Limits on transfer learning from photographic image data to X-ray threat detection

Area of Science:

• Computer Vision
• Machine Learning
• Security Technology

Background:

• X-ray imaging is vital for transport security but faces interpretation bottlenecks.
• Deep Learning (DL) offers automated threat detection but requires extensive labeled data.
• Acquiring large labeled X-ray datasets is challenging compared to photographic data.

Purpose of the Study:

• To assess the feasibility of using photographic data to train X-ray threat detectors.
• To evaluate the extent to which transfer learning can improve X-ray security screening.

Main Methods:

• A dataset of 1901 matched photo-X-ray image pairs was created, including 258 pairs of threat objects.
• Various transfer learning techniques were tested using this dataset.
• A model of threat cue availability was developed to analyze transferability limits.

Main Results:

• Learned appearance features from photos provide a foundation for classifier training.
• Approximately 40% of danger cues are transferable from photos to X-rays, but 60% are not.
• Transfer learning is effective only when X-ray data is extremely scarce (tens of images).

Conclusions:

• Transfer learning from photographic data offers benefits for X-ray threat detection primarily in low-data regimes.
• The utility of transfer learning diminishes significantly as the volume of available X-ray training data increases.
• This study underscores the importance of data quantity in the effectiveness of transfer learning for specialized imaging domains.