Research On Blur Image Restoration And Assessment Method

Digital image restoration and quality assessment are the hotspots in the field of image processing, and are playing increasingly important role in real life. Blur is one of the most common image distortion types. There are many reasons to produce blur images, such as deficiency of imaging system, relative motion between camera and objects and so on. Due to blur image prevailing in various image processing systems, some research has been done on the field of blur image restoration and quality assessment in this paper.Firstly, according to the nonlinear characteristics of neural network, an image restoration method combining region category and back propagation (BP) neural network is presented. Initially, images are divided into three parts, edge region, texture region and smooth region. Then the regions of the training images with the same category form three final training sets. For each training set, a new BP neural network is trained using the set. At last three neural networks are used to restore different regions of the test images. Experimental results show that proposed method can effectively restore the blur images.Secondly, by using blur comparison, a new no-reference blur index for still images is proposed. The key idea is based on the characteristic of human visual perception that we have difficulties in differentiating the same two images with closely blur degree. In new approach, a new re-blurred image is produced by adding blur to the original image. Then for the original image, the local areas around edge points are chosen. According to those local areas, the corresponding statistics in the original and re-blurred images are calculated and summed up respectively. At last the two sums are used to construct a new blur metric. Experimental results on different blur databases show that the proposed method can obtain higher correlations with the subjective quality evaluations.Finally, a novel no-reference blur metric combining characteristics of human visual system (HVS) with structure similarity (SSIM) is proposed with further researching the characteristics of human visual system in blur detection. In the new method, the re-blurred image is produced by convoluting the original image with a gaussian filter. Then a collection of strong edge block around edge point is created. For each pixel block, the SSIM index of it and corresponding pixel block in re-blurred image is calculated. The new blur metric is taken as the average of the SSIM indices. The experimental results demonstrate the effectiveness of the method.