H.264 Quantization Using 16-bit Segmented Arithmetic
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Recent video compression methods, such as H.264, face challenges in reducing computational complexity while maintaining performance, particularly in the implementation of quantization processes that require 32-bit arithmetic, which can be impractical on processors without such memory access.
Innovation Solution
The implementation of simplified 16-bit operations for H.264 video coding by modifying quantization tables based on quantization parameters, reducing the need for 32-bit accesses and using constant delta values to simplify forward quantizations, thereby enabling efficient quantization and inverse quantization processes for 4×4 and 2×2 block sizes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If 32-bit arithmetic operations are used for H.264 quantization, then compression efficiency and accuracy are improved, but device complexity and memory access requirements increase
Solution Approach 1:
The patent divides the 32-bit quantization operation into multiple 16-bit stages. The quantization process is segmented into: (1) initial 16-bit multiplication with quantization matrix, (2) addition of offset, (3) right shift operation, and (4) final rounding. This segmentation allows processors without 32-bit memory access to perform quantization using only 16-bit operations, resolving the contradiction between accuracy and device complexity.
Solution Approach 2:
The patent introduces intermediate variables and temporary storage locations to facilitate the multi-stage 16-bit computation. Intermediate results are stored in 16-bit registers and memory locations, acting as mediators between the input data and final quantized output. This intermediary approach enables accurate quantization computation while maintaining compatibility with 16-bit processor architectures.
2Productivity
If 32-bit memory access is implemented, then quantization performance is improved, but ease of operation and implementation feasibility deteriorate
Solution Approach 1:
The patent employs dynamic right shift operations that adapt based on the quantization parameter (QP) value. The shift amount is determined dynamically during encoding, allowing the system to optimize between compression ratio and computational complexity in real-time. This dynamic approach maintains high productivity across different operating conditions while ensuring ease of implementation on 16-bit processors.
Solution Approach 2:
The patent changes the parameter representation from 32-bit fixed-point to 16-bit fixed-point with adjusted scaling. By modifying the quantization parameter representation and using scaled integer arithmetic, the system achieves comparable quantization performance without requiring 32-bit memory access, thereby improving implementation feasibility.
3Ease of operation
If simplified 16-bit operations are used, then ease of operation and device compatibility are improved, but measurement precision and quantization accuracy may deteriorate
Solution Approach 1:
The patent applies preliminary scaling and offset addition before the final quantization step. By pre-processing the transform coefficients with scaling factors and offset values, the system ensures that the subsequent 16-bit right shift operation produces accurate results. This preliminary action compensates for the reduced precision of 16-bit operations while maintaining compatibility with simple processors.
Solution Approach 2:
The patent incorporates feedback mechanisms in the inverse quantization process, where the quantized coefficients are used to adjust subsequent processing steps. The reconstructed block from 16-bit operations feeds back into the prediction and transformation processes, allowing the system to maintain overall accuracy despite using simplified 16-bit arithmetic throughout the pipeline.
Data Source
AI summary
Low complexity (16 bit arithmetic) H.264 video compression replaces a single quantization table for all quantization parameters with multiple quantization tables and thereby equalizes quantization shifts and round-off additions; this eliminates the need for 32-bit accesses.