The Research Of Super Resolution Image Reconstruction Algorithm Based On The Compress Sensing

Super resolution image reconstruction technology improves image resolution by digital signal processing or other software methods under the condition of existing hardware. So it’s widely applied in many fields of image, such as medical, military, satellite remote sensing and other fields due to the lower cost and good effect technique.This thesis adopted an algorithm of single image super resolution reconstruction based on compress sensing theory through to the analysis and research of the compress sensing theory and existing super resolution methods. Compress sensing theory pointed out, a sparse or compressible signal can get a low-dimensional projection matrix which be measured with a not related sparse matrix, then using the appropriate reconstruction algorithm can recover original signal almost perfectly by a small amount of measurements. In this thesis, we take the compress sensing theory in the reconstruction of industrial and satellite remote sensing image super resolution. The image patches conveyed by sparse linear representations with a over-complete dictionary, learning the two over-complete dictionaries of high and low resolution image using the K-SVD algorithm. Then using low resolution sparse coefficient of low resolution image patch in the low resolution dictionary to find the corresponding high resolution image patch in the high resolution dictionary, and last reconstructing the high resolution image patches to high resolution image.Finally, this thesis puts the forward methods to simulate the experiments. It made a contrast with some other common super resolution reconstruction algorithms and analyses the factors that affect the algorithm at the same time. The results show that the single image super resolution model based on compress sensing established in this thesis can not only fully developing the superiority of the compress sensing theory, but the reconstruction of images in visual effect and PSNR value also has better effect than all other methods.