基于残差网络的中子照相图像无参考质量评价方法研究  

Study on no-reference quality assessment method of neutron radiographic images based on residual network

作　　者：乔双[1] 李俊辉 赵辰一 张天 QIAO Shuang;LI Junhui;ZHAO Chenyi;ZHANG Tian(School of Physics,Northeast Normal University,Changchun 130024,China)

机构地区：[1]东北师范大学物理学院,长春130024

出　　处：《核技术》2021年第7期59-66,共8页Nuclear Techniques

基　　金：国家自然科学基金(No.11905028、No.11275046)资助。

摘　　要：目前中子照相图像的质量主要通过人类视觉系统(Human Visual System,HVS)来评估,而HVS无法作为中子成像系统优化参数的实时辅助。为了能够客观评价中子照相图像质量,以对中子成像系统参数优化提供辅助手段,采用残差网络(Residual Network,ResNet)模型,对中子照相图像进行无参考质量评价(No-reference Image Quality Assessment,NR-IQA)。首先对清晰的自然图像添加不同失真等级和失真类型的噪声,再利用梯度幅度相似性偏差(Gradient Magnitude Similarity Deviation,GMSD)方法对添加了噪声的图像进行质量分数标定来建立实验数据集。最后,通过训练ResNet以实现对中子照相图像的特征提取和质量评价。实验结果表明:模型在实验数据集的测试集和两组真实中子照相图像的质量预测上均有较好的表现,证明了该方法在中子照相图像质量评价上的应用潜力。[Background]The quality of neutron radiographic images is mainly evaluated by human visual system(HVS),but HVS cannot be used as a real-time auxiliary for optimization parameters of neutron imaging systems.[Purpose]This study aims to evaluate the quality of neutron radiographic images by no-reference image quality assessment(NR-IQA)and provide an effective approach for the parameters optimization of neutron imaging systems.[Methods]Firstly,the plain natural images distorted with different distortion levels and types were labelled with quality scores to construct an experimental dataset by the gradient magnitude similarity deviation(GMSD)method.Then,the residual network(ResNet)model was employed to evaluate the quality of neutron radiographic images without reference images.Finally,the goal of extracting features and assessing the quality of neutron radiographic images was achieved by training the ResNet.[Results]The model performs well in the test set of the experimental dataset and the quality prediction of the two group authentic neutron radiographic images.[Conclusions]The proposed quality assessment method could be used for the quality prediction of neutron radiographic images.

关 键 词：中子照相图像 无参考图像质量评价 残差网络 梯度幅度相似性偏差 

分 类 号：TL99[核科学技术—核技术及应用]