Research And Implementation Of Video Super-resolution Reconstruction System

Video has the advantage of vivid and intuitive expression,so it's the main carrier for human obtain information or visual enjoyment.With the development of technology,high-resolution or ultra-high-definition video become the pursuit of industry.Video super-resolution a technology of up sampling a low-resolution video to high-resolution video,and try to restore the video frame which has lost high-frequency details.Improving the accuracy of registration between frames is the key to improving the effectiveness of traditional airspace reconstruction method;Super resolution is an ill-posed problem,and the regularization constraint term is the key factor for getting optimal high resolution frame.In addition,the methods of searching for relationships between low-resolution and high-resolution image patches by machine learning are difficult to designing model structure and tuning the super-parameters.In view of the above problems,combined with image super-resolution technology,the following methods are proposed in this thesis:1)A registration algorithm AKAZE-ILDB(AKAZE-Improved LDB)is proposed,which estimate more accurate inter-frame motion information and lay the foundation for the reconstruction of video image.This algorithm base on the AKAZE algorithm,extracted the features of two images,and using Improved Local Difference Binary(LDB)descriptors to construct feature vectors.Using hamming distance measure the similarity of image features,estimating the motion matrix between the image frames,and complete the image registration.2)Improved incremental video super resolution reconstruction.After registration,The bilateral total variation regularization is used as constraint term,which make a unique and optimal solution.Using L1 normal form as a fidelity term,which measure the degree of similarity between high-resolution after degradation and the input low-resolution image.Maximum constraint term and fidelity term to make the degradation model more accurate.Using sliding window incremental reconstruct the video at last.3)A video super resolution model with improved convolution neural network is proposed.It's an application of the machine learning method in the field of super resolution.In this paper,by improving the network structure and tuning the model super-parameters,based on a large number of high and low resolution image patches,training the network weight parameters,we got a better reconstruction effect compared with the original algorithms.Finally,a video super-resolution prototype system is constructed by method of spatial proposed algorithms are verified.