Feature Bitstream Transform Coding for Adaptive AI Decoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image compression technologies are unsuitable for artificial intelligence services, lacking efficiency and adaptability to machine-oriented tasks.
Innovation Solution
A feature encoding/decoding method and apparatus that abstracts a feature extraction network, enabling efficient encoding/decoding of feature information and adaptively responding to changes in the network, with a recording medium storing and transmitting bitstreams for independent operation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If existing image compression technology is used for human vision, then high-resolution and high-quality image processing is achieved, but it is unsuitable for artificial intelligence services
Solution Approach 1:
The patent segments the image processing pipeline into separate human-oriented compression and machine-oriented feature extraction components. The feature extraction network operates independently on the original image while the compressor handles the visual stream, allowing each to be optimized for its specific purpose without compromising the other.
Solution Approach 2:
The patent introduces feature maps as an intermediary representation between the original image and AI processing. These feature maps contain extracted semantic information that serves as a bridge, enabling AI services to operate on meaningful representations rather than raw compressed images.
2Productivity
If feature extraction network is changed or updated, then better AI task performance is achieved, but the encoding/decoding system must be reconfigured
Solution Approach 1:
The patent creates a universal feature encoding/decoding apparatus that can work with multiple different feature extraction networks. The encoder extracts features using various network architectures (CNN, RNN, Transformer) while the decoder reconstructs images, making the system adaptable to different AI tasks without requiring redesign.
Solution Approach 2:
The patent makes the feature extraction component dynamic and configurable. The system can adaptively select and switch between different feature extraction networks based on the specific AI task requirements, allowing the feature extraction stage to evolve independently from the compression stage.
3Measurement precision
If feature extraction network architecture is modified, then machine task accuracy is improved, but encoding/decoding efficiency decreases
Solution Approach 1:
The patent separates feature extraction from image compression into independent parallel processes. The feature extraction network can use complex architectures for high accuracy while the compression system maintains efficiency by working on the visual stream separately, eliminating the trade-off between accuracy and efficiency.
Solution Approach 2:
The patent merges the feature extraction output with the compressed image reconstruction process. The decoder uses both the compressed visual information and the extracted feature information to reconstruct the image, allowing efficient compression while maintaining high machine task accuracy through feature guidance.
Data Source
AI summary
Provided are a feature encoding/decoding method and device and a computer-readable recording medium generated by the feature encoding method. The feature decoding method according to the present disclosure, which is performed by the feature decoding device, comprises the steps of: obtaining information on transform of feature information and the feature information from a bitstream; and inversely transforming the feature information based on the information on the transform, wherein the information on the transform comprises information on a transform type that is used to transform the feature information.