摘要
为了改善计算机断层扫描(CT)影像重建质量不高的问题,提出一种基于残差注意力聚合对偶回归网络(RAADRNet)的超分辨率CT重建方法。多特征下采样提取模块(MFDEB)通过平均池化、最大池化和卷积运算完成多特征下采样提取,在多特征融合后嵌入通道学习注意力(CLA)和空间学习注意力(SLA),同时并入前级融合特征提取图像的浅层特征。CLA、SLA分别引入通道权重特征学习以及激活函数1+tanh()完成特征提取。残差注意力聚合模块(RAAB)通过CLA嵌入残差网络构成的残差通道学习注意力模块(RCLAB)与SLA构成的空间特征融合模块(SFFB)联合提取图像的深层特征。原始网络在浅层特征与通过亚像素卷积放大的深层特征进行特征融合后完成重建。对偶网络进一步约束重建映射函数的解空间。实验表明,所提算法在重建图像的峰值信噪比(PSNR)和结构相似度(SSIM)上都得到了较好的提升。
A super-resolution computed tomography(CT) reconstruction method based on a residual attention aggregation dual regression network(RAADRNet) is proposed to improve the quality of CT image reconstruction.The multi-feature down-sampling extraction block(MFDEB) is used to complete multi-feature down-sampling extraction by employing average pooling,maximum pooling,and convolution operations,and channel learning attention(CLA) and spatial learning attention(SLA) are embedded after multi-feature fusion.Moreover,the shallow features of an image are extracted by combining the previous fusion features.CLA and SLA respectively introduce channel weight feature learning and activation function 1+tanh() to complete feature extraction.The residual attention aggregation block(RAAB) requires the use of the residual channel learning attention block(RCLAB) composed of a CLA-embedded residual network and the spatial feature fusion block(SFFB) composed of SLA for jointly extracting the deep features of the image.The primal network completes reconstruction after the feature fusion of shallow features and deep features amplified by sub-pixel convolution.The dual network further constrains the solution space of the reconstructed mapping function.Experiments show that the proposed algorithm improves the peak signal-to-noise ratio(PSNR) and structural similarity(SSIM) of the reconstructed image.
作者
范金河
吴静
何茂林
Fan Jinhe;Wu Jing;He Maolin(College of Information Engineering,Southwest University of Science and Technology,Mianyang 621010,Sichuan,China;Sichuan Key Laboratory of Special Environmental Robotics,Southwest University of Science and Technology,Mianyang 621010,Sichuan,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2023年第2期128-136,共9页
Laser & Optoelectronics Progress
基金
特殊环境机器人技术四川省重点实验室基金(13ZXTK07)。
关键词
图像处理
超分辨率计算机断层扫描重建
多特征下采样
通道学习注意力
空间学习注意力
残差注意力聚合
image processing
super-resolution computed tomography reconstruction
muti-feature down-sampling:channel learning attention
spatial learning attention
residual attention aggregation
