视频质量评价主要采用传统的手动提取特征, 再利用机器学习预测视频质量分数, 导致结果不理想。VGG-16网络在特征提取方面具有非常好的稳健性, 借鉴其网络模型, 迁移参数构造出用于端到端的视频质量评价网络。LIVE视频数据库的实验结果表明, 该方法预测的评价分数与主观评价分数具有较高的一致性。其评价指标斯皮尔曼等级相关系数和皮尔逊线性相关系数分别达到了0.867和0.843, 性能优于目前基于手动提取特征进行视频质量评价的大部分算法。

In video quality assessment, most researchers manually extract the features first, and then use machine learning to predict video quality score, which leads to unideal result. Since the VGG-16 net has excellent robustness in feature extraction, we use the network model and migrate parameters to construct the end-to-end video quality assessment network. The experimental results on LIVE video database show that the assessment score of this method is consistent with the subjective assessment score, and its assessment indexes of Spearman rank correlation coefficient and Pearson correlation coefficient reached 0.867 and 0.843, respectively, which indicated that the performance of the proposed method is better than most of the current video quality assessment algorithms based on manual feature extraction.