Transfer Learning CNN for X-ray Threat Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing image processing systems for x-ray security screening face challenges in accurately identifying specific items and threats due to limited availability of training data, requiring large datasets for deep learning convolutional neural networks (CNNs), which is unobtainable in many application domains.

Innovation Solution

The implementation of a deep learning CNN framework that employs transfer learning, retraining a pre-trained CNN with a smaller dataset of x-ray images to identify specific objects and threats, using a machine learning algorithm trained on a specific type of images, such as x-ray volumetric or projection images, to analyze input images and generate models for threat detection.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a deep multi-layer CNN approach is used for image classification, then detection accuracy is improved, but the requirement for large amounts of training data increases

Engineering Contradiction:
Improvedetection accuracyVSAvoidtraining data volume
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent applies transfer learning by pre-training a CNN on a large dataset (ImageNet) before fine-tuning it on the specific x-ray security imaging dataset. This preliminary training action allows the model to learn general feature representations that can be transferred to the target domain, thereby achieving high detection accuracy with limited domain-specific training data

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent modifies the pre-trained CNN model by changing its parameters (weights and biases) through fine-tuning on the specific security imaging dataset. This parameter adjustment allows the model to adapt from general image recognition to specific threat detection while requiring fewer training samples

Inventive Principle:
Principle #35Parameter changes

2Adaptability or versatility

If transfer learning is employed to overcome limited training data, then the system can be optimized for specific application domains, but the overall system complexity increases

Engineering Contradiction:
Improveapplication domain optimizationVSAvoidsystem complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent segments the training process into two distinct phases: pre-training on a general dataset and fine-tuning on the specific security imaging dataset. This segmentation allows the system to leverage pre-trained features while adapting to domain-specific characteristics, achieving adaptability without requiring complete retraining from scratch

Inventive Principle:
Principle #1Segmentation

3Measurement precision

If a pre-trained CNN is fine-tuned for specific application domains, then detection accuracy for specific items is improved, but the training time and computational resources increase

Engineering Contradiction:
Improvespecific item detection accuracyVSAvoidtraining time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

By performing pre-training on a large dataset before fine-tuning, the model already possesses robust feature extraction capabilities. This preliminary action reduces the fine-tuning time required compared to training a model from scratch, as the model only needs to adapt its parameters to the specific domain rather than learning features from zero

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent applies fine-tuning selectively to the later layers of the CNN architecture that are most critical for specific classification tasks, while keeping earlier feature extraction layers fixed or minimally adjusted. This localized adjustment approach reduces computational overhead and training time while maintaining high detection accuracy for specific items

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP3635617B1Systems and methods for image processing
Publication Date: 2024.05.08 LEIDOS SECURITY DETECTION & AUTOMATION INC
  • EP3635617B1 patent drawingFigure 1
  • EP3635617B1 patent drawingFigure 2
  • EP3635617B1 patent drawingFigure 3

AI summary

A computing-device implemented system and method for identifying an item in an x-ray image is described. The method includes training a machine learning algorithm with at least one training data set of x-ray images to generate at least one machine- learned model. The method further includes receiving at least one rendered x-ray image that includes an item, identifying the item using the at least one model, and generating an automated detection indication associated with the item.