No-reference video quality evaluation method based on deep learning

Abstract

The invention discloses a non-reference video quality evaluation method based on deep learning. The method adopts a pre-trained convolutional neural network, a bidirectional GRU network fusing time domain features and a video quality prediction network. The method comprises the steps that: firstly, the pre-trained convolutional neural network extracts content sensing features of each frame, global pooling is used for processing the features, redundant information is discarded, and change information is stored. secondly, dimensionality reduction is carried out on the features by adopting a full connection layer, and the time domain features before and after are fused by using a bidirectional GRU network; and finally, a frame quality score is calculated by using a full connection layer, and then the overall video quality is aggregated to generate a prediction score. The network model provided by the invention fully and effectively considers the content dependence and the time lag effect of the video to realize the evaluation of the video quality score.

Description

technical field [0001] The invention belongs to the field of computer vision, and in particular relates to a no-reference video quality evaluation method based on deep learning. Background technique [0002] Video quality evaluation refers to the perception, measurement and evaluation of the changes and distortions of two pieces of video information with the same main content through specific methods. It has important applications in the fields of video processing, video quality monitoring and multimedia video applications. There are two types of video quality evaluation methods: subjective evaluation and objective evaluation. Subjective evaluation means scoring by means of human visual observation. It can be said to be the method that best reflects the audience's perception of video quality, and it is also the ultimate goal of other objective evaluation methods. However, the subjective evaluation method consumes a lot of manpower and time, and is not suitable for large-sca...

AI Technical Summary

Problems solved by technology

[0005] Although the existing no-reference video quality evaluation methods have achieved certain achievements, they still face various challenges. For example, in some databases, the training samples are not balanced and the number is insufficient, which brings certain difficulties to the training model; There are various types of video distortion, and the standard no-reference method is designed for a specific type of distortion, which has limitations and lacks generality; for natural distortion video databases, it is difficult to achieve better evaluation results, etc.

Images