Quad-DIP for X-Ray Cargo Image Decomposition

Zheng Hu, Qiang Li, Gang Fu, Yuxiang Xing, Li ZhangWorkshops @ CVPR 2021 · DOI

Introduction

Security and logistics operations rely heavily on X-ray scans to identify illicit goods. However, the visual complexity of vehicle structures—such as engines, chassis, and wheels—often obscures the cargo. A significant hurdle in automating this process is the impossibility of supervised learning; obtaining "clean" scans of every vehicle model without cargo for use as labels is logistically unfeasible. Furthermore, even vehicles of the same model exhibit slight variations due to maintenance or positioning during the scan.

The Quad-DIP framework decomposes an input X-ray image into vehicle structure and cargo information.

To address this, the Quad-DIP framework provides an unsupervised deep learning approach to "unmix" these layers. By leveraging structural similarities between vehicles of the same class, the model learns to isolate goods without ever seeing a ground truth image of an empty truck.

Methodology

Quad-DIP evolves the "Deep Image Prior" (DIP) concept into a dual-stage, double-network system. The architecture employs two distinct sub-networks: CarNet, designed to capture the vehicle's structural components, and GoodsNet, focused on isolating the cargo.

Cross-Generation Strategy

The core innovation is a cross-generation strategy that processes pairs of images ($X$ and $Y$) representing the same vehicle type but containing different cargo.

The dual-stage training process utilizing cross-generation to isolate vehicle and cargo components.
  1. Stage 1: The model decomposes both $X$ and $Y$ into their respective vehicle ($C_x, C_y$) and cargo ($G_x, G_y$) components.
  2. Cross-Mixing: The cargo components are swapped to generate "fake" mixed images (e.g., combining the vehicle structure of $X$ with the cargo of $Y$ to create image $U$).
  3. Stage 2: These synthetic images are decomposed again. This creates a self-supervising loop that enforces cycle consistency and significantly reduces "cross-talk," where vehicle parts might otherwise bleed into the cargo layer.

Physics-Informed Pre-processing

X-ray attenuation follows an exponential decay law ($I = I_0 e^{-\mu l}$). By applying a negative logarithm to the raw data, the authors transform the image into an additive space where the relationship $Image = Vehicle + Cargo$ holds true, providing a sound mathematical basis for linear decomposition.

Experimental Results

The framework was evaluated on a real-world dataset of 3,897 top-down X-ray images, including 2,769 cargo tricycles and 1,128 large container trucks. Images were aligned using SIFT and "Demons" registration to ensure structural overlap before processing.

Key findings include:

  • The Common Denominator Principle: By training on image pairs, the network identifies repeating structural patterns as part of the vehicle (CarNet), while transient patterns are assigned to the cargo (GoodsNet).
  • Superior Separation: Quad-DIP successfully isolated intricate vehicle parts, such as dense engine blocks, from overlapping cargo, even when the cargo was placed directly over the engine.
  • Performance: The method outperformed standard unsupervised decomposition techniques (like DDIP), yielding sharper details in the cargo layer essential for downstream automated recognition tasks.

Conclusion and Future Work

Quad-DIP demonstrates that high-quality image decomposition is possible without manual labeling. While the model currently faces challenges in "zero-information" zones where X-rays cannot penetrate dense materials, future work may involve deeper architectures or the integration of GAN discriminators to further refine visual quality.

Summary updated

BibTeX
@inproceedings{hu2021quad,
  author = {Zheng Hu and Qiang Li and Gang Fu and Yuxiang Xing and Li Zhang},
  title = {{Quad-DIP for X-Ray Cargo Image Decomposition}},
  booktitle = {Workshops @ CVPR},
  year = {2021},
  doi = {10.1109/cvprw53098.2021.00491},
}
Figure