Study On Super-resolution Image Restoration Based On Prior Constraints

In the process of image and video acquisition and transmission,such as the limitation of imaging conditions,natural scene changes,time and spatial resolution of imaging equipment and other factors,the imaging and video acquisition system is difficult to obtain the information in the natural scene without distortion.Despite the rapid development of imaging and video acquisition equipment and the improvement of hardware performance,the acquisition and transmission of high-quality and high-resolution images is still limited by many factors in specific application fields such as satellite remote sensing and aerial photogrammetry,industrial imaging monitoring,criminal investigation analysis,medical image analysis,public security video monitoring,video entertainment system and multimedia communication.An effective way to improve the temporal and spatial resolution of images(or sequence images,video)is to adopt a software-based approach(that is,the approach of signal and information processing algorithm)without changing the hardware of the original system.This software-based super resolution image restoration(SRIR)method and technology refer to the fusion of low-resolution images(or video sequences)from the same scene(multiple frames)and the restoration and reconstruction of high-spatial resolution images or high-time-space resolution video sequences.For the still image SRIR,high resolution image restoration and reconstruction can be achieved for single or multiple low-resolution images of the same scene.For SRIR of video sequence,the realization of low-resolution video sequence of the same dynamic scene includes recovery and reconstruction of video sequence with high temporal resolution and spatial resolution.This thesis takes the maximum posterior probability(MAP)estimation method and variational method of finding extreme value of norm as the main line of study,combining the prior modeling as the solution space constraints.Aimed at multi frames,single images and video sequences,the improvement method based on prior(regularization)constraints,the method of adaptive sparse representation combined with regularization constraints,and video super-resolution reconstruction method based on pixel stream and time prior are researched.The research focuses on three aspects:(1)an improved research on regularization of SRIR under solving framework of maximum a posterior(MAP)estimation for multiple low-resolution images;(2)an improved research on sparse dictionary learning based SRIR for single image;(3)an improved research on video SRIR based on pixel stream and time feature prior modeling.The main research and innovation points include:(1)For multiple frame images,a generalized total variation regularization SRIR algorithm based on neighborhood pixel extension is proposed,which overcomes the deficiencies of total variation(TV),bilateral total variation(BTV)regularization and existing generalized total variation(GTV)regularization under MAP estimation solution framework.In order to improve the accuracy of correlation measurement,the algorithm expands from the correlation of geometric distance between adjacent pixels to the double correlation of geometric distance and gray level between adjacent pixels,and designs the cost function of which is different from 1-norm form.In addition,the MM iterative optimization method is also introduced instead of conjugate gradient(CG).Compared with the other algorithms,the results indicate that the proposed algorithm has better ability of suppressing noise and maintaining edge details and higher robustness.(2)For multiple frame images,a regularization SRIR algorithm based on the improved fidelity and adaptive bilateral total variation is proposed.Exiting SRIR algorithms build fidelity item by L₁,L₂,Huber or Gaussian norm under MAP estimation framework,which have the problems of being sensitive to other data and robustness constraint.This proposed algorithm builds fidelity item by Tukey norm to slove heavy tail effect which can be used to deal with complex noises,while the establishment of adaptive weight matrix can further enhance the image detail.By multiple experiments,the results show that the proposed algorithm has advantages in super-resolution image restoration,and has high robustness.(3)For single image(or frame),a super-resolution image restoration algorithm based on adaptive sparse representation combined with regularization constraints is proposed.When building global over-complete dictionary,sparse dictionary learning based super-resolution image restoration method has the problem of sparse representation constraints to different structural image blocks,etc.,in order to improve the accuracy of sparse representation,encoding efficiency,the maintaining of image detail information and so on,combining the sparse representation of image with variational method for finding the extremum of norm function,the traditional sparse dictionary learning based SRIR method is improved by adopting adaptive sparse representation and regularization constraint.Moreover,the sum of absolute differences(SAD)is utilized to measure the pixel structural similarity of image blocks in order to reduce computation and improve computational efficiency.Experiments demonstrate the SRIR performance of the proposed algorithm has advantages in visual effects of the restored image,maintaining edge details,suppressing noise,computational efficiency and so on.(4)For video sequence,a video SRIR algorithm based on pixel stream and time featured prior is proposed.Most of the exiting video SRIR methods have limitations of spatial defuzziness and motion blur elimination and the enhancement of interpolation frame fidelity and so on.In order to effectively increase frame rate and reduce(or eliminate)motion blur in video,under MAP,a single video time-space super-resolution restoration algorithm is proposed,which adopting pixel stream based time-space super-resolution restoration and feature driven based pixel stream time prior.By multiple experiments of single gray video and single color video independently,the results indicate the effectiveness and superiority of the proposed algorithm.