基于高阶深层时空信息的自媒体视频质量评价

No-reference Video Quality Assessment Based on High-order Deep Spatio-temporal Information

自媒体视频中普遍存在的失真扭曲现象,给视频质量评价带来了新的挑战。视觉感知理论研究表明,人类视觉系统中存在感知迭代机制,即视频的感知评价是一个正反向迭代修正的过程。受此启发,本文将该机制引入视频质量评价,提出了基于高阶深层时空信息的质量评价方法。具体而言,本文提出了二阶协方差聚合来进行高阶空域信息的提取,引入快速迭代GRU结构进行深层时域信息建模,而后通过特征层的池化聚合以及多层感知机回归视频得分。实验结果表明,预测结果与人类主观质量评分具有较好的一致性,明显优于现有无参考质量评价算法。

The common distortion phenomenon in self-media videos brings new challenges to video quality assessment.Research on the theory of visual perception shows that there is an iterative mechanism of perception in the human visual system. That is,the perceptual evaluation of the video is a process of forward and backward iterative correction. Inspired by this,this paper introduces this mechanism into video quality assessment and proposes a quality evaluation method based on high-order deep Spatio-temporal information. Specifically,this paper proposes second-order covariance aggregation to extract high-order intra-frame information,introduces a fast iterative GRU structure for deep inter-frame information modeling,and then uses feature layer pooling aggregation and multi-layer perceptron regression to get the video score. Experimental results show that the prediction results are in good agreement with human subjective quality scores,which are significantly better than the existing no-reference quality assessment algorithms.