RWSC-Fusion: Region-Wise Style-Controlled Fusion Network for the Prohibited X-ray Security Image Synthesis

Luwen Duan, Min Wu, Lijian Mao, Jun Yin, Jianping Xiong, Xi LiCVPR 2023 · DOI

Abstract

Automated security systems for prohibited item detection face a chronic shortage of annotated X-ray datasets. This paper introduces RWSC-Fusion, a novel deep-learning framework that synthesizes realistic X-ray security images by fusing prohibited items into ordinary baggage scans. By leveraging physics-inspired loss functions and region-specific style control, RWSC-Fusion generates high-quality training data that respects the unique translucent and multiplicative properties of X-ray imagery, significantly boosting the performance of downstream object detection models.

Introduction

Automated prohibited item detection is critical for transport security, yet deep learning models require massive amounts of annotated data. In X-ray security, data is difficult to acquire because prohibited items appear infrequently in real-world scans. Furthermore, X-ray images are formed by high-energy beams penetrating objects, resulting in a "translucent" appearance where overlapping objects combine multiplicatively. Traditional data augmentation techniques fail to capture these complex visual interactions.

Overview of the RWSC-Fusion synthesis process, showing how prohibited items are realistically fused into original baggage scans.

Methodology

The authors propose the Region-Wise Style-Controlled Fusion (RWSC-Fusion) network, which consists of three primary innovations designed to ensure physical and visual realism.

RWSC-Fusion Module

Unlike standard fusion models that apply a global style, this module focuses specifically on the overlapping region between the prohibited item and the baggage. It learns unique scaling and shifting parameters from feature maps, allowing for fine-grained control over how the foreground and background blend without causing global color shifts or "ghosting."

Edge-Attention (EA) Module

X-ray images often have sparse textures. To prevent the loss of fine details like zippers or shoe textures during synthesis, the EA module employs a sharpening operation—subtracting local mean features—and channel-wise attention to enhance edges.

Architecture of the EA-Dense Block and the Edge-Attention (EA) module designed to enhance fine textures in synthesized X-ray images.

Physics-Inspired Loss Functions

To align the network with the physical principles of X-ray imaging, two specific loss functions are used:

  • Luminance Loss in Logarithmic form (LL): Based on the Beer-Lambert law, this loss ensures the network learns the multiplicative nature of X-ray attenuation.
  • Correlation loss of Saturation Difference (CSD): This maintains realistic saturation levels for pseudo-colored security images, which use color to represent material types (e.g., orange for organic materials).

Key Results and Discussion

The researchers evaluated RWSC-Fusion using a large private dataset and public benchmarks like SIXray and OPIXray. The model outperformed state-of-the-art methods such as U2Fusion and FusionGAN across metrics like PSNR and Edge Intensity (EI).

Crucially, the synthetic data provided a significant boost to automated detection. On a private dataset, adding RWSC-Fusion images to a limited real-world training set improved the Mean Average Precision (mAP) of a YOLOv4 detector by over 3%. On the SIXray benchmark, the model helped recover performance in data-scarce scenarios, raising mAP from 49.22% to 70.89%. Subjective tests also confirmed that the synthetic images were realistic enough to deceive professional security inspectors.

Conclusion

RWSC-Fusion addresses the critical bottleneck of data scarcity in X-ray security. By formalizing X-ray fusion as a multiplicative, region-specific process, the framework generates high-fidelity training data that improves the reliability of automated detection systems. Future work will explore applying these region-wise style controls to other domains, such as medical imaging and multi-exposure photography.

Summary updated

BibTeX
@inproceedings{duan2023rwsc,
  author = {Luwen Duan and Min Wu and Lijian Mao and Jun Yin and Jianping Xiong and Xi Li},
  title = {{RWSC-Fusion: Region-Wise Style-Controlled Fusion Network for the Prohibited X-ray Security Image Synthesis}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2023},
  doi = {10.1109/cvpr52729.2023.02145},
}

Loading related content...

Figure