Prediction-Dependent Residual Scaling for LMCS Video Coding Range Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The current design of Luma Mapping with Chroma Scaling (LMCS) in video coding standards like VVC faces issues with mapped precision of inter prediction samples exceeding the dynamic range of the internal coding depth, leading to inefficiencies.
Innovation Solution
Implementing Prediction Dependent Residual Scaling (PDRS) by converting prediction samples and residual samples into a mapped domain, applying inverse mapping scaling factors, and then converting back to the original domain, while avoiding clipping operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If inter prediction samples are mapped to extended dynamic range (HDR) domain, then video quality is improved, but the mapped precision exceeds the dynamic range of internal coding depth causing overflow and clipping errors
Solution Approach 1:
The patent applies a piecewise linear mapping function that transforms samples from standard dynamic range (SDR) to HDR domain by dividing the input range into multiple segments and applying different linear transformation parameters to each segment. This dimensional transformation allows the mapped samples to occupy a wider dynamic range without causing overflow, as each segment is carefully designed to map within valid output bounds.
Solution Approach 2:
The patent changes the coding parameters by introducing separate mapping parameters (slope and intercept) for different segments of the dynamic range. By adjusting these parameters piecewise across different input ranges, the system can expand the dynamic range for video quality improvement while ensuring that the mapped values remain within the representable precision limits of the internal coding depth.
2Stability of the object's composition
If clipping operations are applied to prevent overflow in mapped domain, then coding stability is maintained, but video quality and compression performance deteriorate
Solution Approach 1:
The patent performs preliminary scaling of the mapped domain samples by inverse mapping scaling factors before adding residuals. This preliminary action ensures that the scaled samples are already within the valid dynamic range, making clipping operations unnecessary and preserving video quality. The inverse mapping is applied in advance to prevent overflow rather than correcting it after the fact.
Solution Approach 2:
The patent introduces inverse mapping scaling factors as intermediary values that bridge the mapped domain and the residual domain. These scaling factors act as a mediator to transform the mapped samples into a range that can be safely combined with residuals without causing overflow, thereby eliminating the need for clipping operations while maintaining coding stability.
3Measurement precision
If prediction samples are scaled by inverse mapping factors before residual addition, then dynamic range accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent applies different inverse mapping scaling factors to different segments of the dynamic range, rather than using a single uniform scaling factor. This local quality approach allows the system to achieve high dynamic range accuracy where needed while using simpler scaling for other regions, thereby balancing computational complexity with precision requirements across different parts of the signal.
Solution Approach 2:
The patent applies inverse mapping scaling to only the necessary components (prediction samples) before residual addition, rather than transforming the entire signal pipeline. This partial action approach achieves the required dynamic range accuracy at the critical point of residual combination while avoiding unnecessary computational overhead in other parts of the encoding/decoding process.
Data Source
AI summary
Methods and devices are provided for rectifying a forward mapping coding bit length issue introduced by LMCS. In one method, a plurality of prediction samples, in a mapped domain, of luma component of a CU that is coded by a CIIP mode under LMCS framework is obtained, a plurality of residual samples, in the mapped domain, of the luma component of the CU is obtained, the plurality of prediction samples in the mapped domain is added to the plurality of residual samples in the mapped domain, resulting in a plurality of reconstructed samples, in the mapped domain, of the luma component, and the plurality of reconstructed samples of the luma component is converted from the mapped domain into an original domain based on a pre-defined plurality of inverse mapping scaling factors.