Asymmetrically distorted 3D video quality assessment: From the motion variation to perceived quality

摘要

The two-stage framework has been widely used in objective image/video quality assessment (IQA/VQA) methods, which includes local quality measurement and global quality pooling. Most of the existing IQA/VQA methods dedicate to local quality degradation modeling. Nevertheless, the significant pooling strategies are often neglected, especially in the quality assessment methods for asymmetrically distorted 3D video sequences. In this paper, motivated by the fact that the perceptual visual information might be affected by the visual masking due to motion suppression, a novel pooling strategy is proposed for visual quality assessment of asymmetrically distorted 3D video sequences by considering the motion variation-based perceptual visual information (MVPVI). In specific, to extract valuable visual information of video sequences, the proposed method computes the motion energy information (MEI) and edge energy information (EEI) by quantifying the frame difference map and the gradient map, respectively. Furthermore, in order to simulate the phenomenon of motion suppression in local motion scenes, where the perceptual information of low-attentional regions might be neglected by the human visual system (HVS), the proposed method first employs the coefficient of variation (CV) to determine the type of motion (local or global) in consecutive frames. Then, the perceptual visual information of the two cases is quantified. Finally, to simulate the role of HVS as an optimal information extractor, the estimated perceptual visual information of each single-view video sequence is used as the weighting factor for the quality pooling of left and right-view video sequences. Extensive comparison experiments conducted on Waterloo-IVC 3D video quality databases demonstrate that the proposed pooling strategy outperforms other relevant pooling strategies, and it achieves better performance than state-of-the-art 3D-IQA/VQA methods when combined with 2D-IQA/VQA methods.