[Paper Review] BSoNet: Deep Learning Solution for Optimizing Image Quality of Portable Backscatter Imaging Systems

Linxuan Li, Wenjia Wei+4 more|arXiv (Cornell University)|Feb 12, 2026

Radiation Shielding Materials Analysis0 citations

TL;DR

BSoNet introduces BSformer and RANet, a self-supervised, Transformer-CNN hybrid framework to enhance image quality in portable backscatter imaging, addressing low photon counts and noise.

ABSTRACT

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.

Motivation & Objective

• Motivate improved image quality in portable backscatter imaging (PBI) systems with inherently low backscatter signals.
• Develop a deep learning framework that combines global (Transformer) and local (CNN) features for denoising and detail enhancement in PBI images.
• Ensure practical applicability by maintaining input-output dimensional consistency across varying scanning conditions via adaptive processing.
• Leverage self-supervised learning to train without clean labeled data, improving robustness to complex noise patterns.

Proposed method

• Propose BSformer, a Backscatter Optimization Transformer that fuses global attention and local CNN features for denoising and detail enhancement.
• Embed FLN and FFN components to enable multi-scale feature extraction and cross-scale feature fusion within BSformer.
• Introduce RANet for Resolution Adaptive processing to adjust image size before optimization and restore original dimensions after processing.
• Apply Noise2Void self-supervised strategy with additional Gaussian noise augmentation to learn from label-free data.
• Train end-to-end within the BSoNet framework to handle diverse scanning conditions and preserve image structure.

Experimental results

Research questions

• RQ1 Can BSformer effectively denoise and enhance backscatter images by combining Transformer-based global context with CNN-based local features?
• RQ2 Does the Resolution Adaptive Network (RANet) maintain image structure and dimensions before and after optimization across varied scanning conditions?
• RQ3 Can self-supervised Noise2Void training with noise augmentation yield robust backscatter image optimization without clean labels?
• RQ4 How does BSoNet perform under different voltages, scanning speeds, and durations using the PBS-140 system?
• RQ5 What are the practical gains in image clarity, recognition, and contrast for PBI applications?

Key findings

• The study introduces BSoNet, combining BSformer and RANet for backscatter image quality optimization.
• BSoNet leverages Noise2Void self-supervised learning with noise augmentation to train without clean labels.
• Data from the PBS-140 system includes 906 raw images, with 760 for training and 146 for testing.
• The framework is designed to adapt to diverse scanning conditions and preserve image structure across resolution changes.