用于稳态视觉诱发电位目标识别的多尺度特征融合卷积神经网络方法

A Multi-Scale Feature Fusion Convolutional Neural Network Approach for Steady-State Visual Evoked Potential Target Recognition

针对传统稳态视觉诱发电位(SSVEP)脑电信号目标识别方法分类精度低、提取特征不充分、方法复杂且耗时等问题,提出一种基于多尺度特征融合卷积神经网络的SSVEP信号分类识别方法(SSVEP-MF)。利用小波变换将多通道SSVEP信号整合转化为二维图像作为输入样本集;建立多尺度特征融合卷积神经网络模型(MFCNN),该模型利用三层二维卷积核实现图像样本不同尺度特征的充分提取,构建多尺度特征融合单元对不同层级特征进行融合,并通过全连接等操作完成模型的训练;将样本集输入到MFCNN模型中实现脑电信号特征自适应提取及端到端分类。所提SSVEP-MF方法能够充分提取信号各层级特征,实现短时间视觉刺激下SSVEP信号的有效识别,并具有较高的目标识别效率。实验结果表明,在1 s刺激时长时,相比传统功率谱密度分析方法、典型相关分析方法以及普通卷积结构方法,所提方法的识别准确率分别提升了18.57%、20.08%及7.03%,有效提高了基于稳态视觉诱发电位范式下脑机接口的信号识别性能。

A multi-scale feature fusion convolutional neural network-based SSVEP signal classification and recognition method is proposed to solve the problems of low classification accuracy,inadequate feature extraction,complex and time-consuming methods of traditional steady-state visual evoked potential(SSVEP-MF)signal target recognition methods.Firstly,the wavelet transform is used to integrate the multi-channel SSVEP signals into two-dimensional images as the input sample set;secondly,a multi-scale feature fusion convolutional neural network model(MFCNN)is established,which uses a three-layer two-dimensional convolutional kernel to achieve sufficient extraction of features at different scales of image samples,constructs multi-scale feature fusion units to fuse features at different levels,and completes the training of the model through operations such as full connectivity;finally,the sample set is input to the MFCNN model to achieve adaptive extraction of EEG signal features and end-to-end classification.The proposed SSVEP-MF method can fully extract the features at each level of the signal,achieve effective recognition of SSVEP signals under short-time visual stimulation,and have high target recognition efficiency.The experimental results show that the recognition accuracy of the proposed method is improved by 18.57%,20.08%and 7.03%,respectively,compared with the traditional power spectral density analysis method,typical correlation analysis method and common convolutional structure method at 1 s stimulus duration,which effectively improves the signal recognition performance of brain-machine interface based on the steady-state visual evoked potential paradigm.