Single-image Based Learning For Blur Removal And Quality Evaluation

The age of digital cameras is dawning, and its revolution is everywhere, so is digital image. However, the advancement of image quality due to camera cannot always keep up with men’s expectation. We refer to reducment in image quality as image degradation; there are many reasons for this, which mainly due to blur and noise. With the miniaturization and cheapness of camera lens, and limited to photographers’level, if the camera doesn’t be hold steadily or the object get out of focus, all of which can reduce the visibility of image, and make us hard to distinguish the detail. Some other stochastic factors also take in inevitable noise into image. Great moment once go may never come back any more, which is difficult to copy. Simple but effective approach to restore image and improve the quality became urgency. Our research thus center about various forms of image blurs.Our main work and innovation include the following:1We look back at background, figure out the difference between image restoration and image enhancement. We introduce the advantage of image restoration methods refer to other restoration methods, and forsee its broad application. Concerning what is image restoration, we compare it with the notion of image enhancement. And then we elaborate the key technology difficulties in image restoration method.2Blind deconvolution of blur image is a most challenging job, the illness of which was mainly due to the number of known less than that of unknown. We set up a framework to blind deconvolution based on a single burred image, and try to add auxiliary information as condition to restrict unknowns, here comes concrete measures:two regularization cost function which involve salient edges as spatial prior to alternately iterate the blur kernel and latent image until convergence, where we can get the blur kernel. Then sparse prior is used to restore the latent image which can preserve more structure detail. In order to reduce noise in blur kernel, we put forward some revisal principle. We propose adaptive kernel size to avoid blindly manual set.3Based on respective mathematical model, we make several rapid processes on several special blur. We find blur parameters (which are motion length and motion direction) in logarithmic spectrum in linear motion blur images, and make precise evaluate combine with error-parameter. Defocus radius can be found in first circle- loop of defocus blur image cestrum. We use Hough transform to extract the edge of rotated blur image, and then the least square was applied to fit the location of the center. Thus pixels on each concentric circle were picked up to one-dimensional restoration respectively. A series of morphology methods can segment blur area from latent area in local blur images, then blind deconvolution framework was applied to restore the blur area. Then merge it with latent area can get the last deblurred image.4For noisy blur images which are more ill, we first utilize directional low-pass filter and inverse Radon transform to restructure the blur kernel. In the denoise and deconvolution stage, we construct two cost functions to denoise and deconvolution alternately, latent image can be obtained when they convergent.5We put forward a criterion of objective assessment which consider the subjective feelings for blur image evaluation. We not only concern some objective measure criteria like saturation, contrast, edge information and power spectral which based on natural statistical laws, but also effectively take in consider of image texture which sensitive to human vision, and a weighted evaluation function is made up. The least square nonlinear regress method is applied to fit the weight factor of each item.