基于神经网络架构搜索与特征融合的小样本脉搏波分类方法

Small-sample Pulse Wave Classification Based on Neural Architecture Search and Feature Fusion

基于深度学习的脉搏波分类依赖大量有标注数据,现有脉搏波带有疾病标注的数据少、标注方法不统一,导致模型准确率低、泛化能力弱。针对此问题,提出一种基于神经网络架构搜索与特征融合的小样本脉搏波分类方法。首先,在并行的双维度拆分卷积分支与因果空洞卷积分支中进行态射搜索,每次搜索结束,获取超网络分支的子网络作为候选网络进行训练评估。双维度拆分卷积分支提取脉搏波横、纵向维度时空特征,因果空洞卷积分支提取脉搏波节律特征。然后,利用特征融合方法整合分支多尺度特征。最后,依据评估指标得到最佳网络模型完成分类。实验结果表明,所提方法在两个小样本脉搏波数据集上准确率为97.04%和95.96%,F1值为97.04%和95.95%,具有较好分类效果。

Classification of pulse wave based on deep learning relied on a large amount of labeled data.However,the existing pulse wave data with disease labels were small and the labeling methods were not uniform,which led to the problems of low accuracy and weak generalization ability of the model.To address this problem,a small-sample pulse wave classification method based on neural architecture search and feature fusion was proposed.Firstly,the morphism search was performed in the parallel super-network branches,the bi-dimensional split convolutional branch and the causal dilated convolutional branch.At the end of each search,the subnetworks in the super-network branches were obtained as candidate networks for training and evaluation.The spatio-temporal features in the horizontal and vertical dimensions of the pulse wave were extracted by the bi-dimensional split convolution branch,and the rhythmic features of the pulse wave were extracted by the causal dilated convolution branch.Then,the branching multiscale features were integrated using the feature fusion method.Finally,the best network model was obtained based on the evaluation index to complete the classification.The experimental results showed that the accuracy of the proposed method on two small sample pulse wave datasets was 97.04% and 95.96%,and the F1 was 97.04% and 95.95%,respectively,which could realize well classification results.