Feature Map Encoding with ROI-Based Compression for AI Tasks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image compression technologies are not optimized for machine tasks and lack efficiency in processing large amounts of image data required for artificial intelligence services.

Innovation Solution

A feature encoding/decoding method and apparatus that distinguishes regions of interest (ROI) and non-ROI in feature maps, using representative values and difference values to enhance encoding/decoding efficiency, generating a compressed bitstream for machine tasks.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If existing image compression technologies are used for machine tasks, then high-resolution and high-quality image processing is achieved, but encoding/decoding efficiency is insufficient for processing large amounts of image data required for artificial intelligence services

Engineering Contradiction:
Improveimage qualityVSAvoidencoding/decoding efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The feature map is divided into multiple channels, and each channel is further divided into multiple regions. This segmentation allows selective encoding of different regions based on their importance to machine tasks, thereby improving encoding efficiency while maintaining necessary quality for AI processing.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Different encoding strategies are applied to different regions within the same feature map. Regions of interest (ROIs) that are important for machine tasks are encoded with higher quality, while non-ROI regions use lower quality encoding, optimizing the balance between overall quality and encoding efficiency.

Inventive Principle:
Principle #3Local quality

2Loss of information

If all regions of a feature map are encoded with equal quality, then comprehensive image information is preserved, but encoding complexity and processing time increase

Engineering Contradiction:
Improvefeature information preservationVSAvoidencoding complexity
Core Design Contradiction:
Loss of informationVSDevice complexity

Solution Approach 1:

Instead of encoding all regions with equal high quality, the method applies partial encoding action only to regions that are important for machine tasks. This reduces the overall encoding complexity while preserving the critical feature information needed for AI processing.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The encoding quality parameter is changed dynamically based on region importance. ROIs are assigned higher quality parameters while non-ROIs use lower quality parameters, reducing overall encoding complexity without losing critical information for machine tasks.

Inventive Principle:
Principle #35Parameter changes

3Productivity

If region-wise selective encoding is applied to improve efficiency, then encoding/decoding speed increases, but complexity of determining ROI and non-ROI regions increases

Engineering Contradiction:
Improveencoding speedVSAvoidregion classification complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The regions are classified into ROIs and non-ROIs before the encoding process begins. This preliminary classification based on importance metrics allows the subsequent encoding to proceed efficiently without complex real-time decisions, improving encoding speed while managing classification complexity in advance.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4686200A1Feature encoding/decoding method and device, recording medium storing bitstream, and method for transmitting bitstream
Publication Date: 2026.01.28 LG ELECTRONICS INC
  • EP4686200A1 patent drawingFigure 1
  • EP4686200A1 patent drawingFigure 2
  • EP4686200A1 patent drawingFigure 3

AI summary

The present disclosure relates to a feature encoding/decoding method, a recording medium storing a bitstream, and a method for transmitting a bitstream. A feature decoding method performed by a feature decoding device, according to one embodiment of the present disclosure, comprises the steps of: obtaining information on a channel within a feature from a bitstream; and reconstructing the channel on the basis of the information on the channel, wherein the information on the channel may include information on whether coding has been performed according to an importance level of each region of the channel.