Research On The Enhancement Algorithm For Images In Video Sequences

Video sequence enhancement is that the method processing less detail information and grainy video sequence to get more number of pixels as while as more detailed video sequence. By the computer processing technology, the application of the research of video sequence enhancement algorithm is very meaningful in the military, medical, pattern recognition, video surveillance applications, for that it can get higher resolution under the condition of limited device and reduce the hardware cost.Video sequence super resolution enhancement algorithm is to get one or more of the high resolution image from a series of low resolution images. The previous algorithms such as robust algorithm, and based on the maximum a posteriori probability (MAP) algorithm, can have good result, because they all have general restrictions based on the following two aspects:first, the video sequence must meet half pixel displacement difference between the images, the final algorithm is therefore difficult to have universal applicability; Second, local motion cannot exist between video sequence, because the result of the super resolution depends on the accuracy of motion estimation and match. Motion estimation algorithms such as optical flow method and SIFT transform, it is difficult to compute each pixel in the image sequence or small batches to establish accurate motion vector in super-resolution calculation, so the traditional methods for local motion sequence treatment effect is very poor.Based on the deficiencies and defects of traditional video sequence, as well as the previous scholars’method, combined with existing algorithm, I do some research and innovation of the following aspects in this paper:1. First of all, I do some processing to blurred image which often appear in video sequences. I improved the previous method of motion blurred image to the estimation of point spread function (PSF) that the defects of inaccurate. At the same time, I applied effectively method to restrain the ringing effect brought by the deconvolution.2. Secondly, I improved the single frame image super-resolution algorithm. Elad and other scholars based on compressed sensing theory put forward K-SVD dictionary training algorithm of image super resolution reconstruction, which used in my method extract high frequency information characteristics to reconstruction. We all know that although K-SVD can reconstruct image texture details, but it is hard to restrain degradation image noise. In order to repress this defect, I combined the method of steering regression kernel with K-SVD method to inhibit the generation of noise by the way of using the nonlinear relationship between pixels.3. Thirdly, I applied the two-dimensional regression kernel theory to the3-d video sequences, namely use the improved super-resolution algorithm on image sequences. The way makes full use of the similarity between front and rear frame image pixels, get more complete information, because that the method using rough motion estimation, the final result is not entirely dependent on the accuracy of motion estimation so that eliminates restrictions of the video selection. Then I apply compression perception theory further to deal with the fuzzy image, get more sophisticated video sequence.4. Finally, in order to accelerate the speed of video processing, while the difference between adjacent frames is very small or no change, I put forward a method to speed up the calculation. According to the principle of video compression, if the image did not change from the front frame, the calculation result of the previous frame super-resolution will directly replace final results. Experiments prove that this method can effectively speed up the whole video processing, makes the video reconstruction have practical possibility.