Research On VVC Deblocking Filtering Algorithm Based On Convolutional Neural Network

On July 7,2020,the Fraunhofer Heinrich Hertz Institute in Germany officially announced the latest generation of H.266/Versatile Video Coding(VVC).Compared with the previous generation H.265/High Efficiency Video Coding(HEVC),the VVC compression efficiency has increased by more than 30%.However,VVC still uses the traditional block-based hybrid coding structure,and it will inevitably encounter block artifacts.Research on the Deblocking Filter(DF)method can effectively reduce the blocking artifact and improve the quality of the reconstructed video image.On the one hand,the traditional deblocking filtering method significantly reduces the blocking artifact,but in the face of increasingly diversified video image content,it is impossible to make intelligent decisions to achieve optimal filtering.On the other hand,due to the dual effects of technological breakthroughs and industry promotion,deep learning has achieved great success in the field of computer vision,especially in terms of classification and prediction.Therefore,this thesis uses these two capabilities of neural networks to design two neural network models to improve the quality of VVC reconstruction images in different coding scenes.The main work of this thesis is as follows:1.The performance of the VVC deblocking filtering algorithm under general test conditions was evaluated,and finds that after deblocking filtering,the quality of the video sometimes drops.After analysis,it is found that the unreasonable threshold value in the deblocking filtering algorithm leads to the over-filtering.Therefore,this thesis designs a four-layer lightweight convolutional neural network,the first two layers are convolutional layers,and the last two layers are fully connected layers.Using the convolutional neural network's ability to learn features of massive data,comprehensively extract feature information of multiple scenes to complete the training of the model.This thesis embeds the offline trained model into the deblocking filtering process,replacing the original fixed threshold decision method,and improving the subjective and objective quality of coding.2.The performance of the VVC deblocking filtering algorithm at low bitrates was evaluated,and finds that for video sequences with chaotic motion and complex textures,the reconstructed image will have serious blocking artifacts.Through analysis,it is found that the image information is seriously lost due to quantization.The VVC deblocking filter algorithm designed based on the idea of spatial interpolation filtering cannot effectively reduce the blocking artifacts.Therefore,with the help of the idea of super-resolution,this thesis proposes a deep learning-based deblocking filtering post-processing module,which can effectively reduce the blocking artifacts and improve the image quality.Specifically,this thesis designs a four-layer convolutional neural network.The network not only inputs image information to be processed,but also inputs coding unit(CU)division information as important feature information into the network.Experimental results show that this model significantly improves the image reconstruction quality with a certain increase in coding complexity.