Research On Algorithms For Super-resolution Image Reconstruction

Super-resolution image reconstruction is a kind of method that satisfies the demand for high quality images. First proposed in the1960s, its goal is to reconstruct one or a series of high resolution (HR) image(s) by exploiting hidden information from one or a sequence of low resolution (LR) original image(s). The significance of super-resolution image reconstruction lies in its ability to reconstruct higher resolution images on top of image acquisition devices with inherently restricted resolution capacity.The main task of super-resolution image reconstruction is to elevate the spatial resolution of images. The basic idea is to find the high-frequency constituents in the high resolution image that was lost from one or a sequence of low resolution image(s) and then perform reasonable modeling and predicting. However, super-resolution image reconstruction is an ill-posed problem, so prior constraints are required to be imposed on its solution. Possible constraints include time-space complementary information among LRs, prior knowledge on the characteristics of the HRs and non-linear mapping relationships between LRs and HRs.Rested on the wavelet transform based pyramid framework, the thesis assumes that the LR images were located at the lower level of the pyramid and reconstructs HR images by solving the higher levels of the wavelet pyramid. The research is conducted on two mainstream approaches, reconstruction based and learning based, with sub-pixel shifts, gradients and self-similarity as prior constraints. The major research work and contributions are listed as follows:(1) Propose a reconstruction based algorithm CSSR (Cycle-Spinning Super-Resolution). CSSR is based on a thorough investigation on the inverse translation operation, the translation direction and amount along with the averaging among multiple LRs, and applies Cycle-Spinning in image denoising to improve the image resolution by eliminating the impacts by artifacts, taking sub-pixel shift as the prior constraint. Simulation results indicate that CSSR algorithm could eliminate Pseudo-Gibbs phenomenon effectively.(2) Propose a learning based algorithm LLGP (Local Lipschitz and Gradient Prior). LLGP incorporates the way that Lipschitz regularization coefficients varies w.r.t different scales into the wavelet pyramid structure first, then takes gradients from a full analysis of detailed information of the edge pixels along horizontal, vertical and diagonal directions as the prior of image characteristics, and finally reconstructs HR by means of gradient descent method. Simulation results show that LLGP algorithm can remarkably enhance the sharpness at edge pixels.(3) Propose a learning based algorithm SS-ANN (Self-Similarity by Artificial Neural Network). SS-ANN constructs a pyramid structure containing images of different scales by wavelet transform of a single image first, then trains ANN taking the similarity of the detailed information among LR-HR image patches with the same and different scales as a prior, subsequently predicts high-frequency constituents lost by quality degrading from the ANN model, figures initial estimation of HR by inverse wavelet transform as the next step, and finally applies error function to constrain the initial estimation. Simulation experiments demonstrate that SS-ANN algorithm can approximate the real HR image much effectively.To sum up, based on the pyramid framework of wavelet transform, the thesis has conducted researches on three kinds of priors:time-space complementary information among LRs, characteristics of HRs and the non-linear mapping relationship between LRs and HRs, applied Cycle-Spinning, Lipschitz regularization and ANN to the modeling and resolving of super-resolution image reconstruction, and proposed novel algorithms whose effectiveness have been well demonstrated by simulation experiments.