Prototype-based representation for cross-database eeg emotion recognition optimization method

Abstract

The application discloses a cross-database electroencephalogram emotion recognition optimization method based on a prototype representation pairwise learning, and comprises the following steps: constructing a prototype representation pairwise learning model, wherein the prototype representation pairwise learning model comprises a feature discriminator module, a prototype feature extraction module and a classifier module; respectively extracting sample features of source domain samples and target domain samples by using the feature discriminator module; extracting prototype features by using the prototype feature extraction module; acquiring interaction features according to the sample features and the prototype features by using the classifier module, calculating the similarity of a sample pair according to the interaction features by using a pairwise learning method, acquiring a label of the sample pair, iteratively updating the prototype representation pairwise learning model according to the similarity of the sample pair and the label of the sample pair, calculating emotion classification performance according to the interaction features, and optimizing the prototype representation pairwise learning model. The method can improve the robustness and generalization, optimize emotion recognition, and improve the accuracy of emotion recognition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electroencephalogram analysis, and particularly relates to a cross-database electroencephalogram emotion recognition optimization method based on prototype representation paired learning. BACKGROUND

[0002] Traditional emotion recognition mainly relies on facial expressions, speech signals, body movements and other non-physiological signals, which are highly subjective and generally have low accuracy. Electroencephalogram (EEG) signals, as a physiological signal, have the characteristics of being difficult to fake and the advantages of superior real-time performance and objectivity, and can provide more direct and objective clues for understanding and evaluating emotional states, so it has attracted more and more attention from researchers in different fields such as computer science, neuroscience and signal processing. In recent years, researchers have devoted to developing emotion recognition models based on electroencephalogram (EEG), using transfer learning and domain adaptation technology, which have shown very good performance in cross-subject emotion recognition tasks within a single database, and have also been successfully applied to large public electroencephalogram emotion datasets such as SEED and SEED-IV. However, due to the large differences between different fields (databases), traditional electroencephalogram emotion recognition mainly focuses on intra-individual or cross-task, and when dealing with cross-database electroencephalogram emotion recognition tasks, in addition to dealing with individual differences, it also needs to face environmental differences, equipment differences and many other factors of different database electroencephalogram acquisition, which leads to a significant decline in the performance of these emotion recognition methods. SUMMARY

[0003] The technical problem to be solved by the present application is to provide a cross-database electroencephalogram emotion recognition optimization method based on prototype representation paired learning, so as to improve the accuracy of emotion recognition.

[0004] To solve the above technical problems, the purpose of the present application is realized by the following technical scheme: a cross-database electroencephalogram emotion recognition optimization method based on prototype representation paired learning is provided, comprising the following steps: constructing a prototype representation paired learning model, the prototype representation paired learning model comprising a feature discriminator module, a prototype feature extraction module and a classifier module; using the feature discriminator module to extract sample features of source domain samples and target domain samples respectively; using the prototype feature extraction module to extract prototype features; using the classifier module to obtain interaction features according to the sample features and the prototype features, calculating the similarity of sample pairs according to the interaction features by a paired learning method, obtaining the labels of the sample pairs, iteratively updating the prototype representation paired learning model according to the similarity of the sample pairs and the labels of the sample pairs, and calculating the emotion classification performance according to the interaction features; and optimizing the prototype representation paired learning model.

[0005] The beneficial technical effect of the present application is that the cross-database electroencephalogram emotion recognition optimization method based on prototype representation pairwise learning of the present application constructs a prototype representation pairwise learning model including a feature discriminator module, a prototype feature extraction module and a classifier module, uses the feature discriminator module to extract sample features of source domain samples and target domain samples respectively, uses the prototype feature extraction module to extract prototype features, uses the classifier module to obtain interaction features according to the sample features and the prototype features, calculates the similarity of a sample pair according to the interaction features through a pairwise learning method, obtains the label of the sample pair, iteratively updates the prototype representation pairwise learning model according to the similarity of the sample pair and the label of the sample pair, and calculates the emotion classification performance according to the interaction features; the prototype representation pairwise learning model is optimized, the robustness and the generalization are improved, the similarity between each sample in the feature space is obtained, a specific model structure and a training method are designed to optimize the emotion recognition, and the accuracy of the emotion recognition is improved. BRIEF DESCRIPTION OF DRAWINGS

[0006] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0007] Figure 1 The flowchart of the cross-database electroencephalogram emotion recognition optimization method based on prototype representation pairwise learning provided by the embodiments of the present application. DETAILED DESCRIPTION

[0008] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0009] Please refer to Figure 1 as shown, Figure 1 The flowchart of the cross-database electroencephalogram emotion recognition optimization method based on prototype representation pairwise learning provided by the embodiments of the present application. The cross-database electroencephalogram emotion recognition optimization method based on prototype representation pairwise learning includes the following steps:

[0010] Step S10, a prototype representation pairwise learning model is constructed, and the prototype representation pairwise learning model includes a feature discriminator module, a prototype feature extraction module and a classifier module;

[0011] Step S20, sample features of the source domain samples and the target domain samples are extracted respectively by using the feature discriminator module;

[0012] Step S30, prototype features are extracted by using the prototype feature extraction module;

[0013] Step S40, interaction features are obtained according to the sample features and the prototype features by using the classifier module, the similarity of the sample pair is calculated according to the interaction features by using the pair-wise learning method, the label of the sample pair is obtained, the prototype representation pair-wise learning model is iteratively updated according to the similarity of the sample pair and the label of the sample pair, and the emotion classification performance is calculated according to the interaction features; wherein, the label of the sample pair includes a real label and a pseudo label, and is a judgment result of the two samples of the sample pair belonging to the same emotion category, and the value is only 0 or 1; when the value of the label of the sample pair is 1, it indicates that the two samples of the sample pair belong to the same emotion category; when the value of the label of the sample pair is 0, it indicates that the two samples of the sample pair do not belong to the same emotion category.

[0014] Step S50, the prototype representation pair-wise learning model is optimized.

[0015] Wherein, the sample is electroencephalogram data in different electroencephalogram emotion databases, and the prototype representation pair-wise learning based cross-database electroencephalogram emotion recognition optimization method constructs a prototype representation pair-wise learning model including a feature discriminator module, a prototype feature extraction module and a classifier module, extracts sample features of source domain samples and target domain samples respectively by using the feature discriminator module, extracts prototype features by using the prototype feature extraction module, obtains interaction features according to the sample features and the prototype features by using the classifier module, calculates the similarity of the sample pair according to the interaction features by using the pair-wise learning method, obtains the label of the sample pair, iteratively updates the prototype representation pair-wise learning model according to the similarity of the sample pair and the label of the sample pair, and calculates the emotion classification performance according to the interaction features; the prototype representation pair-wise learning model is optimized, the robustness and the generalization are improved, the similarity between each sample in the feature space is obtained, a specific model structure and a training method are designed to optimize the emotion recognition, and the accuracy of the emotion recognition is improved.

[0016] Preferably, the step S20 is specifically:

[0017] The feature extractor of the feature discriminator module extracts corresponding sample features of the source domain samples and the target domain samples respectively, that is, the feature extractor of the feature discriminator module extracts domain-invariant sample features of the source domain samples and the target domain samples, and the feature discriminator of the feature discriminator module distinguishes the domain to which the sample features belong, that is, the feature discriminator of the feature discriminator module distinguishes whether the sample corresponding to the sample features belongs to the source domain or the target domain.

[0018] The feature discriminator module can include a feature extractor and a feature discriminator. The feature extractor is configured to extract sample features from the samples. The feature extractor includes two hidden layers. The feature discriminator is configured to distinguish whether the samples corresponding to the sample features belong to the source domain or the target domain. The feature extractor can represent the sample features based on a domain adaptation neural network (DANN) using a domain adversarial training method. The domain adaptation and adversarial training are used to minimize the difference between the sample features of the source domain samples and the target domain samples. The relevant features of the domain-invariant attributes are extracted. The influence of the individual differences of the electroencephalogram signals can be effectively alleviated. The robustness and generalization ability of the model are improved. The difference between the source domain samples and the target domain samples can be minimized by calculating a domain adversarial loss function. The domain adversarial loss function can be a binary cross-entropy loss function, which is represented by formula (1):

[0019]

[0020] In the formula, l disc (·) represents the domain adversarial loss function, X S represents the source domain sample set, X T represents the target domain sample set, represents the i-th sample in the source domain sample set, represents the j-th sample in the target domain sample set, n represents the number of samples in the source domain, m represents the number of samples in the target domain, D(·) represents the domain discriminator, f(·) represents the feature extraction function, and represents the sample feature of the i-th sample in the source domain sample set, represents the sample feature of the j-th sample in the target domain sample set.

[0021] Preferably, the step S30 is specifically:

[0022] The prototype feature extraction module is used to extract the prototype features from the sample features of the source domain samples.

[0023] Each emotion category from the source domain or the target domain corresponds to a prototype feature. The samples belonging to the same emotion category have the same prototype feature. The prototype feature is the most essential attribute of all sample features of the emotion category. The sample features belonging to the emotion category are distributed around the prototype feature. The prototype feature is the centroid of all sample features of the emotion category, that is, the average value of all sample features of the corresponding domain in the emotion category. The extraction of the prototype feature requires the guidance of the emotion category. Since the target domain emotion category is invisible, the prototype feature is extracted from all samples in the source domain. The extraction of the prototype feature can improve the robustness and generalization. The specific steps of the step S30 are:

[0024] The prototype feature extraction module is used to obtain an average vector of sample features of the source domain samples corresponding to each emotion category as a prototype feature corresponding to each emotion category.

[0025] Specifically, the average vector of the prototype features corresponding to each emotion category is obtained by using formula (2):

[0026]

[0027] In the formula, ψ 1:z represents a set of prototype features of each emotion category, ψ 1:z ={ψ1, ψ2, ψ3,..., ψ z}, the elements ψ · in the set are prototype features of each emotion category, z represents the number of emotion categories, represents the i-th sample of the emotion category c in the source domain, represents the number of samples of the emotion category c, and f(·) represents a feature extraction function. represents the sample feature of the i-th sample of the emotion category c in the source domain.

[0028] Preferably, the classifier module in the step S40 is used to obtain interaction features according to sample features and prototype features, specifically:

[0029] A bilinear transformation strategy is used to represent the interaction between the sample features and the prototype features to obtain the interaction features; wherein the interaction features are represented by the sample features and the prototype features. The interaction features can be represented by formula (3):

[0030] Γ=(F·ψ c ·θ)(3)

[0031] In the formula, Γ represents the interaction features, · represents the inner product operation of the matrix, F represents the sample feature, which is obtained by calculating the sample through the feature extraction function f(·), ψ c represents the prototype feature of the emotion category c, and θ represents the bilinear transformation matrix, wherein the parameters of the bilinear transformation matrix are updated iteratively in the model training process. The interaction features Γ represent the probability that the sample belongs to each emotion category. For example, when the interaction features of the sample are [0.22, 0.65, 0.13], since the probability of the second position is the largest, the sample will be classified into the second emotion category. The emotion recognition accuracy can be calculated according to the interaction features.

[0032] Specifically, in the step S40, the similarity of the sample pair is calculated according to the interaction features by using the pair-wise learning method, the similarity of the sample pair is obtained by calculating the cosine similarity according to the interaction features, and the target loss function of the pair-wise learning method can be represented by formula (4):

[0033]

[0034] wherein, l pair denotes the target loss function of the pair-wise learning method, a denotes the number of samples, denotes the interaction feature, i and j denote the serial numbers of the two samples in the sample pair respectively, denotes the similarity between the two samples in the sample pair, and denotes the probability that the two samples belong to the same emotion category, which is between 0 and 1, and the closer to 1 indicates that the two samples in the sample pair are more likely to belong to the same emotion category, and the closer to 0 indicates that the two samples in the sample pair are less likely to belong to the same emotion category. ij denotes the similarity between the two samples in the sample pair, i.e. the similarity between the two samples in the sample pair, and denotes the probability that the two samples belong to the same emotion category, which is between 0 and 1, and the closer to 1 indicates that the two samples in the sample pair are more likely to belong to the same emotion category, and the closer to 0 indicates that the two samples in the sample pair are less likely to belong to the same emotion category. i denotes the probability that the two samples belong to the same emotion category, which is between 0 and 1, and the closer to 1 indicates that the two samples in the sample pair are more likely to belong to the same emotion category, and the closer to 0 indicates that the two samples in the sample pair are less likely to belong to the same emotion category. j denotes the label of the sample pair, which is the result of determining whether the two samples in the sample pair belong to the same emotion category, and the value is only 0 or 1, i.e. μ ij = 1 indicates that the two samples in the sample pair belong to the same emotion category, and μ ij = 0 indicates that the two samples in the sample pair do not belong to the same emotion category. ij

[0035] wherein, the source domain belongs to supervised learning, and the emotion label is visible, so in the source domain, the label of the sample pair is the real label, and the label of the sample pair in the source domain can be obtained by using formula (5):

[0036]

[0037] wherein, R(·) denotes the rounding function, y s denotes the real label obtained by encoding the source domain sample, denotes the similarity between the two samples in the sample pair, and denotes the interaction feature of the source domain sample. ij s

[0038] The target domain belongs to unsupervised learning, and the emotion label is invisible, so in the target domain, the label of the sample pair is the pseudo label, and the adaptive nonlinear dynamic updating threshold is used to calculate the label of the sample pair in the target domain, which is represented by formula (6):

[0039]

[0040] wherein, denotes the similarity between the two samples in the sample pair, denotes the interaction feature of the target domain sample, and denotes the adaptive nonlinear dynamic updating threshold. ij t u d ​​​​​​respectively represent the upper threshold and the lower threshold of the adaptive nonlinear dynamic update, when the similarity of the sample pair exceeds the upper threshold, it can be considered that the sample pair belongs to the same emotion category, and the label of the sample pair is 1, while when the similarity of the sample pair is lower than the lower threshold, it can be considered that the sample pair does not belong to the same emotion category, and the label of the sample pair is 0, and 0 < η d ≤ η u < 1.

[0041] In order to improve the robustness and generalization performance of the model, and improve the quality of the model training data, in the target domain, the sample pairs between the upper threshold and the lower threshold do not participate in the model training, but with the continuous iteration of the model, the threshold will be adaptively and nonlinearly dynamically updated, the upper threshold will gradually decrease, and the lower threshold will gradually increase, so that more and more sample pairs meeting the requirements participate in the training.

[0042] The updating process of the adaptive nonlinear dynamic updating threshold is represented by formula (7):

[0043]

[0044] In the formula, Epoch_size represents the number of training rounds, represents the upper threshold of the threshold at time t, represents the upper threshold of the threshold at time t-1, represents the lower threshold of the threshold at time t, represents the lower threshold of the threshold at time t-1. The threshold at the current time is affected by the threshold at the last time.

[0045] The total target loss function of the prototype representation pair learning model is calculated by formula (8):

[0046]

[0047] In the formula, L pl represents the total target loss function of the prototype representation pair learning model, represents the pair learning loss value of the source domain, represents the pair learning loss value of the target domain, l disc represents the domain adversarial loss value of the feature discriminator module, and α and β represent hyperparameters. Among them, the pair learning loss value of the source domain and the pair learning loss value of the target domain can be calculated according to formula (4) to formula (7), and the domain adversarial loss value l discIt can be calculated according to formula (1). The prototype representation pairwise learning model is trained by the target loss function of the prototype representation pairwise learning model. During the training process, domain adversarial and pairwise learning methods are used to improve the emotion recognition performance of the model. Moreover, during the training process of the model, the labels of the sample pairs in the target domain are completely invisible, so the erroneous labels and noise labels in the target domain have a limited impact on the model performance.

[0048] Preferably, step S50, namely optimizing the prototype representation pairwise learning model, specifically includes:

[0049] Based on the sample features, the Local Maximum Mean Difference (LMMD) method is used to measure the similarity between the source domain samples and the target domain samples in the feature space. The feature discriminator module is optimized to obtain a prototype representation pairwise learning model with local maximum mean difference. The unbiased estimation objective function of local maximum mean difference is expressed by formula (9):

[0050]

[0051] In the formula, Let X represent the feature mapping function. S and X T Let Hk and n represent the source domain sample set and the target domain sample set, respectively. Hk represents the reproducible Hilbert space with feature kernel k, which is the feature mapping space. The feature kernel k is a distance function that defines the distance between samples in the Hilbert space and samples in the source and target domains. z represents the number of emotion categories, and n... c m represents the number of all samples in the source domain with sentiment category c. c This represents the number of all samples in the target domain with sentiment category c. Let represent the probability weight of the i-th source domain sample belonging to the source domain given the emotion category c. This represents the i-th sample in the source domain sample set. This represents the probability weight that the j-th target domain sample belongs to the target domain given the emotion category c. Let represent the j-th sample in the target domain sample set, and LMMD represent the Local Maximum Mean Difference. By obtaining the probability of each sample belonging to the corresponding domain for each emotion category, the feature contributions of the two domains can be balanced, thus learning the features of the target domain more effectively. The sum of the probability weights of all samples belonging to a certain emotion category in the source domain is always 1, i.e. The sum of the probability weights of all samples belonging to a certain emotion category in the target domain is always 1, that is... Measuring the similarity between source and target domain samples, i.e., assessing the similarity of samples based on the spatial distribution of their features, involves calculating the distance between samples in the feature space; the closer the samples are, the more similar they are. Using the Local Maximum Mean Difference (LMD) method to measure the similarity between source and target domain samples in the feature space can reduce the distance between subdomains of the same emotion category in the source and target domains, achieving alignment of subdomains of the source and target domains in the feature space. The classifier module of the pairwise learning model, which uses the LMD prototype, is represented by the Ada classifier, which employs a gradient descent algorithm with an adaptive learning rate (Adam).

[0052] Preferably, step S50, namely optimizing the prototype representation pairwise learning model, specifically includes:

[0053] The contrast domain difference (CDD) is calculated based on the sample features, and the feature discriminator module is optimized to obtain a pairwise learning model of the prototype representation with contrast domain difference; wherein, the objective function of contrast domain difference is expressed by formula (10):

[0054]

[0055] In the formula, z represents the number of emotion categories, and Hk represents the reproducible Hilbert space with feature kernel k.

[0056] Denotes the feature mapping function, d cc (·) represents the intra-domain difference between samples of the same emotion category, d cc '(·) represents the inter-domain difference between samples of different emotion categories. The inter-domain difference and the intra-domain difference are optimized in opposite directions, so that the features of samples of the same emotion category are close to each other in the feature space, while the features of samples of different categories are far apart. The set of predicted labels for samples in the target domain. CDD stands for Contrast Domain Difference. The classifier module of the pairwise learning model with contrast domain difference is called the Ada classifier, which is a classifier using the gradient descent algorithm with the Adam adaptive learning rate.

[0057] Specifically, the classifier module representing the pairwise learning model includes an Ada classifier and an Rms classifier, wherein the Ada classifier is a classifier using the Adam adaptive learning rate gradient descent algorithm, and the Rms classifier is a classifier using the RMSprop gradient descent algorithm.

[0058] Preferably, step S50, namely optimizing the prototype representation pairwise learning model, includes:

[0059] Optimize the classifier module.

[0060] Specifically, the optimized classifier module includes:

[0061] Supervised learning is performed on the feature discriminator module and the classifier module to train the prototype representation pairwise learning model so that the model can store basic classification parameters, enabling the model to obtain task-specific basic emotion recognition performance, and the basic training objective loss function of the prototype representation pairwise learning model is calculated.

[0062] The basic training objective loss function of the prototype representation pairwise learning model is calculated using formula (11):

[0063]

[0064] In the formula, L1 represents the basic training objective loss function of the prototype representation pairwise learning model, l disc This represents the domain adversarial loss of the feature discriminator module. This represents the pairwise learning loss value of the source domain. and Let represent the pairwise learning loss values ​​of the target domain for the Ada classifier and the Rms classifier, respectively. α, β, and γ represent hyperparameters used to assign weights to each loss value; where represents the pairwise learning loss value of the source domain. Pairwise learning loss values ​​of the target domain of the Ada classifier Pairwise learning loss values ​​of the target domain and the RMS classifier The domain adversarial loss value l of the feature discriminator module can be calculated according to formulas (4) to (7). disc It can be calculated according to formula (1);

[0065] Lock the parameters of the feature discriminator module, train the classifier module, and calculate the classifier difference according to formula (12):

[0066]

[0067] In the formula, l diff (X T ) represents the difference loss between classifiers, Mean represents the average loss, and X represents the difference loss between classifiers. T Let x represent the target domain sample set. t Let |·| represent the target domain sample, |·| represent the absolute value function, φ(·) represent the similarity between two samples in a sample pair, and Γ represent the similarity between two samples in a sample pair. Ada and Γ Rms represents the predicted probabilities of the sample by the Ada classifier and the Rms classifier, respectively, with the interaction feature used as the predicted probability of the sample.

[0068] Maximize the classifier difference according to formula (13):

[0069]

[0070] In the formula, min represents taking the minimum value, L1 represents the basic training objective loss function of the prototype representation pairwise learning model, and C Ada and C Rms These represent the Ada classifier and the RMS classifier, respectively. `max` indicates taking the maximum value. diff (X T ) represents the classifier difference, l diff (X T X represents the difference loss between classifiers. T Let L2 represent the target domain sample set, and L2 represent the target loss function for classifier discrepancies, which is always greater than 0. By training the classifier module to maximize classifier discrepancies, more controversial samples distributed at the classification boundary are obtained, thereby improving the accuracy of emotion recognition.

[0071] Lock the parameters of the classifier module, train the feature discriminator module, and minimize the sample feature distribution of the source domain samples and the target domain samples according to formula (14):

[0072]

[0073] In the formula, L3 represents the target loss function, m represents the number of samples in the target domain, z represents the number of emotion categories, and |·| represents the absolute value function. This represents the i-th target domain sample. This represents the predicted probability that a sample belongs to the emotion category c in the Ada classifier. F represents the predicted probability that a sample belongs to the emotion category c in the RMS classifier, and F represents the sample feature. This represents minimizing the feature distribution. By adjusting the mapping of target domain samples in the feature discriminator module, the difference between the two classifiers is minimized, thus minimizing the feature distribution of samples in the source and target domains.

[0074] In summary, the cross-database EEG emotion recognition optimization method based on prototype representation pairwise learning of the present invention constructs a prototype representation pairwise learning model including a feature discriminator module, a prototype feature extraction module, and a classifier module. The feature discriminator module extracts sample features from source and target domain samples respectively; the prototype feature extraction module extracts prototype features; the classifier module obtains interaction features based on sample features and prototype features; the pairwise learning method calculates the similarity of sample pairs based on interaction features to obtain labels for sample pairs; the prototype representation pairwise learning model is iteratively updated based on the similarity and labels of sample pairs; and the emotion classification performance is calculated based on interaction features. Optimizing the prototype representation pairwise learning model improves robustness and generalization. Furthermore, by obtaining the similarity between samples in the feature space and designing specific model structures and training methods, the accuracy of emotion recognition is improved.

[0075] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A cross-database electroencephalogram emotion recognition optimization method based on prototype characterization of pair learning, characterized by, The method comprises the following steps: A prototype representation pairwise learning model is constructed, and the prototype representation pairwise learning model comprises a feature discriminator module, a prototype feature extraction module, and a classifier module; Sample features of source domain samples and target domain samples are extracted by using the feature discriminator module; Prototype features are extracted by using the prototype feature extraction module; Interaction features are obtained from the sample features and the prototype features by using the classifier module, the similarity of a sample pair is calculated from the interaction features by using a pairwise learning method, the label of the sample pair is obtained, the prototype representation pairwise learning model is iteratively updated according to the similarity of the sample pair and the label of the sample pair, and the emotion classification performance is calculated from the interaction features; The prototype representation pairwise learning model is optimized; The optimized prototype representation pairwise learning model comprises an optimized classifier module; the classifier module of the prototype representation pairwise learning model comprises an Ada classifier and an Rms classifier, wherein the Ada classifier is a classifier using an Adam adaptive learning rate gradient descent algorithm, and the Rms classifier is a classifier using an RMSprop gradient descent algorithm. The optimized classifier module comprises: The feature discriminator module and the classifier module are subjected to supervised learning, the prototype representation pairwise learning model is trained, and a basic training target loss function of the prototype representation pairwise learning model is calculated; wherein the basic training target loss function of the prototype representation pairwise learning model is obtained by using the following formula: wherein, represents a basic training target loss function of the prototype feature representation pair-wise learning model, represents a domain adversarial loss of the feature discriminator module, represents a pair-wise learning loss value of the source domain, and respectively represent a pair-wise learning loss value of the target domain of the Ada classifier and a pair-wise learning loss value of the target domain of the Rms classifier, , and represent hyperparameters for giving each loss value a weight.

2. The cross-database electroencephalogram emotion recognition optimization method based on the prototype characterization of pair learning according to claim 1, characterized in that, The sample features of the source domain samples and the target domain samples are extracted by using the feature discriminator module, and the sample features of the source domain samples and the target domain samples are extracted by using the feature extractor of the feature discriminator module, and the feature discriminator of the feature discriminator module is used to distinguish the domain of the sample features. The prototype features are extracted by using the prototype feature extraction module, and the prototype features are extracted from the sample features of the source domain samples by using the prototype feature extraction module.

3. The cross-database electroencephalogram emotion recognition optimization method based on the prototype characterization of pair learning according to claim 2, characterized in that, The interaction features are obtained from the sample features and the prototype features by using the classifier module, and the interaction between the sample features and the prototype features is represented by using a bilinear transformation strategy to obtain the interaction features; the interaction features are represented by using the following formula: ​ 4. The cross-database electroencephalogram emotion recognition optimization method based on prototype characterization of pair learning according to claim 1, characterized in that, ​ ​ wherein, represents an interaction feature, represents an inner product operation of a matrix, represents a sample feature, represents a prototype feature of an emotion class, represents a bilinear transformation matrix.​

Images