Boundary-Weighted Semantic Segmentation for Accurate Object Edges

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional semantic segmentation methods using computer vision deep learning fail to accurately predict boundary pixels, leading to high scores despite inaccurate boundary line predictions due to the use of metrics like dice coefficient loss and IOU, which measure overlapping areas without considering boundary precision.

Innovation Solution

A method that identifies boundary regions in training images, obtains additional training images with enhanced weights for these regions, and trains neural networks using these images to improve boundary segmentation accuracy, employing techniques such as cross-entropy loss and dice coefficient loss, along with Intersection over Union (IOU) for refined training.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If conventional measurement methods (dice coefficient loss, IOU) are used to evaluate semantic segmentation, then overall segmentation coverage is improved, but boundary region accuracy deteriorates

Engineering Contradiction:
Improvesegmentation coverageVSAvoidboundary accuracy
Core Design Contradiction:
Quantity of substanceVSMeasurement precision

Solution Approach 1:

The patent divides the image into two distinct regions: boundary regions and non-boundary regions. By segmenting the evaluation process into region-specific metrics (boundary IOU for boundaries, dice coefficient for non-boundaries), the system can independently optimize for boundary accuracy without compromising overall coverage. This resolves the contradiction by allowing different evaluation priorities for different spatial zones.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies different loss functions and evaluation metrics to different regions of the image. Boundary regions use boundary IOU with higher weighting, while non-boundary regions use dice coefficient loss. This local differentiation allows the system to maintain high boundary accuracy while preserving overall segmentation coverage, directly addressing the technical contradiction.

Inventive Principle:
Principle #3Local quality

2Measurement precision

If boundary region weighting is increased during training, then boundary segmentation accuracy is improved, but training complexity increases

Engineering Contradiction:
Improveboundary accuracyVSAvoidtraining complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent performs preliminary identification of boundary regions using contour detection algorithms before the main segmentation training. By pre-marking boundary pixels and preparing region-specific weights in advance, the system simplifies the training process rather than complicating it, as the boundary regions are already identified and weighted appropriately during training.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent modifies the loss function parameters by introducing region-specific weighting factors (alpha for boundary, beta for non-boundary). These parameter changes allow the model to focus on boundary accuracy through adjustable weights without fundamentally changing the training architecture, thus improving boundary accuracy while maintaining manageable training complexity.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4648007A1Method for improving semantic segmentation accuracy at boundary regions
Publication Date: 2025.11.12 AIVEX CO LTD
  • EP4648007A1 patent drawingFigure 1~2
  • EP4648007A1 patent drawingFigure 3
  • EP4648007A1 patent drawingFigure 4

AI summary

Disclosed is a method for performing object segmentation using a neural network model, the method performed by one or more processors of a computing device according to an exemplary embodiment of the present disclosure. the method may include: obtaining a training image containing an object; training a neural network model to perform object segmentation using the training image; identifying a boundary region of the object included in the training image; obtaining an additional training image based on the training image and the identified boundary region; and additionally training the trained neural network model based on the additional training image.