Research And Implementation Of Single - Frame Image Super - Resolution Reconstruction Algorithm Based On Manifold Learning

Super-resolution (SR) is a kind of technology which uses one or more frames of the low resolution image to restore a higher resolution image,this technology is a software technology which based on signal processing theory.As the super-resolution reconstruction technique can overcome limitation in the resolution of the imaging system without changing the original case, to obtain higher resolution images, the technique does not involve hardware imaging system, it cost lower than any other methods. It has broad applied in remote sensing, high-definition digital TV, medical diagnostics, network transmission and video surveillance and so on.This article select LLE algorithm which based on manifold learning to realize super-resolution reconstruction for single image.The main problem of the algorithm is to construct learning library and to determine field relation.Based on these two issues, we propose corresponding solutions.The main contribution of the paper and the innovation lies on the following aspects:Single image reconstruction method based on LLE algorithm was proposed. Single image SR reconstruction methods based on learning in previous literature have introduced additional high and low resolution images to build a learning library. The similarity of learning library has a direct impact on the reconstructed image reconstruction results. This paper uses only one frame image to excavate the self-similarity and combine LLE algorithm for image SR reconstruction.Learning library construction method was introduced for only one frame image in the LLE algorithm. For SR reconstruction of the single-frame images, this paper introduces an image without the use of additional resources for learning method for constructing single-frame image library, which is self-similar images using image block by sampling the input image to generate fuzzy lower resolution image, and then the composition of low-resolution images with the original input image to build on learning library. This method can not only learn library to build single image, but also solve the learning library to be reconstructed image and the reconstructed image similarity influence the effect of the problem.Feature amount was improved. Current literature shows that the feature information of the image block directly determines the field characteristics of the block and impacts the result of the reconstruction in the methods of learning. Feature normalization (minus the average value of the block) gray value was widely used in current papers,but if there is noise in the image block, it will affect the accuracy of the mean and the characteristics of the image information block. Therefore this paper proposed to use the middle value of the sort block pixel value instead of the the mean value. This method can maintain the advantage of normalized gray feature in the field of information,but also inhibit the noise better.The final results show that the proposed method, without the use of any additional image resources, comparing to the reconstructed image of interpolation and reconstruction methods has a lower MSE and higher peak signal to noise ratio.