基于改进MobileNet的ECG身份识别算法  

ECG Identification Algorithm Based on Improved MobileNet

作　　者：韦以嘉 张烨菲 张显飞[3] 赵治栋[2] WEI Yijia;ZHANG Yefei;ZHANG Xianfei;ZHAO Zhidong(School of Communication Engineering,Hangzhou Dianzi University,Hangzhou 310018,China;School of Cyber Science and Engineering,Hangzhou Dianzi University,Hangzhou 310018,China;School of Electronics and Information Engineering,Hangzhou Dianzi University,Hangzhou 310018,China)

机构地区：[1]杭州电子科技大学通信工程学院,浙江杭州310018 [2]杭州电子科技大学网络空间安全学院,浙江杭州310018 [3]杭州电子科技大学电子信息学院,浙江杭州310018

出　　处：《杭州电子科技大学学报（自然科学版）》2024年第4期37-43,共7页Journal of Hangzhou Dianzi University：Natural Sciences

基　　金：浙江省公益技术研究计划项目(LGG21F020002)。

摘　　要：信息安全在当今社会显得愈发重要,基于心电(Electrocardiogram, ECG)信号的身份识别技术因其“活体”采集的高防伪性,呈现出了独特的优势。为了在移动环境下实现更高效快捷的身份识别,提出了一种基于稀疏卷积(Sparse convolution, SP)和轻量化网络MobileNet的深度迁移识别模型SP-MobileNet。首先对原始ECG信号进行预处理,采用小波软阈值消噪后并将其盲分割成信号片段,采用广义S变换得到ECG时频图作为网络输入;其次构建基于SP-MobileNet的ECG识别模型,引入MobileNet,修改其卷积层为稀疏卷积计算策略,通过迁移学习实现从导联Ⅱ采集的大样本ECG数据训练到指尖采集的小样本ECG识别的无缝连接。实验结果表明,该算法可以高效快捷地进行ECG身份识别,在PhysioNet/Cinc Challenge 2017数据集上分别实现了98.00%的识别准确率和50.4 FPS的推理速度。Information security is becoming more and more important in society today.The identification technology based on the electrocardiogram(ECG)signal presents its unique advantages,due to its outstanding anti-counterfeiting performance on“living”detection.To achieve more efficient identification in mobile scenarios,a depth migration recognition model SP-MobileNet based on sparse convolution and lightweight network MobileNet is proposed.Firstly,the original ECG signal is pre-processed:denoised by wavelet soft threshold,blindly segmented into signal segments,and transformed into ECG time-frequency map as the input of the network by generalized S transform.Then,an ECG recognition model based on SP-MobileNet is built:a sparse convolution calculation strategy is adopted in the convolutional layers of MobileNet,along with transfer learning method,realizing the seamless connection from the large-sample ECG data training collected from Lead II to the small-sample ECG recognition collected by fingertips.Experimental results indicate that the proposed algorithm can perform ECG identification efficiently and quickly,achieving a recognition accuracy of 98.00%and an inference speed of 50.4 FPS on the PhysioNet/Cinc Challenge 2017 dataset.

关 键 词：心电信号 身份识别 轻量型网络 稀疏卷积 迁移学习 

分 类 号：TP309[自动化与计算机技术—计算机系统结构]