Development of neural network-based video preprocessing method to increase the VMAF score relative to source video using distillation
Abstract:
In this work, we consider the problem of creating a video preprocessing method that improves video’s quality score measured by the Video Multimethod Assessment Fusion (VMAF) metric. The paper describes a neural network method for automatic preprocessing of input video, operating in real time. Preprocessing is carried out by a deep neural network based on U-Net architecture. In the course of network training, a trained VMAF approximation is used. The paper describes ways of improving the quality of the final method, namely, adding neural network compression, using SSIM in the loss function, and filtering the training set. The final version of the method increases the VMAF score of the original video by an average of 18% after preprocessing. The developed method demonstrates the flaws of the VMAF quality assessment method that can be used by developers of video processing algorithms to improve the ratings of their methods during automatic comparison carried out using VMAF quality assessment method.
Keywords:
video quality, robustness analysis, neural network distillation, adversarial attacks
