基于深度学习的编码孔径辐射成像去噪方法研究
Research on Deep Learning-Based Denoising Method for Coded-Aperture Radiation Imaging
杨彬艺 1肖宇峰 1严东 1刘锐 1周义枞2
作者信息
- 1. 西南科技大学信息与控制工程学院特殊环境机器人技术实验室
- 2. 西安电子科技大学空间科学与技术学院
- 折叠
摘要
编码孔径成像系统作为一种高精度的放射源定位及成像装置，可以准确定位放射源 的位置并重建其大致形状。针对其存在核监测中面临的实时性与成像质量难以兼顾的问 题，本文提出了一种基于深度学习的对最大似然期望最大化（MLEM）迭代重建的辐射图 像去噪方法。在低计数高噪声及短采集时间的场景下，MLEM重建易产生泊松高斯混合噪 声与背景伪影等，导致重建图像失真和热点偏移。利用Geant4软件对编码孔径成像系统进 行仿真，构建了三种不同难度噪声等级的数据集，并提出了一种融合双通道输入与门控的 改进型DnCNN网络。实验结果表明，在最高难度的噪声环境下，改进后的网络取得了 0.7894 的最高SSIM和28.31 dB 的最高PSNR，相较于经典U-Net分别相对提高了约11.4% 和2.74 dB，同时参数量为0.557 M，约为U-Net规模的1.6%。
Abstract
As a high-precision device for localizing and imaging radiation sources, the coded-aperture imaging system can accurately locate radiation sources and reconstruct their approximate profiles. To address the inherent trade-off between real-time performance and imaging quality in nuclear monitoring, this paper proposes a deep learning-based denoising method for radiation images reconstructed by the Maximum Likelihood Expectation Maximization (MLEM) iterative algorithm. In scenarios with low counting rates, high noise, and short acquisition times, MLEM reconstruction is highly susceptible to Poisson-Gaussian mixed noise and background artifacts, leading to image distortion and hotspot shifting. Using the Geant4 software to simulate the coded-aperture imaging system, datasets featuring three distinct levels of noise difficulty were constructed, and an improved DnCNN network integrating dual-channel inputs and a gating mechanism is proposed. Experimental results demonstrate that under the most severe noise conditions, the improved network achieves the highest SSIM of 0.7894 and the highest PSNR of 28.31 dB. Compared with the classical U-Net, these metrics represent relative improvements of approximately 11.4% and 2.74 dB, respectively. Simultaneously, the parameter size of the model is 0.557 M, which is approximately 1.6% of the U-Net's scale.关键词
伽马相机/深度学习/辐射成像/图像去噪Key words
Gamma camera/Deep learning/Radiation imaging/Image denoising引用本文复制引用
杨彬艺,肖宇峰,严东,刘锐,周义枞.基于深度学习的编码孔径辐射成像去噪方法研究[EB/OL].(2026-09-02)[2026-09-04].https://chinaxiv.org/abs/202609.00004.学科分类
TL8
