Research On Variational Bayesian Image Super Resolution Algorithms Based On Adaptive Prior Models

In the process of obtaining and transporting images, many factors will make the image degrade, such as noising, blurring, down-sampling and deforming etc. It is a practical problem to improve the quality and resolution of images. The super resolution (SR) technology has been preferred by researchers in recent years. In the field of computer vision, this technology can refine images further, and promote the research based on the image content, such as the development from image detection to image recognition. Nowadays, the SR technology has important application prospects in military, medical science, and public safety field etc. The variational Bayesian image SR technology can estimate the high resolution image, motion parameters, and hyper-parameters jointly in a single framework, and it is a milestone for multi-frame image SR technology. The currently used variational Bayesian image SR technology is still far away from satisfaction. For example, it can not preserve edges while removing the noise, or estimate all the algorithmic parameters automatically.Aiming at the above problems, a deep research for establishing image prior model is launched, and the efficiency of the proposed models is analyzed theoretically. Then a series of variational Bayesian image SR algorithms are proposed. Beginning from two types of commonly used images, vehicle plate image and barcode image, to general nature images, a generalized adaptive algorithm has been proposed. The main contents are presented as follows:(1) A Gaussian model and the TV-SAR model based variational Bayesian SR algorithm for vehicle plate image is proposed. Firstly, a Gaussian model is designed to describe the bimodal distribution characteristics of the image grayscale, and the TV-SAR model is used to describe its piece-wise constant feature. Then, these two models are unified into the variational Bayesian framework. An adaptive SR reconstruction method is proposed constrained by the Gaussian model and the TV-SAR model. The experimental results show that the proposed algorithm can improve the reconstruction quality.(2) Inspired that the Gaussian model and the TV-SAR model could not be applied to the barcode image whose edge intensity changes relatively greatly. An edge-preserving (EP) variational Bayesian SR algorithm for barcode image has been proposed. An EP prior model according to the characteristics of barcode image is established and it is applied into the variational Bayesian framework. The discontinuity-adaptive functions are used to measure the local smoothness. On the other hand, the proposed model also takes the anisotropic property into account. The experimental results indicate that the proposed algorithm could adjust the intensity of constraints for the barcode image and improve the contrast and resolution of the reconstructed barcode images.(3) Inspired that the EP model cannot distinguish the weak edges and noise point in nature images, an AHQ model based variational Bayesian SR algorithm for nature image is proposed. By defining an adaptive half quadratic function, an image prior model is built. This function could adjust the smooth constraint intensity according to the local features of the natural image adaptively. The experimental results indicate that the proposed algorithm could preserve the image edges effectively.(4) However, the AHQ method will result in the artificial artifacts when the noise is strong. In the end, a hybrid model based variational Bayesian SR algorithm is proposed for the strong noise. The proposed algorithm is integrated by the AHQ model and a Tikhonov model using weighting function which is estimated by a broken line model. The experimental results show that, in the reconstruction process, the proposed algorithm can assign weights to AHQ model and Tikhonov model adaptively. What's more, the proposed method can not only preserve edges but also avoid artifacts effectively.This research is based on the barcode image and vehicle plate image, and a generalized SR reconstruction method for natural image in strong noise is developed finally.