Region-Based Damage Transfer for Small Visual Inspection Datasets

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing automated visual inspection systems face challenges in generalizing damage detection due to limited datasets and lack of tools for generating diverse images focused on damage regions, leading to poor performance in varying conditions and contexts.

Innovation Solution

A region-based adversarial learning method that uses a generative adversarial network to learn and transfer fine-grained damage representations across multiple contexts and conditions, enhancing model generalization by focusing on damage regions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If automated visual inspection systems use limited customer datasets with damages within similar conditions and context, then the dataset collection is easier and faster, but the model generalization power deteriorates leading to poor detection in field with varying ambient and image conditions

Engineering Contradiction:
Improvedataset collection efficiencyVSAvoidmodel generalization power
Core Design Contradiction:
ProductivityVSAdaptability or versatility

Solution Approach 1:

The system performs preliminary actions by automatically collecting and augmenting training data before model deployment. The automated data collection system proactively gathers images in diverse conditions and applies augmentations beforehand, so when the model encounters varying field conditions, it has already been exposed to similar variations during training, improving generalization without manual intervention.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system creates copies of existing damage images through automated augmentation techniques, generating synthetic training samples that replicate damage characteristics across different conditions, viewpoints, and backgrounds. This allows the model to learn from expanded dataset variations without requiring manual collection of diverse real-world samples.

Inventive Principle:
Principle #26Copying

2Adaptability or versatility

If manual collection of images in diverse conditions and characteristics is conducted, then the model generalization power is improved, but the process becomes time and labor intensive, expensive, and in several cases almost infeasible

Engineering Contradiction:
Improvemodel generalization powerVSAvoiddata collection time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The system performs self-service by automatically collecting, augmenting, and preparing training data without human intervention. The automated pipeline independently handles image collection across diverse conditions, applies transformations, and prepares augmented datasets, eliminating the need for manual data collection while maintaining model generalization capabilities.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system changes parameters of existing images through automated augmentation techniques, systematically varying conditions such as lighting, viewpoint, background, and damage characteristics. This generates diverse training samples by modifying image parameters rather than manually collecting new images, reducing time and resource requirements while improving model adaptability.

Inventive Principle:
Principle #35Parameter changes

3Productivity

If automated tools for image augmentation are used with basic tasks such as crop, rotation or change colors, then the processing speed is improved, but the learning of fine-grain characteristics of damages deteriorates

Engineering Contradiction:
Improveimage processing speedVSAvoidfine-grain damage characteristics learning
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The system segments the image processing task by separating basic geometric transformations from fine-grain damage characteristic learning. While automated tools handle basic operations like cropping and rotation for efficiency, the system additionally implements specialized augmentation techniques that preserve and enhance fine-grain damage characteristics, ensuring both processing speed and learning precision are maintained.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4181078B1Damage transfer method with a region-based adversarial learning
Publication Date: 2026.03.18 HITACHI LTD
  • EP4181078B1 patent drawingFigure 1
  • EP4181078B1 patent drawingFigure 2
  • EP4181078B1 patent drawingFigure 3

AI summary

Example implementations involve systems and methods to create robust visual inspection datasets and models. The novel method learns and transfers damage representation from few samples to new images. The proposed method introduces a generative region-of-interest based adversarial network with the aim of learning a common damage representation and transferring it to an unseen image. The proposed approach shows the benefit of adding damage-region-based component, since existing methods fail to transfer the damages. The proposed method successfully generated images with variations in context and conditions to improve model generalization for small datasets.