BSoNet: Deep Learning Solution for Optimizing Image Quality of Portable Backscatter Imaging Systems

Portable backscatter imaging systems (PBI) integrate an X-ray source and detector in a single unit, utilizing Compton scattering photons to rapidly acquire superficial or shallow structural information of an inspected object through single-sided imaging. The application of this technology overcomes the limitations of traditional transmission X-ray detection, offering greater flexibility and portability, making it the preferred tool for the rapid and accurate identification of potential threats in scenarios such as borders, ports, and industrial nondestructive security inspections. However, the image quality is significantly compromised due to the limited number of Compton backscattered photons. The insufficient photon counts result primarily from photon absorption in materials, the pencil-beam scanning design, and short signal sampling times. It therefore yields severe image noise and an extremely low signal-to-noise ratio, greatly reducing the accuracy and reliability of PBI systems. To address these challenges, this paper introduces BSoNet, a novel deep learning-based approach specifically designed to optimize the image quality of PBI systems. The approach significantly enhances image clarity, recognition, and contrast while meeting practical application requirements. It transforms PBI systems into more effective and reliable inspection tools, contributing significantly to strengthening security protection.

💡 Research Summary

The paper addresses the chronic problem of low‑signal, high‑noise images produced by portable backscatter imaging (PBI) systems, which limit their usefulness for rapid security inspections. Traditional enhancement techniques (spatial filtering, transform‑domain denoising) require manual parameter tuning and cannot cope with the complex, non‑stationary noise inherent to backscatter data. To overcome these limitations, the authors propose BSoNet, a deep‑learning framework that combines a hybrid backbone (BSformer) with a resolution‑adaptive network (RANet) and a self‑supervised training scheme based on Noise2Void.

BSformer merges convolutional neural networks (CNNs) for local texture extraction with a Vision‑Transformer encoder for global context modeling. Input images are first processed by multi‑scale CNN blocks, then reshaped into patches and fed to a multi‑head self‑attention module. The transformer captures long‑range dependencies across the entire backscatter image, while the CNN decoder restores fine‑grained details. This fusion enables the network to simultaneously suppress widespread noise and preserve subtle structural cues that are critical for material discrimination.

RANet solves the practical issue that PBI images vary in size and resolution due to changes in scan speed, tube voltage, and current. It dynamically resizes each input to the fixed resolution required by BSformer, applies the restoration pipeline, and then up‑samples the output back to the original dimensions using sub‑pixel convolution and learned up‑sampling blocks. This ensures consistent performance across diverse operational settings.

Because clean, high‑quality backscatter images are virtually unavailable, the authors adopt a label‑free training strategy. They employ Noise2Void, which masks random pixels and forces the network to predict their values from surrounding context, thereby learning a denoising function without ground‑truth references. To further improve robustness, synthetic Gaussian and impulse noise are added during training, encouraging the model to handle a wide spectrum of real‑world disturbances.

The experimental platform is the PBS‑140 portable backscatter system (max 140 kV, 50 µA, 1 mm resolution). Over 10,000 images were collected under varying voltages, currents, scan speeds, and material types (metal, plastic, composites). Quantitative evaluation using PSNR, SSIM, and mean opinion scores (MOS) shows that BSoNet outperforms classical filters (NLM, BM3D) and recent deep‑learning baselines (DnCNN, UNet) by an average of 4 dB in PSNR, 0.12 in SSIM, and 1.8 MOS points. The gains are especially pronounced in low‑dose settings (e.g., 80 kV, 20 µA) where noise levels exceed 30 %.

Ablation studies confirm that both the transformer component and the adaptive resizing are essential: removing the transformer drops performance by >2 dB, while omitting RANet leads to large variance across different scan configurations. The authors acknowledge that the transformer block is memory‑intensive, limiting real‑time deployment to high‑end GPUs, and that the current implementation handles only 2‑D slices. Future work includes lightweight ViT variants (e.g., Swin‑Transformer), extension to 3‑D volumetric data, and hardware acceleration on FPGA/ASIC platforms.

In summary, BSoNet delivers a substantial improvement in image quality for portable backscatter systems, making them more reliable for rapid threat detection at borders, ports, and industrial sites. Its combination of global‑local feature fusion, resolution adaptability, and self‑supervised learning represents a significant step forward in applying deep learning to low‑photon, high‑noise imaging modalities.