基于改进MobileNet的ECG身份识别算法

ECG Identification Algorithm Based on Improved MobileNet

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.