Research On Intra Prediction Algorithm Of H.266/VVC Based On Neural Network

In recent years,thanks to the dividend of the Internet era,the great-leap-forward development of communication technology and the popularity of intelligent handheld terminals and monitoring equipment,video business shows an exponential explosive growth,which puts forward higher requirements for video coding technology.Based on the hybrid coding framework,in order to further improve the video compression ratio,the next generation video coding standard H.266/VVC has added new coding tools and technologies in several coding modules.Among them,Matrix-based Intra Prediction is adopted in the prediction module.As the first neural network-based coding technology received by VVC,MIP achieves effective intra prediction through off-line training matrix,which makes up for the deficiency of the traditional mode.However,there is still room for optimization in MIP technology in terms of algorithm unity and software and hardware overhead.In addition,other fusion technologies of neural networks and video coding have also become effective tools to improve coding performance.This paper optimizes and innovates H.266/VVC intra prediction based on matrix and fully connected network.The specific work is as follows:1.Optimization of Matrix-based Intra Prediction.In order to reduce the software and hardware overhead and improve the unity of the algorithm,this paper analyzes the problem of input vector dependence in the preprocessing process of MIP coding,and proposes an optimization algorithm based on the removal of input vector dependence.First of all,an optimization algorithm is proposed to remove the position dependence of the input vector on the vector elements.Based on the analysis of the dependence,the algorithm adjusts the corresponding matrix weight by fixing the operation position of the downsampling value,and the mapping relationship between integer matrix and floating point matrix is optimized.This step optimization algorithm realizes the partial unity of the preprocessing process,removes the dependence on the position of vector elements,and provides conditions for subsequent optimization algorithms.On this basis,an optimization algorithm is proposed to remove the dependence of the input vector on the classification index of the coding block,by transforming the pixel expected value to the downsampling vector element,and optimizing the corresponding matrix weight at the same time.In order to complete the unity of the preprocessing process under different classification.The optimization of this step further realizes the overall unity of the preprocessing process and removes the dependence on the classification index of coding blocks.The above two-step optimization algorithm eliminates the dependence of input vector on pixel position and coding block classification index,and then simplifies the construction process of input vector,realizes the unity of MIP algorithm,and is convenient for software and hardware implementation.2.Innovation of intra prediction algorithm based on fully connected network.In order to further combine neural networks with video coding technology,an intra prediction technique based on fully connected network is proposed in this paper.The coding techniques such as block partition,transform quantization and rate-distortion optimization are integrated into the process of network design and training,and a set of multi-mode lightweight intra prediction models based on fully connected networks are trained,which is deployed in VVC reference software.In the process of luma intra prediction,the algorithm competes with the traditional mode and has good coding performance,with the average BDrate saving of Y component in the whole intra mode being 0.12%.