An emotion recognition method based on left and right brain emotion lateralization

Abstract

The application belongs to the technical field of emotion recognition, and in particular to an emotion recognition method based on left and right brain emotion lateralization; the method specifically comprises the following steps: S1, electroencephalogram signal acquisition electrode and processing: obtaining standard electroencephalogram signals; S2, time domain feature and frequency domain feature extraction of electroencephalogram signals of left and right brains: extracting Hjorth and differential entropy features; S3, calculating edge weights of a dynamic graph convolutional neural network based on feature differences of left and right brains: calculating weights by using a CRITIC weight method; S4, calculating an adjacency matrix of the dynamic graph convolutional neural network based on inter-channel correlation of electroencephalogram signals: calculating phase locking values to construct the adjacency matrix; S5, emotion recognition by using an improved dynamic graph convolutional neural network: iteration and training; the application extracts the same features from the time domain and the frequency domain of the left and right brains respectively, thereby better recognizing that the left brain is more sensitive to positive emotions and the right brain is more sensitive to negative emotions, and solving the problem of low accuracy of emotion recognition in the two dimensions AV of valence and arousal.

Description

Technical Field

[0001] This invention relates to the field of emotion recognition technology, specifically to an emotion recognition method based on left-right brain emotional lateralization. Background Technology

[0002] Emotions are human reactions to objects or events in the environment, influencing various aspects of human life such as attention, memory, goal achievement, prioritization, cognitive motivation, communication with others, learning and development, emotional state, and effort motivation. The methods for judging human emotional states include external behavior and internal physiological signals. Because external behavior is often subjective, easily controlled, and potentially deceptive, current mainstream methods of emotion recognition rely on physiological signals. Electroencephalogram (EEG) signals, compared to other physiological signals, record the electrical activity of neurons in the brain, and emotion recognition is achieved through the fluctuations between neurons. Currently, numerous deep learning-based EEG emotion recognition methods have been proposed, but some problems remain to be solved. First, most emotion research ignores the cognitive patterns of human emotions, constructing network models solely from the EEG data itself. Second, there is an over-reliance on artificial features; single features are no longer sufficient for emotion recognition. Finally, there are connections between EEG channels, and most studies have neglected these connections, leading to low accuracy in emotion recognition. Summary of the Invention

[0003] (a) Technical problems to be solved

[0004] To address the shortcomings of existing technologies, this invention provides an emotion recognition method based on left-right brain emotional lateralization, which solves the problems mentioned in the background section.

[0005] (II) Technical Solution

[0006] To achieve the above objectives, the present invention specifically adopts the following technical solution:

[0007] An emotion recognition method based on left-right brain emotional lateralization, the method specifically includes the following steps:

[0008] S1: Electroencephalogram (EEG) signal acquisition electrodes and processing: The other end of the electrode cap is connected to a chip for potential difference recording, and the raw EEG signal is processed to obtain a standardized EEG signal.

[0009] S2: Extraction of temporal and frequency domain features of EEG signals from the left and right hemispheres: Extract Hjorth features from the left and right hemisphere channels of the obtained standard EEG signals in the time domain and differential entropy features in the frequency domain, and then linearly fuse them.

[0010] S3: Calculate the edge weights of the improved dynamic graph convolutional neural network based on the differences in features between the left and right hemispheres: Calculate the weights using the CRITIC weighting method based on the differences in fusion features between the left and right hemispheres obtained in the previous step, and input them into the graph convolutional neural network as edge weights.

[0011] S4: Adjacency matrix of the improved dynamic graph convolutional neural network based on the correlation between EEG signal channels: The phase-locked value of the obtained left and right brain fusion features is used as the adjacency matrix of the improved dynamic graph convolutional neural network.

[0012] S5: Emotion recognition through an improved dynamic graph convolutional neural network: The obtained left and right brain fusion features are used as the point features of the graph, and the weights of the left and right brain fusion features are used as the edge features of the graph. The point features, edge features, and adjacency matrix are used as the input of the improved dynamic graph convolutional neural network for iterative training to update the weights to meet the requirements of emotion recognition.

[0013] Furthermore, in S1, the test subject wears an electrode cap, the other end of which is connected to a chip to record the potential difference. The obtained potential difference is amplified, filtered, and processed by power frequency. The processed EEG signal is then adjusted to an EEG signal suitable for emotion recognition using EEGlab.

[0014] Furthermore, in step S2, the Hjorth parameter is extracted in the time domain to analyze the temporal variation characteristics of the signal, and the EMD empirical mode decomposition method is used to decompose the signal into multiple frequencies. By employing band-limited features to generate nonlinear data sequences, differential entropy features are extracted from the left and right symmetrical electrode pairs (Fp1,Fp2;Af3,Af4;F7,F8;F3,F4;FC5,FC6;FC1,FC2;T7,T8;C3,C4;CP5,CP6,CP1,CP2,P7,P8;P3,P4;PO3,PO4;O1,O2) in the frequency domain. The variation of the power spectrum with frequency is analyzed. Traditional differential entropy features cannot clearly explain the multi-dimensional trends of the signal. The proposed expression for differential entropy is as follows:

[0015]

[0016] in denoted as the variance of the original EEG signal, and dim as the dimension of the three-dimensional EEG data;

[0017] The proposed formula shows different trends in EEG signal changes across different dimensions. The higher the dimension, the more obvious the trend, and the more accurate the feature extraction in the frequency domain is for emotion recognition.

[0018] Furthermore, in S2, the time-domain and frequency-domain features of the left and right hemispheres are fused separately using a splicing method. The features obtained from the two domains are added together, and then the resulting dimension is reduced using principal component analysis and input into a trainable improved dynamic graph convolutional neural network.

[0019] Furthermore, the weights in S3 directly affect the final recognition accuracy. Feature parameters are learned through training an improved dynamic graph convolutional neural network, and the weights of the left and right brain features are calculated using the CRITIC weighting method. The weights are calculated using two parameters: contrast intensity and conflict index. Contrast intensity is represented by the standard deviation; the larger the standard deviation, the greater the fluctuation, and the higher the weight. The conflict index is represented by the correlation coefficient; the larger the correlation coefficient between the indicators, the smaller the conflict, and the lower the weight. In addition, the recognition result is defined as the dimension set of the AV model, which consists of the following dimensions: valence: it refers to whether the feeling is pleasant; arousal: it represents the degree of excitement. This two-dimensional model is defined as 0-9, with greater than 5 indicating positive emotions and less than 5 indicating negative emotions.

[0020] Furthermore, in the improved dynamic graph convolutional neural network in S4, the construction of the adjacency matrix directly affects the emotion recognition effect; different adjacency matrices result in different recognition accuracies. The phase-locked value, as an adjacency matrix, can calculate the changing trend between two different EEG channel signals. The formula for the phase-locked value is as follows:

[0021]

[0022] in Let N represent the instantaneous phase difference at the same time interval, and let N represent the number of experiments. This indicates that complex signals are obtained by using the phase with the help of Euler's formula.

[0023] Furthermore, in S5, emotion recognition is performed using an improved dynamic graph convolutional neural network. The obtained left and right brain fusion features are used as point features of the improved dynamic graph convolutional neural network, and the calculated weights of the left and right brain fusion features are used as edge features of the improved dynamic graph convolutional neural network. The adjacency matrix is ​​calculated using phase-locked values. These three parameters are used as inputs to the dynamic graph convolutional neural network. After network training and weight updates, the optimal accuracy on the AV model is obtained.

[0024] An emotion recognition device based on left-right brain emotional lateralization, characterized in that it comprises:

[0025] The signal acquisition module involves placing an electrode cap on the test subject and connecting the other end to a chip to record the potential difference in order to obtain the raw electroencephalogram (EEG) signal.

[0026] The signal processing module performs power frequency filtering on the obtained raw signal. The data processing directly uses the MATLAB toolbox EEGLAB to adjust and record the data, resulting in standardized EEG signals suitable for emotion recognition.

[0027] The feature extraction module extracts time-domain and frequency-domain features from the processed EEG signals. For the time-domain features, the Hjorth parameter, which reflects signal fluctuations, is selected. For the frequency-domain features, differential entropy is used to analyze the emotional value represented by peak values. These two types of features are linearly fused to obtain left-right brain fusion features, which are used as initial features for subsequent steps.

[0028] The CRITIC weight module is used to calculate the weights of the left and right brain fusion features. This method uses two parameters, contrast intensity and conflict index, to calculate the weights and updates the initial feature to more accurately describe the differences between the left and right brain.

[0029] The emotion recognition module uses the obtained left and right brain fusion features as point features of the improved graph convolutional neural network, calculates the adjacency matrix using the fusion features as the adjacency matrix of the improved graph convolutional neural network, and calculates the weight matrix as the edge features of the improved graph convolutional neural network. At the same time, it serves as the input of the improved graph convolutional neural network to train and update the weights of the left and right brain fusion features.

[0030] Furthermore, the CRITIC weight module specifically includes:

[0031] The left-right brain linear relationship unit calculates the difference in emotion between the two hemispheres based on the obtained left-right brain fusion features. The first parameter is the standard deviation, which is calculated based on the standard deviation of the left-right brain fusion features; a larger standard deviation is assigned a higher weight. The second parameter is the correlation coefficient, which is calculated based on the correlation coefficient of the left-right brain fusion features; a relatively larger correlation coefficient is assigned a lower weight. The differences between the left and right hemispheres are analyzed using these two parameters.

[0032] The weight unit is updated by inputting the processed left and right brain fusion feature weights into the improved graph convolutional neural network for iterative training and weight updates until the optimal accuracy in both valence and arousal dimensions is obtained.

[0033] A computer-readable storage device, characterized in that a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, it implements the steps of the emotion recognition method as described in any one of claims 1 to 4.

[0034] (III) Beneficial Effects

[0035] Compared with existing technologies, this invention provides an emotion recognition method based on left and right brain emotion lateralization, which has the following beneficial effects:

[0036] This invention extracts and fuses temporal and frequency domain features from both hemispheres, allowing for the separate extraction of identical features. This improves the identification of the left hemisphere's greater sensitivity to positive emotions and the right hemisphere's greater sensitivity to negative emotions, thus enhancing accuracy in both valence and arousal dimensions. Based on the standard deviation and correlation coefficient of the differences in left- and right-brain fusion features, the feature weights are calculated as edge features of a dynamic graph convolutional neural network. The phase-locked values ​​of the calculated left- and right-brain fusion features are used as the adjacency matrix of the dynamic graph convolutional neural network. By using the graph's point features, edge features, and adjacency matrix as input to the dynamic graph convolutional neural network, and through a learnable adjacency matrix, emotion recognition becomes more accurate. Combined with updating the fusion feature weights between the left and right hemispheres, this invention addresses the issue of low accuracy in emotion recognition across the valence and arousal dimensions (AV). Attached Figure Description

[0037] Figure 1 This is a schematic diagram of an emotion recognition method based on left and right hemisphere lateralization according to the present invention.

[0038] Figure 2 This is a schematic diagram of the time-domain and frequency-domain features of the EEG signals extracted from the left and right hemispheres in this invention.

[0039] Figure 3 This is a schematic diagram of the edge weight structure of the convolutional neural network for calculating the difference in features between the left and right hemispheres in this invention.

[0040] Figure 4 This is a schematic diagram of the structure for emotion recognition using an improved dynamic graph convolutional neural network, as described in this invention.

[0041] Figure 5 This is a channel visualization result diagram of the present invention;

[0042] Figure 6 The resulting diagram of the adjacency matrix constructed for this invention;

[0043] Figure 7 This is a diagram showing the left brain features of the present invention.

[0044] Figure 8 This is a diagram showing the left brain features of the present invention.

[0045] Figure 9 This is a schematic diagram of an emotion recognition device based on left and right hemispheres of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Example

[0048] like Figure 1-9 As shown in the figure, an embodiment of the present invention proposes an emotion recognition method based on left and right brain emotion lateralization, which specifically includes the following steps:

[0049] S1: Electroencephalogram (EEG) signal acquisition electrodes and processing: The other end of the electrode cap is connected to a chip for potential difference recording, and the raw EEG signal is processed to obtain a standardized EEG signal.

[0050] S2: Extraction of temporal and frequency domain features of EEG signals from the left and right hemispheres: Extract Hjorth features from the left and right hemisphere channels of the obtained standard EEG signals in the time domain and differential entropy features in the frequency domain, and then linearly fuse them.

[0051] S3: Calculate the edge weights of the dynamic graph convolutional neural network based on the differences in features between the left and right hemispheres: Calculate the weights using the CRITIC weighting method based on the differences in fusion features between the left and right hemispheres obtained in the previous step, and input them into the graph convolutional neural network as edge weights.

[0052] S4: Calculate the adjacency matrix of the dynamic graph convolutional neural network based on the correlation between EEG signal channels: Calculate the phase-locked value of the obtained left and right brain fusion features as the adjacency matrix of the dynamic graph convolutional neural network;

[0053] S5: Emotion recognition through an improved dynamic graph convolutional neural network: The obtained left and right brain fusion features are used as the point features of the graph, and the weights of the left and right brain fusion features are used as the edge features of the graph. The point features, edge features, and adjacency matrix are used as the input of the dynamic graph convolutional neural network for iterative training to update the weights to meet the requirements of emotion recognition.

[0054] Step S1, the extraction of temporal and frequency domain features of EEG signals from the left and right hemispheres, is specifically divided into:

[0055] S11: Data acquisition electrodes and processing;

[0056] S12: Preprocessing of EEG data;

[0057] S13: Extracting time-domain features of the left brain, selecting and extracting Hjorth parameter features in the time domain that can analyze signal fluctuations;

[0058] S14: Extract frequency domain features from the left brain, and select to extract differential entropy features in the frequency domain;

[0059] S15: Perform a linear fusion operation on the time domain and frequency domain features of the left brain;

[0060] S16: The operation of the right brain is the same as the operation of S12-S14 above;

[0061] S17: Obtain the fusion characteristics of the left and right hemispheres;

[0062] Step S2, which calculates the weights of the edges of the dynamic graph convolutional neural network based on the differences in features between the left and right hemispheres, specifically includes:

[0063] S21: Initialize features;

[0064] S22: Calculate the standard deviation and correlation coefficient of left and right brain characteristics;

[0065] S23: Assign left and right brain feature weights according to the CRITIC weighting method;

[0066] Step S3, which calculates the adjacency matrix of the dynamic graph convolutional neural network based on the correlation between EEG signal channels, specifically includes:

[0067] S31: Calculate the phase-locking value between channels based on the fusion characteristics of the left and right hemispheres;

[0068] S32: Construct the adjacency matrix using the obtained phase-locked values;

[0069] Step S4: Emotion recognition is performed by segmenting the emotion into parts using an improved dynamic graph convolutional neural network;

[0070] S41: The obtained left and right brain fusion features are used as point features of the graph;

[0071] S42: The obtained left and right brain fusion feature weights are used as edge features of the graph;

[0072] S43: Use point features, edge features, and adjacency matrix as input to a dynamic graph convolutional neural network;

[0073] S44: Perform iterations and training;

[0074] S45: Update the feature weights for the left and right hemispheres;

[0075] S46: Achieve the highest accuracy results in both valence and arousal dimensions;

[0076] In some embodiments, in S1, the test subject wears an electrode cap, the other end of which is connected to a chip to record the potential difference. The obtained potential difference is amplified, filtered, and processed by power frequency. The processed EEG signal is then adjusted to an EEG signal suitable for emotion recognition using EEGlab.

[0077] like Figure 2 As shown, firstly, EEG data is read and processed to arrange symmetrical electrodes. Then, the corresponding electrode channel in the left brain is selected, and the Hjorth parameter is extracted from the data in the time domain as a feature. Simultaneously, the frequency domain is transformed to the frequency domain using Fourier transform, and the EMD empirical mode decomposition method is used to decompose the signal into multiple frequencies. The same frequency band is used to extract differential entropy features as the second feature. Finally, the two features are linearly fused and spliced ​​together for the next step. The same processing method is applied to the right brain electrode channel to obtain two different fusion features for the left and right brains. In some embodiments, in S2, the Hjorth parameter is extracted in the time domain to analyze the characteristics of signal changes over time. In the frequency domain, differential entropy features are extracted from the left and right symmetrical electrode pairs (Fp1, Fp2; Af3, Af4; F7, F8; F3, F4; FC5, FC6; FC1, FC2; T7, T8; C3, C4; CP5, CP6, CP1, CP2, P7, P8; P3, P4; PO3, PO3; O1, O2) to analyze the change law of power spectrum with frequency. Traditional differential entropy features cannot clearly explain the trend of signal change in different dimensions. The proposed expression of differential entropy is as follows:

[0078]

[0079] in denoted as the variance of the original EEG signal, and dim as the dimension of the three-dimensional EEG data;

[0080] The proposed formula shows different trends in EEG signal changes across different dimensions. The higher the dimension, the more obvious the trend, and the more accurate the feature extraction in the frequency domain is for emotion recognition.

[0081] By employing a feature extraction process based on left and right brain channels of EEG signals, more suitable input features for emotion recognition can be obtained, thus solving the problem of low recognition accuracy caused by a single input feature and improving the accuracy of emotion recognition. At the same time, applying the same feature extraction method to the left and right brain channels can better illustrate that the left and right brains have different lateral responses to the same emotions, thereby improving the final recognition efficiency.

[0082] like Figure 3As shown, the obtained left and right brain fusion features are first initialized, and then the weights between the left and right brain features are calculated using the CRITIC weighting method. The weight values ​​are calculated in two ways. The first is to calculate the standard deviation to judge the degree of fluctuation of the EEG signal features. The larger the standard deviation, the more intense the signal fluctuation, reflecting the more significant the emotion in that range, and the greater the weight. The second is to calculate the correlation coefficient. The larger the correlation coefficient between the signals, the smaller the conflict, and the lower its weight. After initializing the feature weights, they are input into an improved dynamic graph convolutional neural network for training, obtaining the accuracy of the AV two dimensions. If the result is the best recognition accuracy, the loop ends; otherwise, it returns to recalculate the left and right brain feature weights until the best recognition effect is obtained. In some embodiments, in step S3, the size of the weights directly affects the final recognition accuracy. The feature parameters are learned through training the improved dynamic graph convolutional neural network, and the size of the left and right brain feature weights is calculated using the CRITIC weight method. In addition, the recognition result is defined as the dimension set of the VA model, which consists of the following dimensions: valence: it refers to whether the feeling is pleasant; arousal: it represents the degree of excitement. This two-dimensional model is defined as 0-9, with greater than 5 being positive emotion and less than 5 being negative emotion.

[0083] By calculating the edge weight features of a dynamic graph convolutional neural network based on the differences in features between the left and right hemispheres, a more accurate method for emotion recognition can be found from the differences between the symmetrical channels of the left and right hemispheres. By updating the weight information of the features of the left and right hemispheres to train the improved dynamic graph convolutional neural network, not only can the multidimensional features of EEG data be statistically analyzed better, but the efficiency of emotion recognition in both arousal and valence dimensions can also be achieved by leveraging the emotional bias between the left and right hemispheres.

[0084] In some embodiments, the improved dynamic graph convolutional neural network in S4, the construction of the adjacency matrix directly affects the emotion recognition effect; different adjacency matrices result in different recognition accuracies. The phase-locked value, as an adjacency matrix, can calculate the changing trend between two different EEG channel signals. The formula for the phase-locked value is as follows:

[0085]

[0086] in Let N represent the instantaneous phase difference at the same time interval, and let N represent the number of experiments. This indicates that complex signals are obtained by using the phase with the help of Euler's formula.

[0087] like Figure 4As shown, firstly, the fused features obtained from the left and right hemispheres are used as the point features of the graph, and the phase-locked values ​​are calculated as the adjacency matrix of the graph. The adjacency matrix is ​​constructed using the obtained phase-locked values. The weights of the left and right hemisphere features are calculated according to the CRITIC weighting method, and these weights are used as the edge features of the graph. At the same time, these three parameters are used as the input of the dynamic graph convolutional neural network. Then, the requirement of improving the accuracy of emotion recognition is achieved by iteratively training the learnable adjacency matrix and updating the weights of the left and right hemisphere features. In some embodiments, in step S5, emotion recognition is performed by an improved dynamic graph convolutional neural network. The three obtained parameters are used as the input of the dynamic graph convolutional neural network, and the network is trained and the weights are updated to obtain the best accuracy on the AV model.

[0088] Emotion recognition is achieved through an improved dynamic graph convolutional neural network. The three parameter features of the graph are used as input to the dynamic graph neural network. The phase-locked value is used as the adjacency matrix to analyze whether the symmetrical channels are in phase synchronization and whether there is an overall upward or downward trend in the spatial domain. By learning the adjacency matrix constructed with the phase-locked value, the correlation between channels in the left and right hemispheres can be better analyzed. The left and right hemisphere fusion feature weights obtained by the CRITIC weight method can obtain the emotion bias from the different responses of the left and right hemispheres to the same emotion. By updating the left and right hemisphere feature weights, the final accuracy of emotion recognition is improved.

[0089] like Figure 9 As shown, in some embodiments, an emotion recognition device based on left-right brain emotion lateralization includes:

[0090] The signal acquisition module involves placing an electrode cap on the test subject and connecting the other end to a chip to record the potential difference in order to obtain the raw electroencephalogram (EEG) signal.

[0091] The signal processing module performs power frequency filtering on the obtained raw signal. The data processing directly uses the MATLAB toolbox EEGLAB to adjust and record the data, resulting in standardized EEG signals suitable for emotion recognition.

[0092] The feature extraction module extracts time-domain and frequency-domain features from the processed EEG signals. For the time-domain features, the Hjorth parameter, which reflects signal fluctuations, is selected. For the frequency-domain features, differential entropy is used to analyze the emotional value represented by peak values. These two types of features are linearly fused to obtain left-right brain fusion features, which are used as initial features for subsequent steps.

[0093] The CRITIC weight module is used to calculate the weights of the left and right brain fusion features. This method uses two parameters, contrast intensity and conflict index, to calculate the weights and updates the initial feature to more accurately describe the differences between the left and right brain.

[0094] The emotion recognition module uses the obtained left and right brain fusion features as the point features of the graph convolutional neural network, calculates the adjacency matrix using the fusion features as the adjacency matrix of the graph convolutional neural network, and calculates the weight matrix as the edge features of the graph convolutional neural network. At the same time, it serves as the input of the improved graph convolutional neural network to train and update the weight matrix of the left and right brain fusion features.

[0095] In some embodiments, the CRITIC weight module specifically includes:

[0096] The left-right brain linear relationship unit calculates the difference in emotion between the two hemispheres based on the obtained left-right brain fusion features. The first parameter is the standard deviation, which is used to calculate the standard deviation of the left-right brain fusion features; a larger standard deviation is assigned a higher weight. The second parameter is the correlation coefficient, which is used to calculate the correlation coefficient between the left-right brain fusion features; a larger correlation coefficient is assigned a lower weight. These two coefficients are used to analyze the differences between the left and right hemispheres.

[0097] The weight unit is updated by inputting the processed left and right brain fusion feature weights into the improved graph convolutional neural network for iterative training and weight updates until the optimal accuracy in both valence and arousal dimensions is obtained.

[0098] A computer-readable storage device, wherein a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, it implements the steps of the emotion recognition method described above.

[0099] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An emotion recognition method based on left-right brain emotional lateralization, characterized in that: The method specifically includes the following steps: S1: Electroencephalogram (EEG) signal acquisition electrodes and processing: The other end of the electrode cap is connected to a chip for potential difference recording, and the raw EEG signal is processed to obtain a standardized EEG signal. S2: Extraction of temporal and frequency domain features of EEG signals from the left and right hemispheres: Extract Hjorth features from the left and right hemisphere channels of the obtained standard EEG signals in the time domain and differential entropy features in the frequency domain, and then linearly fuse them. S3: Calculate the edge weights of the dynamic graph convolutional neural network based on the differences in features between the left and right hemispheres: Calculate the weights using the CRITIC weighting method based on the differences in fusion features between the left and right hemispheres obtained in the previous step, and input them into the improved dynamic graph convolutional neural network as edge weights. S4: Adjacency matrix of the improved dynamic graph convolutional neural network based on the correlation between EEG signal channels: The phase-locked value of the obtained left and right brain fusion features is used as the adjacency matrix of the improved dynamic graph convolutional neural network. S5: Emotion recognition through an improved dynamic graph convolutional neural network: The obtained left and right brain fusion features are used as the point features of the graph, and the weights of the left and right brain fusion features are used as the edge features of the graph. The point features, edge features, and adjacency matrix are used as the input of the improved dynamic graph convolutional neural network for iterative training to update the weights to meet the requirements of emotion recognition.

2. The emotion recognition method based on left and right brain emotion lateralization according to claim 1, characterized in that: In step S1, the test subject wears an electrode cap, with the other end connected to a chip to record the potential difference. The obtained potential difference is then amplified, filtered, and processed using power frequency. The processed EEG signal is then adjusted using EEGLab to a level suitable for emotion recognition.

3. The emotion recognition method based on left and right brain emotion lateralization according to claim 1, characterized in that: In S2, the time-domain and frequency-domain features of the left and right hemispheres are fused using a splicing method. The features from the two domains are added together, and then the resulting dimension is reduced using principal component analysis and input into a trainable, improved dynamic graph convolutional neural network.

4. The emotion recognition method based on left and right brain emotion lateralization according to claim 1, characterized in that: The weights in S3 directly affect the final recognition accuracy. An improved dynamic graph convolutional neural network is used to train and learn feature parameters. The CRITIC weighting method is used to calculate the weights of the left and right brain features. The weights are calculated using two parameters: contrast intensity and conflict index. Contrast intensity is represented by the standard deviation; a larger standard deviation indicates greater volatility, resulting in a higher weight. Conflict index is represented by the correlation coefficient; a larger correlation coefficient indicates less conflict, resulting in a lower weight. The recognition result is defined as the dimension set of the AV model, consisting of the following dimensions: valence (indicating whether the feeling is pleasant) and arousal (representing the degree of excitement). The AV model is defined as 0-9, with values ​​greater than 5 indicating positive emotions and values ​​less than 5 indicating negative emotions.

5. The emotion recognition method based on left and right brain emotion lateralization according to claim 1, characterized in that: In S5, emotion recognition is performed using an improved dynamic graph convolutional neural network. The obtained left and right brain fusion features are used as point features of the dynamic graph convolutional neural network, and the calculated weights of the left and right brain fusion features are used as edge features of the dynamic graph convolutional neural network. The adjacency matrix is ​​calculated using phase-locked values. These three parameters are used as inputs to the dynamic graph convolutional neural network. After network training and weight updates, the optimal accuracy on the AV model is obtained.

6. An emotion recognition device based on left-right brain emotional lateralization, characterized in that, include: The signal acquisition module involves placing an electrode cap on the test subject and connecting the other end to a chip to record the potential difference in order to obtain the raw electroencephalogram (EEG) signal. The signal processing module performs power frequency filtering on the obtained raw signal. The data processing directly uses the MATLAB toolbox EEGLAB to adjust and record the data, resulting in standardized EEG signals suitable for emotion recognition. The feature extraction module is used to extract time-domain and frequency-domain features from the processed EEG signal. The time-domain feature selects the Hjorth parameter, which can reflect the signal fluctuation, and the frequency-domain feature selects the differential entropy feature to analyze the emotional value represented by the peak. The features of these two domains are linearly fused to obtain the left and right brain fusion features, which are used as the initial features for the next steps. The CRITIC weight module is used to calculate the size of the weight matrix of the left and right brain fusion features. It uses two parameters, contrast intensity and conflict index, to calculate the weight matrix. By updating this initial feature, the differences between the left and right brain can be described more accurately. The emotion recognition module uses the obtained left and right brain fusion features as point features of the improved dynamic graph convolutional neural network, calculates the adjacency matrix using the fusion features as the adjacency matrix of the improved dynamic graph convolutional neural network, and calculates the weight matrix as the edge features of the improved dynamic graph convolutional neural network. At the same time, it serves as the input of the graph convolutional neural network to train and update the weight matrix of the left and right brain fusion features.

7. An emotion recognition device based on left-right brain emotion lateralization as described in claim 6, characterized in that, The CRITIC weight module specifically includes: The left and right brain linear relationship unit calculates the difference in emotions between the left and right brains based on the obtained left and right brain fusion features. The first parameter is the standard deviation, which is used to calculate the standard deviation of the left and right brain fusion features. A larger standard deviation is assigned a higher weight. The second parameter is the correlation coefficient, which is used to calculate the correlation coefficient of the left and right brain fusion features. A larger correlation coefficient is assigned a lower weight. The differences between the left and right brains are analyzed through the above two parameters. The weight unit is updated by inputting the processed left and right brain fusion feature weights into the improved graph convolutional neural network for iterative training and weight updates until the optimal accuracy in both valence and arousal dimensions is obtained.

8. A computer-readable storage device, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the emotion recognition method as described in any one of claims 1 to 4.

Images