No-reference Blur Image Quality Assessment Based On Visual Perception

In the process of image collection,the image quality will be affected by camera shake,compression,noise and other factors,among all the distortion types,blur is the most common one,filled with every corner of life,therefore,it is very important to assessment the quality of blur image.The quality assessment of blur image divided into three: full reference,reduced reference and no reference.The original image is not available in the real world,the no-reference image quality assessment algorithm does not need the original image information and has strong applicability.The papers studies the quality assessment of no-reference blur images.The main research contents are as follows:(1)In the no-reference image quality assessment of blur image,the less focus on image structure message and pixels and neighborhood pixels.the image structure information will change in the process of image blur.In this paper,local standard deviation is introduced to measure the change of blur degree,by computer gradient magnitude map of blur image,the input blur image is weighted by the local standard deviation value to get the local perceptual sharpness feature map,finally,the quality of blur image is evaluated by statistic information of pixel value in local perceptual sharpness map.In LIVE and CSIQ image databases,the performance index and time complexity of the proposed algorithm are better than other algorithms.(2)In order to evaluate the quality of blur image more effectively,a no reference blur assessment method based on gradient distortion measure map and salient region map is proposed.First,the reference image is constructed by Gaussian low pass filtering,and gradient similarity is included to obtain the gradient distortion measure map which can finely reflect potential tiny changes in textures and details.Then,the saliency model is utilized to calculate image saliency in which an adaptive method is adopted to calculate the specific salient threshold of blur image,and binarizes the blur image to yield salient region map.To this end,block-wise visual saliency is served as weights to get final image quality.Experimental results on LIVE,CSIQ,TID2008 and TID2013 databases demonstrate the proposed algorithm correlates well with human judgments than the existing traditional method;And increase the performance of BID and CID2013 in the real image database,compared with other algorithms,the performance index is in the middle position,which needs further improvement.meanwhile,its computational complexity is much lower.