基于编码解码结构的微血管减压图像实时语义分割  

Real-time semantic segmentation of microvascular decompression images based on encoder-decoder structure

作　　者：白瑞峰 江山[1] 孙海江[1] 刘心睿 BAI Rui-feng;JIANG Shan;SUN Hai-jiang;LIU Xin-rui(Changchun Institute of Optics,Fine Mechanics and Physics,Chinese Academy of Sciences,Changchun 130033,China;University of Chinese Academy of Sciences,Beijing 100049,China;Department of Neurosurgery,The First Hospital of Jilin University,Changchun 130021,China)

机构地区：[1]中国科学院长春光学精密机械与物理研究所,吉林长春130033 [2]中国科学院大学,北京100049 [3]吉林大学第一医院神经肿瘤外科,吉林长春130021

出　　处：《中国光学（中英文）》2022年第5期1055-1065,共11页Chinese Optics

基　　金：吉林省科技发展计划项目(No.20200404155YY,No.20200401091GX);白求恩医学工程与仪器中心(长春)项目(No.Bqegczx2019047)。

摘　　要：针对真彩色微血管减压图像实时语义分割网络参数量大、语义分割精度低的问题,本文提出了一种适用于微血管减压场景的U型轻量级快速语义分割网络U-MVDNet(U-Shaped Microvascular Decompression Network),该网络由编码解码结构构成。在编码器中设计了轻型非对称瓶颈模块(LABM)对上下文特征进行编码,解码器中引入了特征融合模块(FFM),有效组合高级语义特征和低级空间细节。实验结果表明:对于微血管减压测试集,U-MVDNet在单NVIDIA GTX 2080Ti上的参数量只有0.66 M,平均交并比(mIoU)达到了76.29%,速度达到140 frame/s,且当输入图像尺寸为640×480时,U-MVDNet在嵌入式平台NVIDIA Jetson AGX Xavier上实现了实时(24 frame/s)语义分割。本文方法未使用任何的预训练模型,参数量少且推理速度快,语义分割性能优于其他对比方法,在分割精度和速度上做到了良好的平衡。同时,还可以方便地在嵌入式平台上开发和应用,性能优越,易于部署。Aiming at the problems of large parameters and low semantic segmentation accuracy of real-time semantic segmentation networks for true-color microvascular decompression(MVD)images.This paper proposes a U-shaped lightweight fast semantic segmentation network U-MVDNet(U-Shaped Microvascular Decompression Network)for MVD scenarios,which consists of encoder-decoder structure.A Light Asymmetric Bottleneck Module(LABM)is designed in the encoder to encode context features.Feature Fusion Module(FFM)is introduced in the decoder to effectively combine high-level semantic features and underlying spatial details.Experimental results show that for the MVD test set,U-MVDNet achieves 0.66 M parameters,76.29%mIoU(mean Intersection-over-Union),and 140 frame/s speed on NVIDIA GTX 2080Ti.And when input image size is 640×480,the real-time(24 frame/s)semantic segmentation is realized on NVIDIA Jetson AGX Xavier embedded development board.The proposed network has no pretrained model,fewer parameters,and fast inference speed.The semantic segmentation performance is superior to other comparison methods,and a good trade-off between segmentation accuracy and speed is achieved.Furthermore,U-MVDNet can also be easily developed and applied on embedded platform with superior performance and easy deployment.

关 键 词：微血管减压图像 编码解码 实时语义分割 U-MVDNet 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术] TH691.9[自动化与计算机技术—计算机科学与技术]