Multi-Loss Image Segmentation Training for Continuous Linear Objects

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods fail to accurately and efficiently evaluate the continuity and indefinite shapes of linear-shaped objects in image recognition, particularly in semantic segmentation.

Innovation Solution

A learning apparatus that employs multiple loss functions to evaluate and optimize a model for image segmentation, including a first loss function for identifying features, a second loss function for evaluating continuity, and optionally a third loss function for emphasizing contours, through backpropagation, to generate a model that accurately recognizes linear-shaped objects.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If semantic segmentation using deep learning is used to improve recognition resolution for linear objects, then the resolution of recognition is improved, but the ability to accurately evaluate continuity and indefinite shapes deteriorates

Engineering Contradiction:
Improverecognition resolutionVSAvoidevaluation accuracy of continuity and shapes
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The loss function is segmented into multiple independent components: a first loss function for basic segmentation accuracy, a second loss function for continuity evaluation, and a third loss function for contour emphasis. Each component addresses a specific aspect of linear object recognition, allowing the model to optimize each feature separately while maintaining overall performance.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent changes the parameters of the loss function by introducing multiple evaluation metrics with different mathematical formulations. The second loss function uses element product calculations to evaluate continuity, while the third loss function applies contour weighting to emphasize boundary features. These parameter changes enable the model to simultaneously optimize for resolution, continuity, and shape accuracy.

Inventive Principle:
Principle #35Parameter changes

2Device complexity

If a single loss function is used for model training, then the training process is simple, but the model cannot accurately evaluate multiple features simultaneously

Engineering Contradiction:
Improvetraining process complexityVSAvoidfeature evaluation accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

Multiple loss functions are merged into a unified training framework where the total loss is the sum of individual loss components. This allows the model to evaluate multiple features (segmentation accuracy, continuity, contour definition) simultaneously while maintaining a coherent training process. The gradient from each loss function is backpropagated together to update the model parameters.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The loss function system is designed to be multi-functional, where a single training mechanism serves multiple purposes: basic segmentation, continuity evaluation, and contour emphasis. This universal approach allows one training process to optimize multiple aspects of linear object recognition without requiring separate training procedures for each feature.

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Productivity

If conventional loss functions are used for linear object recognition, then the training is efficient, but the model fails to capture continuity features of linear shapes

Engineering Contradiction:
Improvetraining efficiencyVSAvoidcontinuity recognition accuracy
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The second loss function acts as an intermediary between the basic segmentation loss and the final output, specifically evaluating the continuity feature. It calculates element products of segmentation masks at different positions to determine how well the model preserves the continuous nature of linear objects. This intermediary evaluation guides the model to maintain continuity without disrupting the overall training efficiency.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4583044A1Training device, image segmentation device, training method, and program
Publication Date: 2025.07.09 OMRON CORP
  • EP4583044A1 patent drawingFigure 1~2
  • EP4583044A1 patent drawingFigure 3
  • EP4583044A1 patent drawingFigure 4

AI summary

A learning apparatus configured to train a model in an image segmentation apparatus configured to perform segmentation of a recognition target in an image using the model, the learning apparatus including: a first evaluation value acquisition unit configured to acquire a first evaluation value calculated by a first loss function configured to evaluate a feature configured to identify the recognition target in the image; a second evaluation value acquisition unit configured to acquire a second evaluation value calculated by a second loss function configured to evaluate a continuity of the recognition target identified in the image; and a learning execution unit configured to perform learning to optimize the model by performing backpropagation based on the first evaluation value and the second evaluation value.