Blind Image Blind Drop Qualitative Parameters Estimation And Super Resolution Reconstruction Algorithm

Image super-resolution reconstruction has a very wide range of applications in many areas such as clinical diagnosis, public safety, satellite remote sensing and so on, it can be an effective solution to raise the limit of improving imaging resolution from hardware device and further improve the image quality which is caused by the degradation process. After a series of study, many scholars have put forward a lot of practical algorithms of super-resolution reconstruction, but most reconstruction algorithms are built on the condition of degradation parameters were well knownd which limited the range of applications. So scholars should be focus on estimating the unknown parameters and researching the blind super-resolution reconstruction. This article focuses on the blind super-resolution reconstruction of multiple images and it is consisted of noise estimation, blur kernel estimation, registration parameter estimation, blind super-resolution reconstruction.Our main achievements as follows:1) This paper analysed the over estimated problem of traditional RAP method which based on BayesShink and proposed two new algorithm of high-precision image noise estimation named FB-RAP and RFB-RAP. The FB-RAP’s basic idea is to find a flat block in the noise image according to the principle of edge detection, then estimate an approximate range of the noise intensity, after that we use wavelet transform threshold to denoise the flat block for each of the noise intensity, then create a denoised residual autocorrelation power matrix, and finally estimate the noise variance according to the matrix. In order to overcome the influence of texture and edge, we use the random sampling theory to improve our algorithm.The experiment shows that our algorithms can estimate the noise automaticly and effectively.2) We deeply analysed the traditional algorithm of blur kernel blind estimation and verified a fact that step edges in the image after convoluted with the blur kernel will form a projection profile in a certain direction, and the profile is mirror symmetrical to the blur kernel in that direction of projection profile. According to this theory we designed and implemented a blur kernel estimation algorithm based on Radon transform. This algorithm is consisted of extracting fuzzy edges, establishing a rule of edge filtering and importing the priori knowledge of blur kernel and source image. At last we implement an iterative algorithm to estimate the blur kernel.3) By combination of mutual information and a framework of image registration algorithm, we designed and implemented a algorithm of registration parameters blind estimation based on mutual information. This algorithm is consisted of interpolating, searching strategy, geometric space transform, similarity measure and so on. At last we introduced an open source project ITK, and used the ITK kernel to estimate the registration parameters.4) In order to achieve super-resolution reconstruction for multiple images, this paper designed and implemented a reconstruction algorithm based on regularization and elaborated the entire solving process. Finally combined with noise estimation, blur kernel estimation and registration parameter estimation, we provided a scheme of blind SRR for multiple images.