Video Coding Transform Index Probability Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional video coding schemes face inefficiencies in modeling statistical variations in residuals due to the use of a single fixed transform, leading to increased bitrate costs in encoding the transform index for Rate Distortion Optimization (RDO) searches.
Innovation Solution
A method and device for picture encoding and decoding that determine the probability of each transform being used for encoding a picture block, using a convolutional neural network to select and encode an index efficiently, thereby reducing bitrate by employing a truncated unary code for the transform index.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If multiple transforms are used to model statistical variations in residuals, then compression efficiency is improved, but bitrate increases due to encoding additional transform index syntax elements
Solution Approach 1:
The patent applies preliminary action by using a convolutional neural network to predict the most likely transform index before encoding. This prediction allows the encoder to use truncated unary coding with fewer bits for the most probable transforms, reducing the average bitrate cost while maintaining the benefit of multiple transforms for modeling statistical variations in residuals
Solution Approach 2:
The patent changes the parameter representation by using probability-based truncated unary coding instead of fixed-length coding for transform indices. By dynamically adjusting the code length based on predicted probabilities from the neural network, the system optimizes the trade-off between compression efficiency and bitrate consumption
2Measurement precision
If exhaustive Rate Distortion Optimization search is performed to select the best transform, then coding accuracy is improved, but computational complexity increases
Solution Approach 1:
The patent performs preliminary action by using a convolutional neural network to pre-predict the optimal transform index before the exhaustive RDO search. This prediction guides the RDO process, allowing the system to maintain coding accuracy while reducing the effective search space and computational complexity by focusing on the most probable transform candidates
Solution Approach 2:
The patent introduces an intermediary convolutional neural network between the residual block and the RDO search process. This neural network intermediary provides probability estimates that guide the subsequent exhaustive search, effectively mediating between rapid prediction and thorough optimization to balance accuracy and complexity
Data Source
AI summary
A decoding method is disclosed. The decoding method comprises:determining an index from a bitstream, the index identifying a coding element in a set of a plurality of coding elements;decoding a picture block from the bitstream using the coding element of the set identified by the index.In a specific embodiment, determining the index comprises:determining, for each coding element in the set, a probability that the coding element was used for encoding the picture block; anddetermining the index responsive to the determined probabilities.