No-Reference Video Quality Assessment Using Statistical Features Along Temporal Trajectory

Abstract

In this paper, we propose a novel no-reference (NR) video quality assessment (VQA) algorithm. First, each frame in the distorted video is transformed into wavelet domain and decomposed to form oriented band-pass responses. The obtained subband coefficients are then utilized to extract a series of statistical features of distortions. These statistical features are stacked to form a vector, which is a statistical description of all the distortions in the frame. We utilize the feature vector across images to perform classification and mapping to quality scores, and then a combination is performed to obtain the frame quality in wavelet domain. Next, to evaluate the temporal quality, we propose a motion-compensated approach based on block and motion vector. Finally, the quality of each frame is pooled along temporal trajectory to obtain the overall quality of the distorted video. The proposed algorithm is tested on LIVE video database and the result shows that it outperforms the full-reference (FR) Peak Signal-to-Noise Ratio (PSNR), and is indeed competitive with the modern popular Structural Similarity (SSIM).