基于卷积神经网络的辐射图像降噪方法研究  

Radiation Image Denoising Based on Convolutional Neural Network

作　　者：孙跃文[1,2] 刘洪 丛鹏 李立涛[1,2] 向新程 郭肖静[1,2] 

机构地区：[1]清华大学核能与新能源技术研究院,北京100084 [2]核检测技术北京市重点实验室,北京100084 [3]中国海关管理干部学院,河北秦皇岛066004

出　　处：《原子能科学技术》2017年第9期1678-1682,共5页Atomic Energy Science and Technology

摘　　要：为了抑制探测器中统计涨落引起的噪声,提出了一种基于卷积神经网络的辐射图像降噪方法。该方法利用引入残差网络结构的卷积神经网络模型,对训练集中的辐射图像样本进行了训练,拟合出含噪声图像和无噪声图像的映射关系。实验结果表明,本文方法在降低统计噪声的同时保留了图像的细节。与传统的降噪方法相比,本文方法在量化指标和视觉效果上均有较大的改善。In this paper,a denoising method for statistical noise in detector based on convolutional neural network was proposed.Using a convolutional neural network model with residual blocks,the method training the radiation image samples in the training dataset and the mapping function of image with noise to image without noise was found.The experiment result shows that the method can reduce the statistical noise while maintaining the image details.The method delivers superior performance in both quantitative parameter and visual feeling compared with other traditional methods.

关 键 词：辐射图像 图像降噪 卷积神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]