Image Quality Assessment Considering Visual Saliency

Developing image quality assessment algorithms which consider the Human VisualFeatures has been the main trend nowadays. Visual Saliency, the high-level feature, isnow being integrated into the assessment progress. Existing results have shown thatadding the saliency maps which are obtained from eye-tracking experiment can improvethe performance of image quality assessment algorithms. However, due to thecomplexity of Human Visual System, there are no algorithms that can model visualsaliency accurately and output saliency maps which are similar to eye-tracking data.To study the difference between the predicted saliency maps and the ground-truthand to have a better knowledge of how visual saliency influence the quality assessmentperformance, in this thesis, firstly, an eye-tracking experiment is conducted, obtainingthe saliency maps of all images in LIVE database. Then, four state-of-the-art saliencymodels are selected to compare with the eye-tracking data, proving that the existingmodels are of poor quality and cannot reflect the visual saliency features of human eyes.The comparative study further proves that the added value brought by saliencymodel in assessment algorithms rises with the improvement of model accuracy. What’smore, the experiment also show that the added value is image-content based. Therefore,whether or not integrating saliency models into assessment algorithms should considerthe image content. By doing so, we can maximize the performance gain with a minimalcomplexity.Based on the above conclusions, a visual saliency model based on contrast isproposed. This model only use luminance contrast feature extracted by GaussianPyramids, resulting in a lower complexity. Then the proposed model is adaptivelyintegrated into the SSIM and PSNR algorithms. The experiment results show that bothtwo algorithms obtain a performance gain which is the same compared with the addedvalue brought by eye-tracking experiment.