Incremental learning for detection in X-ray luggage perspective images.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Convolutional neural networks (CNNs) excel at X-ray contraband detection but struggle with new classes and data imbalance.
• Offline training hinders real-time adaptation and leads to catastrophic forgetting of previously learned classes.

Purpose of the Study:

• To develop an incremental learning method for online continuous learning of contraband detection models.
• To enable real-time detection of new contraband classes without retraining and mitigate catastrophic forgetting.

Main Methods:

• Parameter compression using distillation to maintain old class identification.
• Incremental learning of area proposal and object detection subnetworks for new class recognition.
• A novel loss function designed to prevent catastrophic forgetting and ensure stable detection.

Main Results:

• The proposed model stably learns new contraband classes, even with uneven data distribution.
• Parameter compression ensures stable identification of previously learned classes.
• The new loss function effectively prevents catastrophic forgetting.

Conclusions:

• The incremental learning method offers a robust solution for real-time contraband detection in dynamic environments.
• The approach enhances the adaptability and performance of security screening systems.
• This research addresses critical limitations in current deep learning models for contraband detection.