Research On Blind Image Restoration Based On Sparse Constraint

Improving the hardware performance for image system always can't afford to the requirement of practical application for the image definition and high resolution. Image system is affected by atmosphere turbulence, CCD under-sampling, system noise, and the relative motion between target object and camera. Thus the high frequencies details will lose when it is higher than the cut-off frequency of the image system and the image will blur. The method of restoring images can remove blur effectively. Recently, restoring image has a widely use in deep space detecting, military reconnaissance, and public security. From a mathematical point of view, restoring the sharp image from the blur image is an ill-posed problem. How to restore the true scene from the blur images? Not only does it need to restoring technology, but also the method of assessing quality of images. Because the assessment of images quality can provide the assessing method and index for the performance of removing images blur. And it can assess the advantages and disadvantages of reconstruction algorithm objectively. Finally, it will promote the perfection for the reconstruction algorithm.The thesis mainly works on two aspects of assessment of image and image blind deblur.(1) From the aspect of image blind deblurring. This thesis adopts the reconstruction method of the blind image based on the sparse prior constraints. This method mainly concludes two parts. The one is the assessment of muti-scale kernel based on the norm ratio constraints, the other is image restoration method based on the improved hyper-Laplacian. The detailed procedures are: Firstly. In order to suppress noises and accented edges, decompose the muti-scale image, and use the bilateral filter and shock filter filtering in each layer of the decomposed image. Secondly, use the prior information of ratio of l1/l2 as the regular constraint item of estimating the sharp image to replace kernel and sharp image to obtain the estimated kernel of relatively precised. Finally, adopt the regular constraint item which contains hyper-Laplacian priors one to restore the original image. In order to reduce the ringing effect in image restoration, this thesis begin to use the kernel which has been estimated to do the convolution processing in image to get the blur image before restoring clear image. Then use the different weighted values weight the edge region and central region. We propose adaptive kernel size to avoid blindlymanual set.(2) In the aspect of image quality assessment. Under normal conditions, The factors which makes image quality reduced mainly are the noise and blur. It will make for eliminate the distortion through selecting the appropriate reconstruction methods to assess the factors which reduces the image quality. This thesis also comes up with the no-reference image quality assessment method based on the local phase congruency. In the evaluation of the noise, the concrete steps are as follows. Firstly, a LPC map of degraded image is constructed and the image edge is extracted by modifying the noise threshold. Secondly, the edge is removed from the LPC map. Then, the noise level can be quantified by the remaining noise information and little “residual” information of the LPC map. During estimating the blur, As long as calculate the value of local phase coherence and the weight, then put the value in the expression of blur evaluation to compute the assess value of blur. The experiment proved that the method proposed by the thesis can assess the image quality objectively, and assessment result is accord to the subjective assessment results.For every image system, high quality input images are indispensible, in order to get excellent results, which shows the importance of image restoration and assessment. In this thesis, some generalization and further research based on it have been done, the experiments also show the effectiveness of the methos proposed.