Grid-Based Instance Segmentation for Faster Fine-Grained Masks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing instance segmentation methods are time-consuming, require a large amount of memory, and lack sufficient granularity in segmentation.

Innovation Solution

An image segmentation method that involves sending an input image to a network, obtaining a foreground target box, dividing the region of interest into grids, predicting features for each grid, and combining this with semantic features to achieve instance segmentation, reducing memory and time requirements.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional image segmentation models (FCN, U-Net, etc.) are used, then basic segmentation functionality is achieved, but they ignore contextual information and local features leading to poor segmentation performance in complex scenes

Engineering Contradiction:
Improvesegmentation performanceVSAvoidmodel structure complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The model divides the image processing task into multiple stages: global context extraction through pooling operations, local feature extraction through convolutional layers, and feature fusion. This segmentation of processing functions allows the model to capture both global and local information without requiring a monolithic complex architecture.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces global context information as an additional dimension by applying max pooling and average pooling operations that aggregate spatial information across the entire image. This global context is then concatenated with local features, adding a new dimensional aspect to the feature representation that enhances segmentation accuracy.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Measurement precision

If more comprehensive feature extraction methods are used to improve segmentation accuracy, then segmentation performance improves, but computational cost and model complexity increase

Engineering Contradiction:
Improvesegmentation accuracyVSAvoidcomputational cost
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The model performs preliminary global context extraction through pooling operations before the main segmentation process. By pre-computing and storing global max and average pooled features, the model avoids redundant computations during the segmentation process, reducing overall computational cost while maintaining accuracy.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent merges global context features (from pooling operations) with local features (from convolutional layers) through concatenation. This merging allows the model to leverage both types of features simultaneously, achieving high segmentation accuracy without requiring separate independent processing streams that would increase computational overhead.

Inventive Principle:
Principle #5Merging (Combining)

3Device complexity

If existing models are used without contextual information, then model simplicity is maintained, but obvious errors occur in segmented results

Engineering Contradiction:
Improvemodel simplicityVSAvoidsegmentation reliability
Core Design Contradiction:
Device complexityVSReliability

Solution Approach 1:

The patent introduces global context features as an intermediary between the input image and the segmentation output. These pooled features act as a mediator that provides contextual information to guide the segmentation process, improving reliability without requiring fundamental changes to the basic model architecture.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4214635B1Image segmentation method, electronic equipment, and storage medium
Publication Date: 2026.04.29 ZHEJIANG DAHUA TECH CO LTD
  • EP4214635B1 patent drawingFigure 1~2
  • EP4214635B1 patent drawingFigure 3~6
  • EP4214635B1 patent drawingFigure 7~9

AI summary

Disclosed are an image segmentation method, a training method for an image segmentation network, an electronic equipment, and a storage medium. The method includes: sending an input image to an image segmentation network; obtaining a first foreground target box of the input image; obtaining a first region of interest and a first region-of-interest feature map of the input image based on the first foreground target box; dividing the first region of interest into grids, predicting a corresponding feature of each grid in the first region of interest based on the first region-of-interest feature map, obtaining a semantic feature of each pixel in the first region of interest; and obtaining an instance segmentation result based on the corresponding feature of each grid in the first region of interest, information of the first foreground target box, and the semantic feature of each pixel in the first region of interest.