US20260261707
2026-09-03
Electricity
H04N19/593
The patent application introduces a method and apparatus for video coding that refines predicted signals of intra prediction using deep learning. The approach involves adaptively selecting a neural network-based prediction model from multiple pre-trained models based on specific block information. This information includes the intra prediction mode, block size, and quantization parameter of the current block. The method generates input data using reconstructed reference samples surrounding the current block and uses this data to produce a prediction block through the selected neural network-based model.
Video data inherently demands significant hardware resources due to its size compared to audio or still images, necessitating compression for efficient storage and transmission. Existing video compression techniques like H.264/AVC, HEVC, and VVC have achieved notable improvements in coding efficiency. However, with increasing image sizes, resolutions, and frame rates, there's a need for more advanced compression techniques. Traditional intra prediction methods rely on fixed rules that may not effectively handle complex content, prompting the exploration of deep learning technologies for enhanced prediction block generation.
Conventional intra prediction methods face limitations due to their reliance on fixed rules and limited reference sample information, often resulting in suboptimal prediction blocks for complex content. Deep learning-based methods have been proposed to address these issues, offering refined predictions by leveraging neighboring reference samples. However, these methods often lack adaptability and may not utilize existing prediction block information effectively. The proposed solution in the patent seeks to improve upon these limitations by integrating deep learning with adaptive model selection based on comprehensive block information.
The application details a video decoding device that includes an entropy decoder, intra predictor, predicted signal refinement unit, and an adder. The entropy decoder extracts block information and residual values from a bitstream. The intra predictor generates a prediction block using reconstructed samples and the intra prediction mode. The refinement unit adaptively selects a deep learning model based on block information to refine the prediction block. Finally, the adder combines residual values with the refined block to reconstruct the current block. This process enhances video coding efficiency and quality by leveraging deep learning for prediction refinement.
The disclosed method and apparatus aim to significantly enhance video coding efficiency and quality by refining intra predicted signals through adaptive deep learning techniques. This approach not only improves the accuracy of prediction blocks but also optimizes the use of existing block information, leading to better compression and video quality. It addresses the growing need for advanced compression technologies capable of handling the increasing demands of high-resolution and complex video content.