MFA-net: Object detection for complex X-ray cargo and baggage security imagery.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Security Systems

Background:

• Current deep convolutional networks for X-ray security focus solely on baggage scans.
• This limits detection accuracy for other domains like cargo screening.
• A generalized approach is needed for comprehensive threat detection.

Purpose of the Study:

• To propose a novel object detection method, MFA-net, for efficient contraband detection in both baggage and cargo X-ray security scans.
• To enhance the accuracy and robustness of threat detection systems across diverse security domains.
• To overcome limitations of existing frameworks in handling object scale variations and occlusions.

Main Methods:

• Developed MFA-net with three key modules: multiscale dilated convolutional module, fusion feature pyramid network, and auxiliary point detection head.
• The multiscale dilated convolutional module addresses object-scale variant issues using dynamic feature selection.
• The fusion feature pyramid network enhances multiscale object recognition and alleviates occlusion problems.
• An auxiliary point detection head improves localizability without additional ground-truth data.

Main Results:

• MFA-net was evaluated on two large-scale datasets: SIXray (baggage) and CargoX (cargo).
• The proposed method demonstrated superior performance compared to state-of-the-art object detectors in both baggage and cargo domains.
• The plug-and-play modules significantly improved detection capabilities on X-ray security images.

Conclusions:

• MFA-net provides an effective solution for contraband detection in diverse X-ray security screening applications.
• The proposed modules can be integrated into existing object detectors to boost their performance on X-ray security images.
• This research advances automated inspection systems in transport security by enabling more accurate and generalized threat detection.