundancy,and then obtained a low complexity dictionary,and the dictionary is applied to super-resolution reconstruction,we can get a better results.The experimental results,when compared with many start-of-the-art denoising algorithms,show that the proposed algorithm has a certain degree of improvement both in subjective visual effects and objective PSNR and SSIM values.