Research On Lightweight CNN Loop Filtering Technology Based On AV1 Video Coding

Video compression is one of the core issues in the field of multimedia communications.The current topic of "5G+HD+AI" has triggered a new round of revolution in the field of video compression.Compared with traditional video coding methods,neural network-based video coding can learn signal features and internal associations from video big data.It can represent the video image signal more efficiently and compactly.Taking neural network-based in-loop filtering methods for example,compared with the traditional loop filtering algorithms,they have achieved more significant gains.However,it also brings super high complexity to the codec.For example,the neural network model has a huge number of parameters and long inference time.Moreover,the video content in practical applications is rich and diverse.It is difficult for the network models which are trained offline to adapt to various video changes.Consequently,its generalization ability is limited.In response to the above problems,this thesis proposes a low-complexity,image content-adaptive loop filter algorithm based on neural networks.It can improve the quality of the reconstructed image and the coding efficiency significantly.The main innovations of this article are as follows:1.This thesis proposes a Guided CNN loop filtering algorithm with low complexity to adaptive image content.Based on the unfiltered image,the algorithm proposes to construct a subspace through CNN.The projection of the original image signal in the subspace is used to approximate the original signal itself.The algorithm estimates the projection of the original image through the least square method in the constructed subspace.The CNN model used by the algorithm only contains 5k parameters,and uses strategies such as wide-dimensional activation to improve the learning ability of the model.Experimental results show that this method can bring 4.32%and 2.18% BD-Rate improvements for AV1 intra-frame and inter-frame coding,respectively.2.For the proposed Guided CNN algorithm,this article discusses and optimizes the coding block size,the number of channels,and the depth of the model.The optimized Guided CNN uses 5k parameters.Experimental results show that under the intra-frame and inter-frame encoding configurations,BD-Rate gains of 5.08% and 2.94%are obtained respectively.Moreover,this article carried out generalization experiments,ablation experiments and complexity analysis on the Guided CNN algorithm,which prove the robustness and superiority of Guided CNN.3.For very low bit rate coding scenarios,this thesis proposes to use CNN-based up-sampling coding.For high-resolution images,low-resolution images are obtained through traditional simple down-sampling methods for encoding.In the loop filtering stage,the low-resolution image is up-sampled through CNN and restored to highresolution.Under the condition that the aspect ratio is 2:1,compared with the traditional super-resolution model of AV1,experiments show that the proposed method can bring10.01% and 10.08% gains,respectively.