Research On Video Super-resolution Based On Bisecting K-means Clustering And Improved Nearest Feature Line

Due to various environmental noises and factors such as the resolution of the camera itself,the images captured by the camera may be of low resolution and it is difficult to meet the requirements of practical applications.As a result,super resolution technology came into being,used to improve the resolution of low-resolution images or low-resolution video.For video super-resolution algorithm based on image block clustering and key frames,the time-cost of clustering on large-scale image blocks using the K-means algorithm is large,and the super-resolution algorithm uses less neighboring image blocks for super-resolution reconstruction.The problem of reconstructing high-resolution frames lacks high-frequency details.This thesis focuses on the following two aspects of research.(1)In order to solve the above problem that the video super-resolution algorithm uses the K-means clustering algorithm to perform image block clustering with a large time overhead.In this paper,an image block clustering algorithm based on bisecting K-means is designed to realize fast construction of high and low resolution image block training sample database.(2)According to the better local structural similarity of the low-resolution image block and its neighboring image block to be reconstructed,this algorithm introduces the concept of the nearest feature line in the process of video super-resolution reconstruction,and expands the neighboring image block sample library.For the problem that the similarity between the projected point image block generated by the nearest feature line method and the neighboring image block on the feature line is not high,the nearest feature line method is improved,and the unreasonable projected point image block is replaced with a simple reject principle.A neighboring image block with better similarity between the feature line and the image block to be reconstructed is selected as a projection point image block.In this paper,the nearest neighbor feature line method is used to expand the neighboring image block sample database and improve the expression ability of the sample database.From the sample library,the projected point image blocks with higher local similarity and low-resolution image blocks are selected and reconstructed,and high-resolution frames with richer high-frequency details are reconstructed.This algorithm performs super-resolution reconstruction of low-resolution video containing keyframes received from the network to obtain corresponding high-resolution video.Experimental results show that compared to the video super-resolution algorithm based on image block clustering and key frames,the proposed algorithm improves the PSNR of the high-resolution frames reconstructed from the fifteenth frame of the News video by 1.3 db.The algorithm reconstruction time of each frame is reduced by275.9 seconds.This algorithm achieves better video frame reconstruction effect and improves the time efficiency of reconstructing high-resolution frames,which greatly reduces the amount of data transmitted by the video on the network and meets the requirements of practical applications.