Image Segmentation Model Training for Scale-Invariant Precision

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image segmentation methods rely on shallow features like pixel values and grayscale differences, leading to low precision and accuracy, especially when targets have subtle edges or differences, and existing methods like SegAN improve segmentation accuracy but still face challenges with scale variations.

Innovation Solution

A method for training an image segmentation model using an encoder and decoder, where feature extraction is performed on sample and rescaled images, calculating class activation graphs, and training the decoder with a weighted sum of class activation graph losses and rescaled losses to ensure segmentation results remain consistent across different scales.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If traditional threshold method is used for image segmentation, then the method is simple to implement, but the segmentation precision is low

Engineering Contradiction:
ImproveEase of implementationVSAvoidSegmentation precision
Core Design Contradiction:
Ease of manufactureVSMeasurement precision

Solution Approach 1:

The patent transforms the image to multiple scales and processes them through different network depths, changing the scale parameter to improve segmentation precision while maintaining computational feasibility through parameter optimization

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If deep feature extraction is used to improve segmentation accuracy, then the segmentation precision is improved, but the computational complexity increases

Engineering Contradiction:
ImproveSegmentation precisionVSAvoidComputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent divides the image into multiple scales and processes each scale through the network separately, then combines the results. This segmentation approach allows using deep features for precision while managing complexity by processing different scales independently rather than requiring extremely deep networks on the full resolution image

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces the scale dimension by transforming the image to multiple resolutions. This adds a dimensional approach to feature extraction, allowing the model to capture features at different levels without requiring excessive network depth at a single scale

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

3Stability of the object's composition

If scale transformation is applied to ensure scale invariance, then the segmentation consistency across scales is improved, but the processing time increases

Engineering Contradiction:
ImproveSegmentation consistencyVSAvoidProcessing time
Core Design Contradiction:
Stability of the object's compositionVSLoss of time

Solution Approach 1:

The patent performs scale transformation and feature extraction at multiple scales during the training phase in advance. The model learns scale-invariant features beforehand, so during inference, the pre-learned knowledge can be applied without requiring extensive real-time processing at multiple scales

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4220555B1Training method and apparatus for image segmentation model, image segmentation method and apparatus, and device
Publication Date: 2025.12.10 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP4220555B1 patent drawingFigure 1
  • EP4220555B1 patent drawingFigure 2
  • EP4220555B1 patent drawingFigure 3

AI summary

A method for training an image segmentation model, including: performing feature extraction on a sample image and a rescaled image to obtain a sample image feature of sample image and a rescaled image feature of the rescaled image; performing class activation graph calculation to obtain a sample class activation graph of the sample image and a rescaled class activation graph of the rescaled image; calling the decoder to decode the sample image feature to obtain a sample segmentation result of the sample image, and calling the decoder to decode the rescaled image feature to obtain a rescaled segmentation result of the rescaled image; calculating a class activation graph loss and a rescaled loss based on the sample class activation graph, the sample segmentation result, the rescaled class activation graph, and the rescaled segmentation result; and training the decoder based on the class activation graph loss and the rescaled loss.