Research And Realization Of Super-resolution Image Reconstruction Algorithms

In most imaging applications such as surveillance, recognition, transmission and medical application, high-resolution (HR) images are always desired. However, during the process of image acquisition, the observed images are low-resolution (LR) images owing to many limited factors. Although traditional image interpolation can increase the image size, it can't produce the high frequency components lost in capturing the images. The traditional image restoration can improve the visual effect, but it can only use the information of one input image, the effect of restoration is greatly limited. One promising approach is super-resolution (SR) image reconstruction, which refers to the technology of reconstructing an HR image from observed multiple LR images acquired from the same scene for facilitating better visualization. It can overcome the inherent resolution limitation of LR image systems and obtain the additional information from each LR image. Therefore, during the last two decades, this technology has been widely developed.The reconstruction algorithms are deeply studied in this paper. First we introduce the theory foundation of the problem and set up a suitable image degradation model, then the SR technology is generally composed of three steps: image interpolation, motion estimation and image reconstruction. Some important reconstruction algorithms are discussed, and the POCS algorithm is analyzed. The original image, convex sets, frequency spectrum characteristics of each PSF function, projection operator and iterative cycle of POCS algorithm are discussed in detail. Based on these discussions, an improved algorithm based on adaptive threshold is proposed. The proposed algorithm can adaptively modify the threshold in order to apply constraints on the solution and restrain the noise caused by projection by taking full advantage of the statistical property of projection residue.Experimental results show that the reconstructed image quality of our algorithm is improved according to both subjective and objective evaluations.