Research Of Motion-blurred Image Super-resolution Recovery

Motion-blurred images which are cased of relative motion are numerous in real life. Peoplealways want to obtain high-quality and high-resolution images. However, the existing imageacquisition technology and the processing technology as well as external factors make the finalresults obtained quite unsatisfactory. So how to get high-resolution clear images has been hot anddifficult in the field of image processing.The content of this study is motion-blurred image super-resolution recovery. Compared to thegeneral super-resolution reconstruction, what we have to solve is not only to increase the resolution,but also to pay attention motion blur processing. Simply put, our study is the whole processing ofmotion-blurred image super-resolution recovery.First, this paper systematically discussed several common motion-blurred forms, from whichwe select a common motion-blurred form--uniform linear motion blur as the research object. Theblur kernel estimation has been studied in detail. This type of motion-blurred kernel has theblurring length and the blurring direction. Traditional differential-based method is easy to beinterference by the shape and texture of the image. There are a lot of limitations in the application.This paper presents a new method which use local variance to select some good image blocks toestimate the blurring direction. Experiments show that such improvement can significantlyimprove the estimation accuracy.It can be said that the estimated blur kernel is an importantprerequisite for our study contents, because the blur kernel estimation accuracy is directly related tothe work of the back.Then, the paper discusses the multi-frame super-resolution restoration problem, detailing theestimated stability problems, including a stable estimator, stable data bound term and stableregularization term. Stable estimates in order to minimize the influence of outliers. Here, we addthe natural properties of the currently popular image to stability data constraints. Experiments showthat our approach has increased significantly in visual and the peak signal-to-noise ratio.At the same time, this paper discusses the single-frame super-resolution restoration problem.As to obtain the actual multi-frame images with sub-pixel motion are very difficult. Most cases,only single frame of the image is useful. In this paper, we divide the problem into two steps. One isthe de-convolution to deblurred, the other is the improvement of the resolution.Of the two steps, we mainly discussed the de-convolution which is based on sparse priori.Here outliers cannot be ignored, too. Because outliers can damage the existing model assumptions that its model is linear blur model and noise is substitute to the Gaussian white noise. The proposedsolution is to divide the image pixels into outliers and non-outliers and treat them separately. Weuse a machine learning of the EM algorithm to do the classification. The experimental results showthat this approach can effectively solve de-convolution with outliers. Finally, this papersummarizes the research work and future prospects.