Research On Emotion Recognition Based On Parallel Computing

Empowering machine emotions and the research on emotion recognition is a hot topic in the contemporary era.EEG data must be obtained using EEG technology,but with the development of hardware technology,in order to obtain as much as possible EEG information and higher accuracy data,multi-channel(32,64,or 128-channel)high-resolution EEG is usually used Acquisition equipment.By increasing the number of channels and sampling frequency of the EEG acquisition device,on the one hand,the cost of the EEG acquisition device is increased,and the design difficulty and operation complexity are also increased.On the one hand,a large amount of EEG data will cause excessive calculation and affect emotions.Real-time recognition.Processing a large number of EEGs quickly has become a research focus of emotion recognition based on EEG signals.In feature extraction algorithms based on EEG signals,wavelet packet decomposition and wavelet decomposition are commonly used.In order to solve the complexity of decomposition and reconstruction using the traditional Mallat algorithm,this article applies the concept of "semi-wavelet packet" based on the Mallat algorithm to form a "semi-wavelet packet decomposition" algorithm that combines wavelet decomposition and wavelet packet decomposition.It solves the problem that wavelet decomposition is only valid for lowfrequency signals and the wavelet packet decomposition is redundant.For the first time,the convolution process of the semi-wavelet packet and the improved Mallat algorithm was used to decompose and reconstruct the EEG signals in the DEAP database.On the basis of ensuring accuracy,the optimized Mallat algorithm achieved a high rate of signal decomposition and low complexity of the decomposition algorithm Compared with the unoptimized Mallat algorithm,the optimized Mallat algorithm only needs 3/14 of the calculation amount of the traditional Mallat algorithm.Using the parallel algorithm to accelerate the optimized Mallat algorithm,the acceleration effect is more obvious.Based on the above-mentioned optimized Mallat algorithm,this article uses NVIDIA's GPU to accelerate the optimized Mallat algorithm using CUDA to obtain data in the five frequency bands of Delta,Theta,Alpha,Beta,and Gamma.Since the Alpha and Gamma frequency bands have a greater impact on emotion recognition,feature values are extracted from the Alpha and Gamma frequency bands.The principal component analysis algorithm is used to select features and CUDA is used for acceleration.To collect data from one subject's EEG experiment,the size of the data matrix is 32 * 7680.CUDA accelerates the decomposition and reconstruction of the Alpha band,extracts the features of the Alpha band,and selects the features of the Alpha band.The acceleration ratios are 2.37,0.49,and 3.92,respectively.Comparing the effects of CUDA acceleration of matrices of different sizes under the same conditions,it is concluded that the larger the EEG signal matrix,the better the CUDA acceleration effect,and the advantages of parallel computing will be further manifested.After extracting and selecting EEG signals in the DEAP database,training data required for emotion recognition is obtained.Extract the sample data provided by the online?ratings.xls table in the DEAP database and use BP neural network training to get the emotions corresponding to valence,arousal,anddominance in the emotion wheel.Define the emotion wheels 1 to 8 as positive emotions and 0,9 to 16 as negative The sentiment is defined as 1.Replace the original tags of the DEAP database with positive or negative emotions.The new labels are combined with the feature values extracted from the DEAP database to compose new data and the labels are used to classify sentiment through a multi-prediction deep Boltzmann machine.The multi-prediction depth Boltzmann machine used to recognize positive and negative emotions has an accuracy rate of 88.3%.