基于3D卷积神经网络的无参考视频质量评价

无参考视频质量评价（NR-VQA)在无法获得原始高质量视频参照的前提下，对失真视频的视觉质量进行定量度量.常规NR-VQA方法通常针对特定失真类型设计，或者与人的主观感受存在偏差.首次将3D深度卷积神经网络（3D-CNN)引入到了视频质量评价中，提出了一种基于3D-CNN的无参考视频质量评价方法，可以适用于非特定失真类型的NR-VQA.首先，通过3D块来有效学习和表征视频内容的时空特征.其次，对常规的3D卷积网络模型进行改进，使其适用于视频质量评价的任务.实验结果表明，所提出的方法在多种失真类型和多个测试指标上，与人的主观感知一致性较高.作为无参考视频质量评价方法，其性能与许多全参考评价方法具有可比性，同时比主流的NR-VQA方法具有更快的运行速度，这使得所提模型在实际中具有更好的应用前景.

No Reference Video Quality Assessment Based on 3D Convolutional Neural Network

No reference video quality assessment (NR-VQA) measures distorted videos quantitatively without the reference of original high quality videos. Conventional NR-VQA methods are generally designed for specific types of distortions, or not consistent with human''s perception. This paper innovatively introduces 3D deep convolutional neural network (3D-CNN) into VQA and proposes a 3D-CNN based NR-VQA method, which is universal for non-specific types of distortions. First, the proposed method utilizes 3D patches to learn spatio-temporal features that represent video content effectively. Second, the original 3D-CNN model is modified which is used to classify videos to make it adapt to VQA task. Experiments demonstrate that the proposed method is highly consistent with human''s perception across numerous distortions and metrics. Compared with other state-of-the-art no-reference VQA methods, the proposed method runs much faster while keeping the similar performance. As a no-reference VQA method, it is even comparable with many of the state-of-the-art full-reference VQA methods, which provides the proposed method with better application prospects.