Image Codec Profile Syntax for Machine-Oriented Compression
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
An image encoding/decoding method and apparatus that utilize a high-level syntax to define profiles based on neural network-based and machine analysis-based profiles, enabling adaptive encoding/decoding of specific coding tools and resources tailored to service purposes and device performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If existing image compression technology is used, then high-resolution and high-quality image processing for human vision is achieved, but encoding/decoding efficiency for machine learning services is insufficient
Solution Approach 1:
The patent introduces dynamic profile tier level (PTL) adjustment mechanisms that allow the encoding system to adaptively select compression profiles based on machine learning task requirements. The PTL can be dynamically modified through syntax elements in the bitstream, enabling the system to transition between different compression quality levels depending on the specific machine learning application needs, thereby improving both efficiency and adaptability.
Solution Approach 2:
The patent changes the parameter structure by introducing expanded profile tiers that include additional parameters for machine learning optimization. These parameter changes enable the system to represent image data in ways that are more suitable for neural network processing, such as modified color spaces, resolution levels, and compression ratios that optimize for machine vision tasks rather than human perception.
2Adaptability or versatility
If a unified compression profile is used, then simplicity is maintained, but adaptability to different machine learning applications is reduced
Solution Approach 1:
The patent segments the compression profile into distinct tiers (basic profile tier and expanded profile tier) with different levels of complexity. The basic tier provides simple compression for lightweight machine learning tasks, while the expanded tier offers enhanced compression capabilities for more demanding applications. This segmentation allows the system to select appropriate profile complexity based on specific application requirements, balancing adaptability with implementation simplicity.
Solution Approach 2:
The patent implements a nested profile structure where the expanded profile tier contains all the features of the basic profile tier plus additional machine learning optimization features. This nesting approach allows for incremental complexity addition, where systems can start with the basic profile and expand to more complex profiles only when needed, thereby managing overall system complexity while maintaining high adaptability.
3Productivity
If high-quality compression is applied, then image quality for human vision is improved, but encoding time and computational resources increase
Solution Approach 1:
The patent implements dynamic quality adjustment where the compression quality level is not fixed but can be adjusted based on the specific machine learning task requirements. The system can dynamically switch between faster, lower-quality compression and slower, higher-quality compression depending on the computational resources available and the precision requirements of the downstream machine learning model, thereby optimizing the trade-off between encoding speed and image quality.
Data Source
AI summary
The present application discloses image encoding/decoding method and apparatus based on high level syntax defining profile, and recording medium having bitstream stored therein. The image decoding method according to an embodiment of the present disclosure comprises obtaining profile tier level (PTL) information of an image and determining a profile and a tier of the image based on the PTL information, wherein the PTL information includes a profile indicator indicating a type of the profile for the image, wherein the profile type includes a basic profile or an expanded profile, and wherein the expanded profile includes a neural network (NN)-based profile and a machine analysis-based profile.