End-to-End Lossless Stereo Image Compression Network
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for lossless stereo image compression either suffer from information loss due to non-linear transformations and quantization, or fail to fully exploit inter-view correlation, leading to sub-optimal coding performance.
Innovation Solution
An end-to-end lossless compression network is employed, which includes a multi-scale coding structure that derives multi-scale auxiliary representations and establishes hierarchical dependencies between the input stereo images and these representations, using an auto-encoder network to estimate probability distributions and perform entropy coding.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If non-linear transformations and quantization are applied for compression, then compression ratio is improved, but information loss occurs
Solution Approach 1:
The patent replaces traditional mechanical compression operations (non-linear transformations and quantization) with a learned end-to-end neural network compression model. This substitution allows the system to achieve compression while preserving lossless reconstruction capability, as the neural network learns optimal compression representations without the inherent information loss of quantization operations.
Solution Approach 2:
The patent changes the parameter representation by using learned probability distributions from the neural network instead of fixed quantization parameters. The entropy coding module uses these learned parameters to achieve efficient compression, while the reversible nature of the parameter transformations ensures no information loss during the compression-decompression cycle.
2Device complexity
If traditional compression methods are used, then implementation simplicity is maintained, but inter-view correlation is not fully exploited
Solution Approach 1:
The patent segments the compression task into specialized modules: a neural network compression model for learning representations, an entropy coding module for efficient bitstream generation, and a decomposition module for multi-scale analysis. This segmentation allows each module to be optimized for its specific function while working together to fully exploit inter-view correlation in stereo images.
Solution Approach 2:
The patent introduces an entropy coding module as an intermediary between the neural network compression model and the final bitstream. This intermediary efficiently encodes the learned representations by exploiting statistical correlations, particularly inter-view correlations in stereo images, thereby improving coding performance without compromising implementation feasibility.
Data Source
AI summary
A mechanism for processing video data is disclosed. The mechanism determines to apply an end-to-end lossless compression network to compress an input stereo image pair {xL, xR} into bitstreams {bL, bR}. A conversion can then be performed between a visual media data and a bitstream based on the end-to-end lossless compression network.