X-ray Image Synthesis for Overlapping Article Recognition

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current X-ray baggage inspection systems face challenges in accurately detecting dangerous articles, especially when items overlap, due to limitations in existing image recognition technologies, which require extensive training and are costly to implement, and are not adaptable to varying combinations of articles.

Innovation Solution

An X-ray image processing system that uses deep learning and image synthesis techniques to generate learning data, allowing for high-accuracy recognition of overlapping articles by constructing a co-occurrence data table and article data table, and employing a processing apparatus with X-ray sensors to differentiate between high and low energy data, enabling the synthesis of images and generation of learning models.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If deep learning is used for image recognition, then recognition accuracy can be improved, but a large amount of learning data is necessary which increases preparation time and cost

Engineering Contradiction:
Improverecognition accuracyVSAvoidlearning data preparation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system performs preliminary actions by automatically generating synthetic learning data through virtual X-ray image synthesis before actual deep learning training begins. The learning data generation unit creates diverse training samples by combining article images and X-ray transmission data in advance, eliminating the need for time-consuming manual data collection and preparation while enabling high-accuracy recognition model training

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system creates copies by synthesizing virtual X-ray images that replicate real inspection scenarios. The image synthesis unit generates synthetic learning data that copies the characteristics and patterns of actual X-ray images, allowing the deep learning model to train on numerous synthetic examples without requiring equivalent real-world data collection time

Inventive Principle:
Principle #26Copying

2Extent of automation

If existing image recognition technology is used, then automation can be achieved, but it cannot accurately recognize overlapping articles in X-ray images

Engineering Contradiction:
Improvedetection automationVSAvoidoverlapping article recognition accuracy
Core Design Contradiction:
Extent of automationVSMeasurement precision

Solution Approach 1:

The system introduces an intermediary approach by using X-ray transmission amount data as a mediator between the image synthesis unit and the deep learning unit. This intermediary data layer enables the synthetic images to accurately represent overlapping article scenarios, allowing the automated recognition system to distinguish overlapping articles with high precision that conventional automated systems cannot achieve

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The system applies parameter changes by utilizing dual-energy X-ray transmission data with different energy levels to synthesize images. By varying the energy parameters and combining transmission data from multiple energy levels, the system generates synthetic images that reveal different material properties and overlapping structures, enabling accurate automated recognition of overlapping articles

Inventive Principle:
Principle #35Parameter changes

3Reliability

If manual inspection by trained inspectors is performed, then dangerous articles can be detected, but it requires highly-trained inspectors which increases cost and reduces flexibility

Engineering Contradiction:
Improvedetection reliabilityVSAvoidinspector training requirements
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The system implements self-service by enabling the inspection system to automatically generate its own learning data and train its recognition models without human intervention. The learning data generation unit autonomously creates synthetic training samples, and the deep learning unit automatically improves recognition accuracy through iterative training, eliminating the need for human inspectors while maintaining high detection reliability

Inventive Principle:
Principle #25Self-service

Applied Scientific Principles

This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.

Function Achieved in This Case

Improves the accuracy of image recognition in X-ray inspections by automatically generating learning data that accounts for overlapping articles, reducing the burden on inspectors and enhancing the efficiency of baggage screening processes.

Implementation Method 1

an X-ray measurement value indicating a transmission amount of the X-ray with respect to the article

Methodology Applied
Scientific EffectX-ray transmission: X-Ray

Data Source

PatentEP3772722B1X-ray image processing system and method, and program therefor
Publication Date: 2024.05.08 HITACHI SOFTWARE ENG
  • EP3772722B1 patent drawingFigure 1
  • EP3772722B1 patent drawingFigure 2
  • EP3772722B1 patent drawingFigure 3

AI summary

Information on an area specified as having an article is acquired, synthesis is performed with X-ray transmission amounts acquired by a plurality of sensors in the same area (for example, the same background and the same luggage), and material information is estimated from the X-ray transmission amount synthesized again to generate a color image. As a synthesis target, for example, combination information of articles accumulated during operation is used, and articles having many combinations are used as combination of articles included in the same luggage. A color image generated by synthesis is used as learning data.