Object-Granularity Image Segmentation for Precise Language Queries

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing language-driven precise instance segmentation methods face challenges in accurately locating instance targets among congested targets of the same category and predicting masks that are not overflowed to adjacent targets, leading to inaccurate target positioning and mask prediction.

Innovation Solution

Introduce a target integrity constraint in cross-modal feature fusion by using neural networks to fuse image features with language encoding at the object granularity, activating instance regions on a per-target basis, and updating the feature extraction network based on the similarity between text features and embedding vectors.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If cross-modal feature fusion is used to activate regions correlated with language encoding, then the model can process natural language inputs, but the instance target positioning accuracy deteriorates and masks are overflowed to adjacent targets

Engineering Contradiction:
Improvenatural language input processingVSAvoidinstance target positioning accuracy
Core Design Contradiction:
Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent segments the image feature map into multiple candidate regions corresponding to different objects, then processes each region independently through the neural network to generate separate embedding vectors. This segmentation allows the model to distinguish between multiple targets of the same category and assign accurate positions without mask overflow, while still maintaining the ability to process natural language inputs.

Inventive Principle:
Principle #1Segmentation

2Quantity of substance

If conventional semantic segmentation models are used, then the model can predict labels for targets, but the model cannot distinguish between different targets of the same category

Engineering Contradiction:
Improvetarget category coverageVSAvoidinstance discrimination accuracy
Core Design Contradiction:
Quantity of substanceVSMeasurement precision

Solution Approach 1:

The patent generates different embedding vectors for different candidate regions, allowing each region to have its own unique representation based on its local visual characteristics and relationship with the language input. This local quality differentiation enables the model to distinguish between different targets of the same category while maintaining comprehensive category coverage.

Inventive Principle:
Principle #3Local quality

3Productivity

If language-driven instance segmentation is applied to congested targets, then the model can process multiple targets, but the mask prediction accuracy deteriorates and adjacent targets are affected

Engineering Contradiction:
Improvemulti-target processing capabilityVSAvoidmask prediction accuracy
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent divides the image into multiple candidate regions and processes each region independently through the neural network, generating separate embedding vectors for each. This segmentation approach allows the model to handle multiple congested targets simultaneously while maintaining high mask prediction accuracy for each individual target, preventing mask overflow to adjacent targets.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4592866B1Data processing method and apparatus
Publication Date: 2026.03.11 HUAWEI TECH CO LTD
  • EP4592866B1 patent drawingFigure 1A~1C
  • EP4592866B1 patent drawingFigure 1D~2
  • EP4592866B1 patent drawingFigure 3

AI summary

A data processing method is applied to image processing, for example, image segmentation or target detection. The method includes: obtaining a first image feature corresponding to an image and a text feature corresponding to a text; obtaining a plurality of second embedding vectors through a neural network based on a plurality of preset first embedding vectors and the first image feature, where each second embedding vector corresponds to one candidate region of a target object, and each second embedding vector and the first image feature are used to be fused to obtain one corresponding second image feature; and determining, based on a similarity between the text feature and the plurality of second embedding vectors, a weight corresponding to each second embedding vector, where a plurality of weights are used to be fused with t a plurality of second image features, to determine a prediction region corresponding to the target object. In this application, the image feature is changed from being at a pixel granularity to being at a target object granularity, and pixels belonging to a same target are considered as a whole to be fused with language encoding, so that processing precision of a model can be improved.