Research On Image Super-resolution Reconstruction Algorithm Based On Dictionary Learning

With the rapid development of information society,the demand for high-quality,high-definition images are growing.Image super-resolution reconstruction technology is to reconstruct high-resolution images from one or a sequence of low-resolution images in the same scene by using the information processing methods,with the help of computer software technology.Compared with multi-frame image reconstruction,single-frame reconstruction uses only one low-resolution image,whose actual requirements and application range are wider.Therefore,this paper mainly do research on the single image super-resolution reconstruction algorithm.Currently,there are three categories of image super-resolution reconstruction techniques:interpolation-based,reconstruction-based and learning-based.The learning-based approach introduces machine learning theory,which draws into external high-frequency information and effectively enhances the details of image reconstruction.It is a research hotspot of the image field in recent years.Therefore,this thesis focuses on the learning-based super-resolution reconstruction algorithm.Main contents are as follows:(1)Among the super-resolution reconstruction algorithms based on dictionary learning,the natural features of all image patches cannot be represented through a universal over-complete dictionary,besides,due to the disadvantages of the timeliness of algorithm being affected by the low efficiency of the reconstruction algorithm caused by sparse coding in the reconstruction stage,image super-resolution reconstruction based on multi-dictionary learning and linear regression is proposed.The clustering algorithm is adopted to classify the sample feature patches,so that the image patches with similar structures are clustered together,and then dictionary training is performed to obtain multiple compact sub-dictionaries.By using linear regression,the mapping matrix from low resolution to high resolution space can be directly calculated,which can effectively improve the reconstruction efficiency of the algorithm and make full use of the information contained in the training image patches.Experiments show that both the quality of the image and the efficiency of the algorithm achieve better results in image reconstruction.(2)The effective use of image features will be conducive to the final reconstruction effect.Based on the self-similarity features of images,combined with the matrix low-rank theory,image super-resolution reconstruction based on low-rank theory and dictionary learning is proposed.The image can be divided into low-rank part and sparse part by means of low-rank sparse decomposition.Most of information in the image exists in the low-rank part,which is reconstructed with the proposed multi-dictionary and linear regression algorithm.And for the sparse part contains only a small amount of difference information and only bicubic interpolation is used to complete the reconstruction.Finally,the two partial reconstruction results are combined as the final super-resolution reconstruction image.After the sparse part is interpolated and enlarged,the complex learning reconstruction is not further executed;thereby the influence of noise and other factors on the reconstruction process in the original image is reduced.