多尺度残差挤压和激励的双U舌图分割网络  

Multi-size Residual Squeeze and Excitation Attention Double U Network for Tongue Image Segmentation

作　　者：梁淑芬[1] 解竞一 吴岑 秦传波[1] LIANG Shu-fen;XIE Jing-yi;WU Cen;QIN Chuan-bo(School of Electronics and Information Engineering Wuyi University,Jiangmen 529020,China)

机构地区：[1]五邑大学电子与信息工程学院,广东江门529020

出　　处：《五邑大学学报（自然科学版）》2024年第2期55-63,共9页Journal of Wuyi University(Natural Science Edition)

基　　金：广东省国际和香港、澳门和台湾高端人才交流项目(2020A1313030021);2021年广东省研究生教育创新项目(YJS-SFJD-21-02);2022江门城市基础和理论科研类科技项目(2022JC01025)

摘　　要：在图像分割中,单次卷积和频繁的池化操作容易产生冗余信息或遗漏关键信息.本文设计了一种多尺度的残差挤压和激励注意力的双U形分割网络(MRSEA-DUNet)来解决上述问题.首先,该网络由两个U形的网络组成,分别是预编码网络和精确分割网络.为避免频繁的卷积和池化操作导致信息丢失或产生无效信息,提出了具有不同大小感受野的阶梯卷积模块(SCM),并采用并行结构,可以在不同尺度上捕获更丰富、更详细的特征.其次,还设计了一种残差挤压和激励注意力模块(RSEAM),可以通过空间和通道提高有效特征增益,消除冗余信息,并且提高了模型的整体鲁棒性.最后,为了减少了降采样操作的数量,简化了纵向复杂度.实验结果表明,本文MRSEA-DUNet模型的精度、Jaccard系数和Dice系数分别达到0.995 4、 0.979 4和0.989 5,均优于其他7种主流模型,优化了分割效果.In image segmentation,single convolution and frequent pooling operations very easily produce redundant information or miss key information.This paper designs a multi-size residual squeeze and excitation attention double U segment network(MRSEA-DUNet)to solve the problems.The network consists of two U-shaped networks,namely a precoding network and a precise segmentation network.The network has the following main advantages:(1)In order to avoid frequent convolution and pooling operations leading to information loss or invalid information,we propose a stepped convolution module(SCM),which adopts a parallel structure design and can capture richer and more detailed features at different scales,with receptive fields of different sizes.(2)This paper also designs a residual squeeze and excitation attention module(RSEAM),which can improve the useful feature gain through spaces and channels.It can not only eliminate some redundant information but also improve the overall robustness of the model.(3)Finally,we reduce the number of downsampling operations and simplify the vertical complexity.This paper verifies the performance of the MRSEA-DUNet in 2D tongue image databases. In the tongue image database, the Accuracy, Jaccard coefficient, and the Dice coefficient of the MRSEA-DUNet reach 0.995 4 , 0.979 4 , and 0.989 5 , respectively,which are better than those of the other seven models. Experimental results show that our method is superior to the original U-Net and other state-of-the-art methods in two different datasets.

关 键 词：医学影像分割 挤压与激励 注意力机制 感受野 多尺度 残差机制 

分 类 号：U271.91[机械工程—车辆工程]