Research On Facial Emotion Decoding Based On Real-time Functional Magnetic Resonance Imaging

The interpretation of brain emotion plays an important role in brain-computer interaction for higher cognition,so decoding the emotion of facial visual stimuli based on neural signals becomes an important part of brain visual processing research.Functional magnetic resonance imaging(fMRI)has become the mainstay of studying visual information processing in the brain due to its high temporal and spatial resolution advantages.At present,the research of visual information processing based on fMRI mainly focuses on the semantic and content analysis of visual stimuli perceived by the brain.In this paper,the fMRI signal of visual cortex is used to express facial emotion and decode emotion type in real time.It has important theoretical significance and practical value for the research direction of BCI and visual reconstruction of brain information.Focusing on the key issues of facial visual stimulus emotion decoding based on fMRI,combined with the principles of information processing in the visual cortex and deep learning,incremental learning and other methods,this paper conducts specific research on the fast registration of fMRI brain functional images,visual cortex voxel feature extraction for facial emotion analysis,real-time facial emotion decoding methods and other aspects.The main work is as follows:(1)According to the facial emotion decoding problem studied in this paper,we designed and carried out the related fMRI experiment,established the actual data sample library of facial emotion decoding;according to the characteristics of facial emotion expression in visual cortex,we realized the data preprocessing based on surface cortex,the localization of individual visual functional areas and the generation of visual cortex features,which provided a reliable data basis for the subsequent facial decoding analysis based on visual cortex brain activity.(2)Functional image registration with high accuracy and speed is a key prerequisite for constructing real-time decoding models based on voxel features.In this paper,a fast fMRI image registration method based on U-net is proposed.The traditional registration method relies on the similarity of image features to optimize the registration parameters of the target estimation function image,and the efficiency is relatively low when the data scale is relatively large.In this paper,we introduce the Unet depth network model to learn the geometric position change features of the brain in functional images from a large number of samples,and realize fast fMRI image registration and head movement correction across trials based on depth network.The experimental results show that the proposed method can achieve cross-trial head movement correction of fMRI data,and the registration time is significantly reduced compared with the traditional method,which plays an important role in improving the processing efficiency of real-time fMRI facial emotion recognition.(3)In order to extract the facial emotion information from the visual cortex of fMRI brain,a visual cortex voxel selection and feature extraction method based on Principal Component-Max Correlation(PC-MR)is proposed.The main difficulty in expressing different facial emotions through visual cortical voxel features is that the feature dimension is large,while the emotional information is more refined and weak.In this paper,based on the surface cortex of the brain to obtain the primary visual area and the voxel features of higher visual areas related to facial recognition,on this basis,we select the voxels with the largest difference between voxels by the principal component maximum correlation method,and retain more information entropy between voxel features by the maximum correlation-minimum redundancy method,so as to reduce the feature dimension while preserving the effective information in the visual cortex area as much as possible.Experimental results show that the PC-MR method proposed in this paper can select effective voxels and extract facial emotionrelated features.(4)Aiming at the problem of real-time processing of fMRI facial emotion decoding,a real-time updating method of SVM incremental classifier based on error sample triggering is proposed.In this paper,we first aim at the problem that the training time increases with the increase of training set data in real-time decoding,only the newly added samples are studied by the classical SVM incremental learning method to obtain the current update;secondly,on the basis of the classical incremental learning method,the classification interface is updated only for the error samples,which further improves the computational efficiency.Finally,the experimental results show that the proposed method can achieve stable decoding state quickly and achieve classification accuracy close to the classical method.It is stable and better than the classical algorithm model in computing time.