Towards More Efficient Security Inspection via Deep Learning: A Task-Driven X-ray Image Cropping Scheme

Area of Science:

• Computer Vision
• Artificial Intelligence
• Security Technology

Background:

• X-ray imaging is crucial for security screening in public spaces.
• Deep learning enhances automated object detection in X-ray scans, reducing labor costs.
• Detecting small objects in varied X-ray images is challenging due to size, aspect ratio, and background noise.

Purpose of the Study:

• To develop an efficient and effective deep learning-based object detection method for X-ray luggage inspection.
• To address the challenges of varied image scales, aspect ratios, and object locations in X-ray imagery.
• To introduce a novel dataset for benchmarking X-ray detection algorithms.

Main Methods:

• Proposed a two-stage Task-Driven Cropping (TDC) scheme for adaptive X-ray image cropping.
• Utilized a task-specific deep feature extractor to identify and preserve task-relevant regions.
• Developed the SIXray-D dataset with accurate annotations for supervised X-ray detection model training.

Main Results:

• The TDC scheme effectively improved the performance of popular deep learning detection algorithms.
• Achieved better mean Average Precision (mAP) scores for object detection in X-ray images.
• Demonstrated a reduction in processing time for X-ray inspection tasks.

Conclusions:

• The proposed TDC scheme offers a significant improvement for deep learning-based object detection in X-ray security applications.
• The TDC method enhances both the accuracy and efficiency of luggage inspection systems.
• The SIXray-D dataset provides a valuable resource for advancing research in X-ray image detection.