基于深度学习的辐射图像超分辨率重建方法  

Super-resolution Method for Radiation Image Based on Deep Learning

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

出　　处：《原子能科学技术》2017年第5期890-895,共6页Atomic Energy Science and Technology

摘　　要：安全检查系统中,数字化X射线摄影技术获得的辐射图像空间分辨率较低,影响图像的视觉效果。为了对单幅低分辨率辐射图像的空间分辨率进行提升,提出一种基于深度学习的超分辨率重建方法。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,拟合出低分辨率图像和高分辨率图像的映射关系。实验结果表明,与传统的超分辨率重建方法相比,本方法在量化指标和视觉效果上均有较大的改善,且具备较快的处理速度。研究结果表明,深度学习方法在辐射图像处理中有较大的潜力。In the security check system, the spatial resolution of radiation image generated by digital radiography is often so low that reduces the image quality. In this work, a super-resolution method based on deep learning was proposed. Using the convolution neural network with residual block, the method trained the radiation image sample in dataset and found the mapping function of low-resolution image to high-resolution image. The experiment result shows that the super-resolution method can deliver supe- rior performance compared with other traditional methods while maintaining an excellent speed. The study result indicates the great potential of deep learning in radiation image processing.

关 键 词：辐射图像 超分辨率重建 深度学习 

分 类 号：TL8[核科学技术—核技术及应用] TP391.4[自动化与计算机技术—计算机应用技术]