Prediction Residual Vector Quantization Beyond Transform Coding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video compression techniques face inefficiencies in transform search processes, particularly for prediction residuals with high frequencies, leading to high bitrates and prolonged processing times.
Innovation Solution
Implementing a codebook-based approach that bypasses transforming prediction residuals to the frequency domain by using a lookup table to find the closest matching prototypical prediction residuals, encoded via a gain and shape index, reducing the need for transform coding.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If transform coding is used for prediction residuals with high frequencies, then encoding accuracy is improved, but bitrate increases and processing time is prolonged
Solution Approach 1:
The residual block is segmented into multiple sub-blocks, and different transform types are applied to different sub-blocks based on their local frequency characteristics. This allows accurate encoding of high-frequency regions while using simpler encoding for low-frequency regions, thereby maintaining encoding accuracy while reducing overall bitrate.
Solution Approach 2:
The transform type is dynamically selected for each sub-block based on local characteristics such as gradient magnitude and frequency content. This dynamic adaptation enables the encoder to apply appropriate transform complexity where needed, improving accuracy for high-frequency components while avoiding unnecessary processing for low-frequency components, thus reducing bitrate.
2Measurement precision
If transform coding is used for prediction residuals with high frequencies, then encoding accuracy is improved, but processing time is prolonged
Solution Approach 1:
The residual block is divided into sub-blocks, allowing parallel processing of different regions. High-frequency sub-blocks undergo full transform coding while low-frequency sub-blocks use simplified coding, reducing overall processing time while maintaining accuracy where needed.
Solution Approach 2:
The transform process is dynamically adjusted based on local frequency characteristics. Regions requiring high accuracy (high-frequency areas) receive full transform processing, while other regions use faster simplified coding, thereby reducing total processing time while preserving encoding accuracy in critical areas.
3Measurement precision
If transform search is performed for all blocks, then encoding accuracy is improved, but device complexity increases
Solution Approach 1:
The block is segmented into sub-blocks, and transform search is performed only on sub-blocks with high-frequency characteristics. This selective approach maintains encoding accuracy for complex regions while avoiding unnecessary search complexity in simple regions, thereby reducing overall device complexity.
Solution Approach 2:
The transform search process is dynamically applied based on local block characteristics. High-frequency sub-blocks undergo comprehensive transform search to ensure accuracy, while low-frequency sub-blocks skip or simplify the search process, reducing device complexity while maintaining accuracy where required.
Data Source
AI summary
A gain index representing a magnitude of a residual block for a current block is decoded. A shape index identifying a unit-norm vector that represents a pattern of the residual block is decoded. The residual block is decoded based on the gain index and the shape index. Decoding the residual block may include multiplying a gain value selected using the gain index by the unit-norm vector selected using the shape index to obtain the residual block. A sign bit value may be decoded and decoding the residual block may be further based on the sign bit value. The current block is reconstructed based on the residual block.