PDSXray: A Benchmark Dataset for Pseudocoloring-Driven Domain Adaptation in Security X-ray Inspection

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Security X ray imaging datasets primarily address hardware-induced domain shifts.
• Pseudocoloring driven domain shift (PDS) from varying pseudocoloring schemes is understudied.

Purpose of the Study:

• To introduce a benchmark dataset for security X ray inspection domain shift analysis.
• To specifically address PDS and endogenous domain shifts.
• To facilitate research on model generalization and robustness.

Main Methods:

• Constructed a novel benchmark dataset with two subsets: synthetic (isolating PDS) and aggregated (integrating public datasets).
• Applied controlled hue transformations for synthetic PDS evaluation.
• Standardized annotations for reproducibility and comparability.

Main Results:

• The dataset enables independent evaluation of PDS.
• It supports verification of cross-domain and cross-device model generalization.
• Provides a benchmark for object detection algorithms under domain shifts.

Conclusions:

• The dataset addresses a critical gap in security X ray imaging research.
• It offers a valuable resource for developing more robust object detection models.
• Facilitates advancements in diverse X ray security inspection scenarios.