No-Reference Stereoscopic Video Quality Assessment Based on Spatial-Temporal Statistics

Stereoscopic video quality assessment (SVQA) has become the necessary support for 3D video processing while the research on efficient SVQA method faces enormous challenge. In this paper, we propose a novel blind SVQA method based on monocular and binocular spatial-temporal statistics. We first extract the frames and the frame difference maps from adjacent frames of both left and right view videos as the spatial and spatial-temporal representation of the video content, and then use the local binary pattern (LBP) operator to calculate spatial and temporal domains' statistical features. Besides, we simulate binocular fusion perception by performing weighted integration of generated monocular statistics to obtain binocular scene statistics and motion statistics. Finally, all the computed features are utilized to train the stereoscopic video quality prediction model by a support vector regression (SVR). The experimental results show that our proposed method achieves better performance than state-of-the-art SVQA approaches on three public databases.