Interest Detection Method Research Of Single-Trial Fixation Related Potential In Guided Search Visual Task Using Deep Learning

The introduction of manual annotation into the target detection system and the use of human experience and cognitive ability is expected to break through the bottleneck of traditional artificial intelligence technology in detecting uncertain targets,blocking targets and difficult to describe targets,but it also limits the detection efficiency to some extent.Fixation Related Potential(FRP)is a kind of EEG induced by fixation events reflecting the human interested in cognitive behavioral process of electrical components,which induced by interested with the interested target of FRP has the obvious difference between the time domain waveform.The brain inductance based on FRP is interested in detecting technology,through computing devices of artificial tagging images evoked brain electrical signal preprocessing and classification,which can effectively obtain information related to the interested target in the form of without manual operation to improve the working efficiency of the manual annotation.Single-trial Classification is a key step in the interest detection technology based on EEG.In recent years,deep learning has been proved to be more accurate than the traditional shallow machine learning model in computer vision and natural language processing due to its ability to automatically acquire rich information.At present,studies on the classification methods of deep learning in the field of brain computer interface are mainly aimed at Motor Imagery,Steady State Visually Evoked Potential and other EEG signals,while few studies are aimed at FRP monitoring.In summary,the deep learning method of single-trial FRP classification is studied in this paper,and the following four aspects are mainly carried out:1.The FRP data were collected and pretreated.In this paper,the experiment of visual searching stimulus paradigm is designed,and the EEG and eye movement data of the subjects are collected synchronously in the self-built data acquisition system.By preprocessing the collected data,the single-trial FRP components of each subject were extracted.2.To solve the problem that the importance relationship between the convolutional kernels is not considered in the existing convolutional neural network structures for EEG classification,a single-trial FRP classification method based on the convolutional neural network with weighted assignment of kernels is proposed.By adding the feature channel weight allocation module to the existing network structure,the pertinence of the network model for feature extraction is improved.The experimental results show that compared with the existing convolutional neural networks for EEG signals,the proposed method can effectively improve the accuracy of single-trial FRP classification results.3.To solve the problem that it is difficult to effectively learn the second-order statistical feature of EEG signals in current research methods,a single-trial FRP classification method based on deep Riemann was proposed.By combining the geometric properties of symmetric positive definite Riemannian manifold space,multi-layer bilinear transformation is used to automatically extract and classify the second-order statistical features of EEG signals.The experimental results show that compared with the traditional Riemannian geometry-based shallow model method,the proposed method improves the results of single-trial FRP classification statistically significantly.4.To solve the problems of low signal-to-noise ratio(SNR)restricted depth Riemann network classification performance,a single-trial FRP classification method based on second order pooling depth Riemannian network with convolution feature is proposed.Firstly,the original EEG data were trained by convolutional neural network in the first stage.The multifrequency EEG features were acquired and the signal-to-noise ratio of the features was improved.Secondly,the convolution feature map is pooled in order to obtain its correlation information.Finally,the second stage of ground training is carried out by using the deep Riemannian network to improve the expression ability of convolution feature in Riemannian manifold space.The experimental results show that compared with the convolutional neural network using only first-order statistical features or the deep Riemann network using only second-order statistical features,the proposed method significantly improves the accuracy of single-trial FRP classification based on these two methods,and shows good robustness in the absence of training data.