AI Damage Detector Training Samples from Labeled Vehicle Images

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing AI-based vehicle damage assessment systems face challenges in accurately detecting damages due to interference from additional information in input images, requiring sufficient positive and negative samples, and are prone to low accuracy, especially when not trained with similar damages.

Innovation Solution

A system that generates positive and negative image samples from labeled images using target bounding boxes and internal bounding boxes to train a target detector, allowing it to identify damages independently of the underlying vehicle parts by focusing on material-based damage detection.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If AI models are used for automated damage detection, then processing time is reduced and labor costs are decreased, but detection accuracy deteriorates due to interference from additional information in input images

Engineering Contradiction:
Improveprocessing timeVSAvoiddetection accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent extracts and removes additional information (background elements, other objects) from the input image to create a cleaned image that contains only the damaged vehicle. This extraction process eliminates interference from non-damage-related elements, allowing the AI model to focus solely on detecting damage characteristics, thereby maintaining high detection accuracy while preserving the automated processing speed benefit.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent segments the input image into relevant regions (damaged areas) and irrelevant regions (background, other objects). By dividing the image and selectively processing only the relevant segments for damage detection, the system reduces computational interference while maintaining processing efficiency. The segmentation allows the AI model to concentrate computational resources on damage-affected regions only.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If manual survey and damage assessment are performed by claim adjusters, then detection accuracy is maintained through expert evaluation, but processing time increases significantly and labor costs increase

Engineering Contradiction:
Improvedetection accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent creates a synthesized copy of the damage assessment process using AI models trained on labeled images. Instead of requiring physical presence of claim adjusters, the system uses digital copies of damage data from labeled images to train the AI model. This copying approach enables automated replication of expert assessment logic, maintaining detection accuracy while eliminating the time and cost constraints of manual surveying.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The AI model performs self-service damage assessment by automatically analyzing input images without requiring human intervention. The system uses its own trained capabilities to detect, locate, and assess damage, replacing the need for claim adjusters to physically inspect vehicles. This self-service mechanism maintains assessment accuracy through continuous learning from labeled data while dramatically reducing processing time and labor requirements.

Inventive Principle:
Principle #25Self-service

3Measurement precision

If sufficient positive and negative samples are generated for training, then detection accuracy is improved, but the complexity of sample generation and data processing increases

Engineering Contradiction:
Improvedetection accuracyVSAvoidsample generation complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent introduces labeled images as an intermediary medium between raw input images and training data generation. These labeled images serve as a mediator that pre-identifies damage regions and provides ground truth annotations. By using labeled images as an intermediary, the system simplifies the sample generation process - instead of requiring complex algorithms to create training samples from scratch, the system leverages the pre-processed labeled data to efficiently generate both positive and negative samples for training.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3847579B1Method and computer-readable storage medium for generating training samples for training a target detector
Publication Date: 2025.09.03 ADVANCED NEW TECHNOLOGIES CO LTD
  • EP3847579B1 patent drawingFigure 1A
  • EP3847579B1 patent drawingFigure 1B
  • EP3847579B1 patent drawingFigure 2

AI summary

Embodiments described herein provide a system for facilitating image sampling for training a target detector. During operation, the system obtains a first image depicting a first target. Here, the continuous part of the first target in the first image is labeled and enclosed in a target bounding box. The system then generates a set of positive image samples from an area of the first image enclosed by the target bounding box. A respective positive image sample includes at least a part of the first target. The system can train the target detector with the set of positive image samples to detect a second target from a second image. The target detector can be an artificial intelligence (AI) model capable of detecting an object.