Super-resolution Reconstruction For Single Image Based On Sparse Representation And Nonlocal Mean

With the continuous improvement of technology and living standards,we have more imaging products.Images and multimedia has become a very important medium of information transmission.However,in some processes such as imaging,storage,and so on,both the external lighting environment and the inherent defects of the imaging device itself,the quality of the images we often obtain is unsatisfied.This type of image is missing a lot of details of its own,in order to restore this information to access to high-resolution images.There are two general methods: First,by improving the quality of imaging equipment,but this will result in the high cost,and the current technology also has greater limitations;Second,to find the software method instead,that is,image super-resolution Reconstruction technology.Image super-resolution reconstruction technology in recent years has become a hot topic,and the image super-resolution reconstruction technology based on sparse representation is favored by scholars.The super-resolution reconstruction techniques based on sparse representation are generally focused on the construction and updating of dictionaries.Few people are concerned with the effects of the high-frequency edges of the training images used in dictionary construction on dictionary training.At the same time,using the image super-resolution technology based on sparse representation to reconstruct the image in some edge texture details can still greatly be improved.Based on the above two shortcomings,this paper mainly done two aspects of the work:First,we propose an image super-resolution method based on Gaussian Laplacian and sparse representation.The method is based on sparse representation of the image super-resolution technology framework,the main content is about a few aspects:(1)Image training set only need high-resolution images,low-resolution image training set obtained by high-resolution image blur and degradation.(2)The extraction of the edge and texture features of the image is carried out.(3)The extracted high-frequency feature of the edge is used to carry out the training of the high and low resolution complete dictionary.(4)Reconstruction of low-resolution test set images.Second,a non-local mean image super-resolution reconstruction based on direction feature is proposed.The main source of the idea is to find that there are many similar directional texture features in image.Such as the edge of a leaf,clothes lines,etc.,the direction of the image information is everywhere,and after the first step in the reconstruction of the image in the texture details could be improved.The direction feature information of the image is extracted by the curve transform,and then the non-local mean algorithm is fused to fine reconstruct the image.Experiments show that after reconstructing the image,both in the peak signal to noise ratio,structural similarity in objective evaluation indicators,or in the visual details,have achieved good results.