基于仿真数据的辐射成像分类学习方法研究  

Research on digital radiography classification learning method based on simulation data

作　　者：陈雪睿 赵中玮 孙跃文[1,2] 李光超 丛鹏 CHEN Xuerui;ZHAO Zhongwei;SUN Yuewen;LI Guangchao;CONG Peng(Institute of Nuclear and New Energy Technology,Tsinghua University,Beijing 100084,China;Key Laboratory of Nuclear Detection Technology,Beijing 100084,China)

机构地区：[1]清华大学核能与新能源技术研究院,北京100084 [2]核检测技术重点实验室,北京100084

出　　处：《核技术》2019年第3期13-20,共8页Nuclear Techniques

基　　金：国家核工业科学基金项目(No.20154602098)资助~~

摘　　要：工业辐射成像领域中,系统获取的数据往往难以直接用于区分被检物体性质。分类学习作为机器学习的重要方向之一,可以充分挖掘不同类型的数据特征,在辐射成像领域有着巨大的应用潜力。然而工业辐射成像图像内容物和照射条件复杂,通过实际测量获取分类学习所需的完备训练集样本成本高、时间长。基于此背景提出使用数值模拟方法获得相应场景下的仿真数据,并通过仿真数据建立分类学习训练集的方法。分别采用基于集成学习的袋装树方法和基于K近邻分类方法对由仿真数据建立的训练集进行训练,抽取部分仿真数据作为测试集、实际工业数字辐射成像数据作为验证集对模型进行准确性验证。袋装树方法对测试集和验证集数据预测准确率分别为99.6%和81.25%;K近邻分类方法对测试集和验证集数据预测准确率分别为89.1%和50%。结果表明:袋装树方法对基于仿真数据的辐射成像分类学习具有较好的效果。[Background] As one of the important directions of machine learning, classification learning has great potential in industrial digital radiography, which can fully exploit different types of data features. Classification learning requires a large amount of labeled training data to train the prediction models. In view of the complex conditions of industrial digital radiography, obtaining complete training set samples by practical experiments is expensive and inefficient.[Purpose] This study aims to obtain complete training set data accurately and quickly by using simulation data.[Methods] First of all, numerical method was used to generate simulation data in corresponding scene. Then the training set for classification learning was established to be trained by using prediction model based on the Bagging Trees method and the KNN method. Finally, some of simulation data were assigned as test data set, and real industrial digital radiography data were used as verification data set to evaluate prediction models.[Results] The prediction accuracy of bagging trees method for the test set and verification set data is 99.6% and 81.25% respectively whilst KNN method for the test set and verification set data is 89.1% and 50% respectively.[Conclusion] The results show that the bagging trees method has a good effect on classification learning of radiography imaging based on simulation data.

关 键 词：数字辐射成像 仿真数据 机器学习 

分 类 号：TL99[核科学技术—核技术及应用]