Neural Network-Based Intra- and Inter-Frame Prediction Algorithms

The increasing demand for high-resolution, high-frame-rate video has intensified the complexity of traditional coding standards, hindering real-time encoding. We propose a unified CNN-based model for intra- and inter-prediction that shares a common architecture with separate parameter sets. Experiments show that our model reduces encoding time by 96.67%, incurring a BD-BR increase of 9.59% and a BD-PSNR loss of 0.331 dB; for all-intra tasks, it achieves a 93.98% time reduction with a 6.13% BD-BR increase and 0.287 dB BD-PSNR loss. A Verilog-based hardware implementation and simulation validate its feasibility for hardware deployment.