Research On Video Super-resolution Algorithm Based On Maximum A Posteriori

High-resolution video frames are reconstructed by mixing the relevant information in previous and next frames in video super resolution technology. In multiple areas, such as HDTV, network video, security and protection control, military reconnaissance, real-time video transmission, video super resolution has wide application prospect. Super-resolved results are influenced by reconstruction algorithm and the accuracy of the degrading parameter. Maximum a Posteriori (MAP) based video super resolution could eliminate the ill-posedness of high-resolution frame reconstruction through introducing prior constraints, and also realize the co-estimation of high-resolution frame and degrading parameters including blur kernel and motion vector.The research background and significance of video super resolution are described in this paper. The research status are also classified and summarized. The degrading model and the performance constrants of video super resolution are demonstrated.An improved video super resolution MAP reconstruction algorithm is proposed in this paper. Each motion vector is calculated using layed-pyramid optical flow method with luminance and gradient constant assumption. The high-resolution frames are reconstructed using bilateral total variant regularization. A guided current frame data term is added to the cost function. The transformation matrix an the transpose matrix are estimated by means of poposed bi-directional motion estimation. To deal with the imperfect motion vector, the error-controlling mechanism is proposed. The difference between reference frame and the motion compensated current frame is adopted as the evaluation criterion, and the confidence mask and confidence function are therefore created to weaken error iterative transmission. Experimental results show, better reconstructed high-resolution frames and more robustness to estimation errors are obtained by proposed improved MAP reconstruction method.For the case with unknown degrading parameter, an hybrid MAP model based video super resolution algorithm is further proposed. The high-resolution frames, the blur kernel and the motion vectors are alternately estimated by maximum their marginal probability. The estimation process is improved. The motion vectors between low-resolution frames are first calculated and interpolated to high-resolution lattice. The blur kernel estimation and high-resolution frame reconstruction is conducted before refreshing the motion vectors. A blur kernel MAP estimation method is proposed, in which coordinate decomposition is not used. Considering the unknown parameter, we propose a hybrid regularization reconstruction method with improved L1and L2 norm. The improved L1norm is adopted in the initial process to increase the robustness. The L2norm is then used to improve the reconstruction performance after the degrading parameters are estimated more accurately. Experimental results show that the degrading parameters could be estimated and more image details are restored using proposed method.