Full Reference Image Quality Assessment Via Sparse Representation

Owning to the popularity of consumer electronics and the internet, thetechnology of digital image and digital video is developing fast. For them, imagequality assessment (IQA) is a crucial issue. IQA aims to estimate the quality of theimage. For practical application, IQA can be utilized as a benchmark to compare theperformance of different image processing algorithms. Also, IQA can be applied tooptimize the performance of video communication systems. To tackle the IQAproblem, different objective image quality assessment algorithms are proposed.The challenge of image quality assessment lies in two aspects:(1) formulatingperceptual meaningful features to effectively express the distortion level and (2)finding a way to pool them into a single quality score. To reduce these two problems,a two-step full reference IQA algorithm is proposed in this paper. In the first step, theimage is divided into patches. Local quality assessment for each patch is extracted asfeature to express the distortion level of the image. Motivated by the research ofsparse representation model and human striate cortex, we believe that the sparserepresentation model can effectively model the statistics of the local image patch andthe sparse representation model can extract features effectively. Thus, in the first step,the sparse representation model is employed to extract the perceptual meaningfulfeatures for the IQA problem. In the second step, machine learning technique isemployed to mimic the complex characteristics of human vision system which poolsthe local distortion measurements obtained into a single image quality score. Morespecifically, the kernel ridge regression (KRR) is adopted to learn the relationshipbetween local quality measurements of the patches and the quality score of the wholeimage. By this way, once the local quality assessments for the test image are obtained,the objective image quality assessment for the test image can be calculated by theKRR model.To validate the effectiveness of the proposed algorithm, experiments are performed on five independent image quality assessment databases, e.g., LIVE,TID2008, IVC, CSIQ, and Toyama-MICT. Experimental results demonstrate theeffectiveness of the proposed algorithm and show that the proposed algorithm iscompetitive or even better than most state-of-the-art image quality assessmentalgorithms.