Humans pay more attention to the regions of interest in an image,so distortion in those areas is more likely to affect their subjective quality scores.However,traditional image quality assessment(IQA)methods do not take into account the difference in the attention received by different regions in the image,resulting in a lower degree of fitting between the predicted score and the subjective quality score.Aimed at the above problems,a saliency-map-based dual-stream image quality assessment(SDS-IQA)method to highlight the area of interest in the image was proposed.A dual-stream hierarchical structure composed of a saliency map branch and an original image branch was used to realize multi-scale distortion perception of images from both the whole and the focus,reflecting the importance difference of features in all dimensions through dual attention.In the feature extraction stage,SDS-IQA used spatial attention in the saliency-map branch to reflect the difference in attention in the scale space,strengthened the distortion information expression of the original image branch through spatial attention weight,and used gated attention to strengthen the interaction between channels in the feature fusion stage,so that the attention difference between channels could be reflected during fusion and the key characterization of distortion in the region of interest was eventually realized.Ex-perimental results show that the Pearson linear correlation coefficient of this method reaches 0.976,0.896 and 0.865 on three synthetic datasets(LIVE,TID2013,CSIQ),and 0.869 and 0.877 on two authentic datasets(LIVEC,KonIQ-10k),respectively,proving that the prediction results of SDS-IQA have good fit with human subjective as-sessment.
维普
万方数据