Integerized Interpolation Filters for Accurate Video Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video coding technologies face inefficiencies due to floating-point operations in interpolation filter design, leading to inconsistent results across different computing architectures and increased computational costs, which are exacerbated by the loss of desirable filter properties during integerization using rounding methods.
Innovation Solution
A method for integerization of interpolation filters that minimizes interpolation error by generating filter candidates based on ceiling and floor values of floating-point coefficients and selecting the filter with the lowest error metric, preserving desirable filter properties.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If floating-point interpolation filters are used for video coding, then interpolation accuracy is improved, but computational complexity and inconsistency across different computing architectures increases
Solution Approach 1:
The patent replaces floating-point arithmetic operations with integer arithmetic operations in the interpolation filter. This substitution maintains the essential filtering function while eliminating the computational complexity and architectural inconsistency associated with floating-point operations. The integerized filter coefficients are designed to provide accurate interpolation results using only integer math, thereby resolving the contradiction between precision and computational burden.
2Measurement precision
If floating-point interpolation filters are used for video coding, then interpolation accuracy is improved, but consistency across different computing architectures deteriorates
Solution Approach 1:
The patent changes the numerical representation parameter of the filter coefficients from floating-point to integer format. This parameter change ensures consistent behavior across different computing architectures because integer arithmetic is deterministic and architecture-independent, unlike floating-point arithmetic which can exhibit variations due to different IEEE 754 implementations and rounding behaviors. The integerized coefficients maintain interpolation accuracy while guaranteeing architectural consistency.
3Device complexity
If simple integerization methods are applied to interpolation filters, then computational complexity is reduced, but prediction accuracy deteriorates
Solution Approach 1:
The patent applies preliminary action by pre-computing and optimizing the integerized filter coefficients during the filter design phase. Instead of using simple truncation or rounding during runtime, the coefficients are carefully designed in advance to minimize quantization error and preserve the desirable properties of the original floating-point filters. This preliminary optimization ensures that the integerized filter achieves both low computational complexity and high prediction accuracy.
Solution Approach 2:
The patent employs feedback in the filter design process by evaluating the performance of integerized coefficients against the original floating-point reference and iteratively adjusting them to minimize the error. This feedback mechanism ensures that the final integerized filter coefficients maintain prediction accuracy comparable to the floating-point version while enabling efficient integer-only computation.
Data Source
AI summary
A video encoder encodes a video into a video bitstream. The video encoder accesses a set of frames of the video and performs inter prediction for the set of frames using a set of integerized interpolation filters to generate prediction residuals to be encoded into the video bitstream. The set of integerized interpolation filters are generated by integerizing a set of interpolation filters, each of the set of interpolation filters having floating-point filter coefficients. For each interpolation filter, two integerized filter coefficient values are generated for each filter coefficient and a set of filter candidates are generated based on the two integerized values for each filter coefficient. An error metric for each filter candidate is calculated and an integerized interpolation filter having the lowest error metric is selected for the interpolation filter from the set of filter candidates.