摘要
医学CT图像的超分辨率重构研究具有较大的实用价值。针对CT图像由于设备等原因存在的细节模糊,边缘不清晰、感知质量差等问题,提出一种多次上下采样的深度方格卷积网络。通过上下采样的二维结构,拓宽网络宽度与深度,增强不同尺度信息的深层依赖关系,促进不同尺度下的信息交互,从而充分利用原始图像信息重构出更多的高分辨率细节信息。采用全局深度联结与局部残差相结合的方式,将浅层网络信息反馈至深层网络,实现全局网络信息共享,提高训练时浅层网络特征映射在深层网络中的利用率,突出深度网络训练优势。实验结果表明,通过峰值信噪比与结构相似性指数将本文模型的重构结果和当前最先进的模型结果进行比较,该模型能恢复出最优的高分辨率图像,同时得到较高的重构图像感知质量。
There is great practical value in the research on super-resolution reconstruction of medical CT images.Aiming at the problems of CT image due to equipment and other reasons,such as fuzzy details,unclear edges and poor perceived quality,in this paper,a deep square convolution network with multiple ups and downs was proposed.Through the two-dimensional structure of up-sampling,the network width and depth were widened;the deep dependence of different scale information was enhanced;and the information interaction at different scales was promoted;so that more high-resolution details were reconstructed by using the original image information.The combination of global depth association and local residuals was used to feedback shallow network information to the deep network;the global network information sharing was realized;the utilization of shallow network feature mapping was improved in deep network during training;and the advantages of deep network training was highlighted.The experimental results showed that the peak signal-to-noise ratio and structural similarity index were compared the reconstruction results of the model with the current state-of-the-art model results.The optimal high-resolution image could be recovered and higher weight construct image perception quality was obtained.
作者
曹洪玉
刘冬梅
付秀华
张静
岳鹏飞
CAO Hong-yu;LIU Dong-mei;FU Xiu-hua;ZHANG Jing;YUE Peng-fei(School of Optoelectronic Engineering,Changchun University of Science and Technology,Changchun 130022)
出处
《长春理工大学学报（自然科学版）》
2020年第1期51-57,共7页
Journal of Changchun University of Science and Technology(Natural Science Edition)
基金
吉林省重大科技攻关专项(20190302095GX)。
关键词
上下采样
方格卷积网络
局部残差
全局联结
up and down sampling
square convolutional network
local residuals
global association
