一种基于卷积神经网络的VVC去压缩伪影半盲方法

A Semi-blind Method for VVC Compression Artifact Reduction Based on CNN

多功能视频编解码(Versatile Video Coding,VVC)是新一代视频编解码标准,拥有较好的压缩性能,能够达到较高的压缩比。但是,编码过程中的变换、量化等操作,不可避免地在视频解码时引起一定程度的压缩伪影,导致解码视频质量降低,影响用户的视觉体验。目前,基于卷积神经网络(Convolutional Neural Network,CNN)的VVC的压缩伪影去除算法并不多,且大部分算法是在默认量化参数已知的情况下建立去伪影模型,对于不知道量化参数的盲场景,这些算法不太适合。直接设计全盲算法是复杂困难的,且性能有限。针对这一情况,提出了一种半盲方法用于去除VVC解码视频中的压缩伪影,该方法比全盲的方法更加灵活且能够达到更好的性能,比非盲方法更加实用。该方法设计出一种分类网络来预测重建视频的量化参数,预训练一些压缩伪影去除模型,根据预测的量化参数为重建视频选择对应的模型以去除压缩伪影。实验结果证明了该算法的有效性。

Versatile Video Coding(VVC)is a new generation video coding standard,which has good compression performance and achieves high compression ratio.However,the operations such as transformation and quantization in encoding inevitably cause some degree of compression artifacts in decoding,which leads to the degradation of decoded video quality and affects user’s visual experience.At present,there are not many algorithms for VVC compression artifact removal based on CNN,and most algorithms need to know quantization parameters to build compression artifact reduction model.For blind scenes with unknown quantization parameters,these algorithms are not suitable.It is complicated and difficult to design a blind algorithm directly,and its performance is limited.To address this issue,a semi-blind method is proposed to remove the compression artifacts of decoded videos for VVC.This method is more flexible and achieves better performance than the completely blind method,and is more practical than the non-blind method.This method designs a classification network to predict the quantization parameters of decoded videos,pre-trains some compression artifact reduction models,and selects the corresponding models according to the predicted quantization parameters to remove the compression artifacts.Experimental results show the effectiveness of the algorithm.