Blurred Images Quality Assessment Based On Free Energy And Sparse Representation

Image quality has become a more and more important issue to be addressed in reality.The reason may be summarized into two aspects,one is that with the development of our life,the requirements of the image quality from people increase as well.On the other hand,the image quality can't be guaranteed during image processing or transmission.For example,the image is always compressed due to the bandwidth restrictions for transmission.Therefore,image quality assessment(IQA)is always used to monitor the image quality or employed as the performance measure for image processing algorithms.Generally,objective IQA methods are categorized into three classes on the basis of the access to the reference image,which are full reference(FR),reduced reference(RR)and no reference(NR)respectively.This paper is focused on RR IQA and the investigation of the quality assessment problem of a special kind of images,which refers to the blurred images.Specifically,we combine the free energy theory in neuroscience with the sparse representation to achieve this goal.Firstly,the free-energy principle indicates that the brain tries to account for the input image with an internal generative model and the discrepancy between the image and its model-explained version,which can be measured by free energy,is related to the image's perceptual quality.To approximate the free energy of an image and ease the calculation,we employ the common tools of sparse representation,which has been not only applied in different image processing fields but also inherits a biological support in neuroscience.Accordingly,we define a visual distance between the blurred image and its original image in free energy to evaluate the quality of the blurred image.Different from previous methods,such as edge information based and frequency domain based ones,this paper considers the internal generative model in our brain,which is well grounded in neuroscience.The proposed quality scheme belongs to reduced reference methods,which needs few information from the original image.Experiments are carried out on three public databases,that is,LIVE,TID2013 and CSIQ.It is demonstrated that the proposed method works in high consistency with subjective human scores and outperforms representative image quality assessment approaches.