Combining Visual Saliency and Binocular Energy for Stereoscopic Image Quality Assessment

With the flourishment of 3D content, the loss of quality of the stereoscopic images has been a large problem while being received by human beings. We develop a new metric in this paper to automatically assess the quality of stereoscopic images with the guidance of reference images. Visual saliency (VS) has been largely explored by researchers in the past decade to find out which areas of an image attract most attention of the viewers. We use the similarity of the VS map between original and distorted images as one of the quality-aware features since the degradation of VS map of the images can depict the quality loss in a certain degree. Meanwhile, gradient magnitude (GM) is enriched with image information, and GM similarity is exploited as another feature. While the difference of binocular energy between original and distorted versions reflects the severities of distortion, it can also act as weights between stereo pairs to simulate the binocular perception properties. Therefore, we introduce the difference of binocular energy as part of the features. The depth/disparity information between stereo pairs contains much properties of stereoscopic vision, and we extract features from disparity map. Finally, in order to take advantage of all the features, we utilize support vector machine based regression module to derive the overall quality score. Experimental results show that the proposed algorithm can assess the image quality in a manner of high consistency with human judgments.