基于神经网络的伽马射线定位算法及CUDA计算  

Gamma Ray Positioning Algorithm Based on Neural Network and CUDA Calculation

作　　者：彭敬 王磊[1] 卢位 杜雨昆 李浩炫 孟娜 邹科峰 PENG Jing;WANG Lei;LU Wei;DU Yukun;LI Haoxuan;MENG Na;ZOU Kefeng(College of Nuclear Technology and Automation Engineering,Chengdu University of Technology,Chengdu 610059,China)

机构地区：[1]成都理工大学核技术与自动化工程学院,成都610059

出　　处：《核电子学与探测技术》2024年第4期663-672,共10页Nuclear Electronics & Detection Technology

基　　金：国家自然科学基金项目(41874121)。

摘　　要：为实现便携式核辐射成像场景下利用神经网络进行伽马射线定位的快速GPU计算,提出了一种基于CUDA的伽马射线定位计算方法。首先描述了伽马射线探测器模型、MURA成像系统和神经网络射线定位模型,然后对FCNN射线定位模型进行并行计算分析,在Jetson Xavier NX平台下实现了基于CUDA的伽马射线定位计算,并验证了该计算方法的准确性和对其计算效率进行对比分析。实验结果表明,对于10000组事件数据,设计的基于CUDA的射线定位计算与CPU射线定位计算结果几乎一致,其实际计算耗时11.5 ms,约为CPU射线定位计算的加速比的364倍。In order to realize the fast calculation in GPU of gamma ray positioning using neural network in portable nuclear radiation imaging,We have proposed a gamma ray positioning calculation method based on CUDA.Firstly,the gamma ray detector model,MURA imaging system and neural network ray positioning model are described.Then,the parallelism analysis of FCNN ray positioning model is carried out,and an gamma ray positioning calculation method based on CUDA is implemented on the Jetson Xavier NX platform.Finally,the accuracy of the calculation method is verified,at the same time,the efficiency of the calculation method is compared.The experimental results for 10000 event data sets show as follows:the ray positioning calculation based on CUDA designed in this paper is almost consistent with the CPU ray positioning calculation,and its actual calculation takes 11.5 ms,in addition the acceleration ratio is about 364 times compared with the CPU ray positioning calculation.

关 键 词：辐射成像 射线定位 神经网络 推理计算 CUDA 

分 类 号：TP[自动化与计算机技术]