Image Quantization Selection for Human and Machine Perception
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image compression technologies are not optimized for machine tasks, leading to inefficiencies in processing large amounts of image data required for artificial intelligence services.
Innovation Solution
An image encoding/decoding method that performs quantization optimization based on the image purpose, selecting a quantization method suitable for machine tasks, and encoding/decoding images for human or machine perception, with a focus on Video Coding for Machine (VCM) bitstreams.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing image compression technologies are used, then high-resolution and high-quality image processing is achieved for human vision, but the technologies are not suitable for artificial intelligence services
Solution Approach 1:
The patent changes the quantization parameters based on image purpose (human perception vs. machine perception). For machine perception tasks, the system uses different quantization settings that optimize for machine learning model performance rather than human visual quality, thereby adapting the compression technology to different application scenarios
Solution Approach 2:
The system dynamically selects quantization methods based on the specific machine task requirements. The quantization parameters are not fixed but are adjusted according to the image purpose information, allowing the compression technology to adapt its behavior based on whether the output is for human viewing or machine processing
2Measurement precision
If quantization is optimized for human perception, then image quality is improved, but encoding/decoding efficiency for machine tasks is reduced
Solution Approach 1:
The patent applies different quantization parameters based on the target audience. When the image is intended for machine perception, the system uses quantization parameters that balance compression efficiency with machine learning performance requirements, rather than optimizing solely for human visual quality
Solution Approach 2:
The system segments the image processing into different paths based on purpose. By identifying whether the image is for human or machine perception through purpose information, the system applies appropriate quantization methods to each segment, improving overall efficiency by avoiding unnecessary optimization for human vision when machine tasks are the priority
3Device complexity
If a single quantization method is used for all images, then device complexity is reduced, but encoding/decoding efficiency is compromised
Solution Approach 1:
The system dynamically selects from multiple quantization methods based on image purpose information rather than using a fixed single method. This dynamic selection allows the system to achieve high encoding/decoding efficiency for machine tasks while maintaining manageable complexity through automated decision-making
Solution Approach 2:
The system uses feedback from image purpose information to select the appropriate quantization method. By incorporating purpose information into the quantization decision process, the system achieves efficient encoding/decoding performance tailored to specific machine learning tasks without requiring manual intervention or overly complex configuration
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present disclosure relates to an image encoding/decoding method and device for optimizing quantization according to image use, a method for transmitting a bitstream, and a bitstream storage medium. The image data encoding method according to one embodiment of the disclosure comprises the steps of: acquiring, from a bitstream, image use information for image data; and performing inverse quantization on the basis of the image use information to reconstruct the image data, wherein the image use information can indicate human perception and/or machine perception of the image data.