Emotion Estimation Device

Abstract

To provide an emotion estimation device that can improve the accuracy of emotion estimation. [Solution] The emotion estimation device 100 includes an electroencephalogram (EEG) signal input unit 111 into which time-series data of the subject's EEG signal is input, an estimation unit 112 that estimates the subject's emotion in the target time interval by inputting at least the EEG signal data in a target time interval which is a predetermined time interval, and the EEG signal data in the preceding time interval which is a predetermined time interval immediately preceding the target time interval, into a regression model, and an emotion output unit 113 that outputs the estimated emotion.

Description

[Technical Field]

[0001] This disclosure relates to an emotion estimation device. [Background technology]

[0002] Technologies for estimating human emotions are known. For example, Patent Document 1 discloses a computer system that estimates human emotions using biological data such as electroencephalograms and an emotion estimation model. Non-Patent Document 1 discloses subjective evaluation of images. [Prior art documents] [Patent Documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2023-089729 [Non-patent literature]

[0004] [Non-Patent Document 1] Benedek Kurdi, Shayn Lozano, and Mahzarin R. Banaji, “Introducing the Open Affective Standardized Image Set (OASIS)”, Behav Res (2017), 49, pp.457-470, 2017. [Overview of the project] [Problems that the invention aims to solve]

[0005] Thus, while various technologies for estimating emotions have been developed in recent years, there is a need to improve the accuracy of emotion estimation.

[0006] This disclosure is made against the backdrop of the circumstances described above, and aims to provide an emotion estimation device that can improve the accuracy of emotion estimation. [Means for solving the problem]

[0007] One aspect of the present disclosure for achieving the above objective is an emotion estimation device comprising: an electroencephalogram (EEG) signal input unit that receives time-series data of a subject's electroencephalogram (EEG) signal; an estimation unit that estimates the subject's emotions in the target time interval by inputting at least the EEG signal data in a target time interval, which is a predetermined time interval, and the EEG signal data in the preceding time interval, which is a predetermined time interval, immediately preceding the target time interval, into a regression model; and an emotion output unit that outputs the estimated emotions.

[0008] In one embodiment described above, the regression model may be a model whose weights are learned using explanatory variables that include the correlation coefficients for each frequency component of multiple brainwave signals of a person in the k-th time interval of a predetermined time length (where k is an integer of 2 or more) when the person views the image, the correlation coefficients for each frequency component of multiple brainwave signals of the person in the (k-1) time interval of the predetermined time length prior to the k-th time interval, and the true value of an emotion index corresponding to the image.

[0009] In one embodiment described above, the predetermined time length may be a time length adjusted based on the estimation accuracy of the regression model.

[0010] In the above embodiment, the regression model may be a ridge regression model.

[0011] In one embodiment described above, the regression model may be a Lasso regression model. [Effects of the Invention]

[0012] According to this disclosure, it is possible to provide an emotion estimation device that can improve the accuracy of emotion estimation. [Brief explanation of the drawing]

[0013] [Figure 1] This is a block diagram showing an example of the configuration of an emotion estimation system according to an embodiment. [Figure 2]A diagram schematically showing a regression model used for emotion estimation. [Figure 3] A graph showing the emotion estimation result by an emotion estimation device. [Figure 4] Experimental results when estimating emotion using only the current frame interval in the explanatory variables are shown. [Figure 5] A graph showing the relationship between the time length of the frame interval and the correlation between the true value and the estimated value. [Figure 6] A graph showing the emotion estimation result by an emotion estimation device.

Embodiments for Carrying Out the Invention

[0014] Hereinafter, embodiments will be described with reference to the drawings. FIG. 1 is a block diagram showing an example of the configuration of an emotion estimation system 10 according to an embodiment. As shown in FIG. 1, the emotion estimation system 10 includes a plurality of electrodes 200, an electroencephalograph 300, and an emotion estimation device 100. Details of the configuration of the emotion estimation system 10 will be described later.

[0015] In the present embodiment, as an example, the emotion estimation system 10 estimates the emotion of a subject who has seen a presented image. More specifically, the emotion estimation system 10 estimates the transition of the emotion of the subject who has seen each of the images that vary dynamically, that is, each of the images that are switched over time. For this reason, the emotion estimation system 10 can also be referred to as a dynamic decoder device.

[0016] First, the experimental environment for emotion estimation using the emotion estimation system 10 will be described. In the experiment, a plurality of electrodes 200 are attached to the head of the subject. Then, images in the rest state and images in the task state are alternately and repeatedly presented to the subject with the electrodes 200 attached. Specifically, the subject views a display on which an image in the rest state or an image in the task state is displayed. Here, the image in the task state is an image aimed at inducing an emotion in the subject, and two types of images (negative images and positive images) are prepared. Specific examples of the two types of images used as the images in the task state, that is, negative images and positive images, will be described later. The image in the rest state is an image not aimed at inducing an emotion and is an image aimed at giving the subject a rest. Specifically, an image with a cross drawn in the center of a white background was used as the image in the rest state. In the experiment, the presentation of the image in the rest state for 30 seconds and the presentation of the image in the task state for 30 seconds are repeated for the subject. Here, in the presentation of the image in the task state, one of the negative image or the positive image is randomly selected and presented to the subject each time. As an example, the probability of selecting a negative image and the probability of selecting a positive image are both set to 0.5.

[0017] In this embodiment, Russell's emotional circle diagram (Russell's emotional circle model) was used as an indicator to represent emotions. This emotional circle diagram defines the emotional space with a horizontal axis representing valence and a vertical axis representing arousal. More specifically, the first quadrant of this space is occupied by emotions such as joy and happiness, the second quadrant by emotions such as fear and anger, the third quadrant by emotions such as sadness and boredom, and the fourth quadrant by emotions such as relaxation and ease. In a study shown in Non-Patent Literature 1, which conducted subjective evaluations of images, it was confirmed that images depicting snakes (hereinafter referred to as snake images) are located in the second quadrant, and images depicting forests (green plants) (hereinafter referred to as forest images) are located in the fourth quadrant. In this experiment, the second and third quadrants of Russell's Circle of Emotions were interpreted as negative, and the first and fourth quadrants as positive. The snake image was used as the negative image, and the forest image as the positive image.

[0018] Next, the configuration of the emotion estimation system 10 will be explained in detail. Multiple electrodes 200 are placed at various positions on the subject's head. In this experiment, electrodes 200 are placed at 15 positions as defined by the International 10-20 method: Fp1, Fp2, F3, Fz, F4, T7, C3, Cz, C4, T8, P3, Pz, P4, O1, and O2. In the experiment, dry active electrodes (g.SAHARA electrodes) were used as each electrode 200. The electroencephalogram (EEG) signals detected by each electrode 200 are input to the electroencephalograph 300. The electroencephalograph 300 measures the EEG signals detected by the electrodes 200. The electroencephalograph 300 used in this experiment was a g.USBamp (manufactured by g.tec, 16 channels) and measures EEG signals at a sample frequency of 128 Hz. The electroencephalograph 300 may perform signal processing. In this experiment, the electroencephalograph 300 includes a notch filter (60Hz) and a bandpass filter (0.5-30Hz), and performs signal processing using these filters. The output data from the electroencephalograph 300 is input to the emotion estimation device 100. The data output by the electroencephalograph 300 is received by the electroencephalogram signal input unit 111 of the emotion estimation device 100, which will be described later. As described above, the subject is repeatedly presented with images of a rest state and images of a task state alternately, and the subject's electroencephalogram signals during this time are measured by the electrodes 200 and the electroencephalograph 300. The time-series data of the electroencephalogram signals for each electrode 200 is output from the electroencephalograph 300 and input to the emotion estimation device 100.

[0019] As shown in Figure 1, the emotion estimation device 100 includes a processor 110, a memory 130, and an input / output interface 150. Thus, the emotion estimation device 100 has the functionality of a computer.

[0020] The input / output interface 150 is used to communicate with other devices. In this embodiment, the data output by the electroencephalograph 300 is input to the emotion estimation device 100 via the input / output interface 150.

[0021] The memory 130 is composed of, for example, a combination of volatile memory and non-volatile memory. The memory 130 is used to store programs executed by the processor 110 and data used for various processes. In the example shown in Figure 1, the model storage unit 131 is implemented by the memory 130, but it may be implemented by any storage device other than the memory 130.

[0022] The processor 110 reads and executes a program from the memory 130. This enables the processor 110 to implement the functions of the electroencephalogram (EEG) signal input unit 111, estimation unit 112, emotion output unit 113, and model learning unit 114, which will be described later. The processor 110 may be, for example, a microprocessor, MPU (Micro Processor Unit), CPU (Central Processing Unit), or GPU (Graphics Processing Unit). The processor 110 may also include multiple processors.

[0023] The program, when loaded into a computer, includes a set of instructions (or software code) for causing the computer to perform one or more functions as described in the embodiments. The program may be stored on a non-temporary computer-readable medium or a physical storage medium. Examples, but not limited to, include random-access memory (RAM), read-only memory (ROM), flash memory, solid-state drive (SSD) or other memory technologies, CD-ROM, digital versatile disc (DVD), Blu-ray® disc or other optical disc storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage devices. The program may be transmitted over a temporary computer-readable medium or a communication medium. Examples, but not limited to, include temporary computer-readable medium or a communication medium that includes electrically, optically, acoustically or otherwise propagating signals.

[0024] The following describes the electroencephalogram signal input unit 111, estimation unit 112, emotion output unit 113, model learning unit 114, and model memory unit 131.

[0025] The electroencephalogram (EEG) signal input unit 111 receives time-series data of the subject's electroencephalogram (EEG) signals. Specifically, the EEG signal input unit 111 receives time-series data of the subject's EEG signals output by the electroencephalograph 300. The EEG signal input unit 111 also receives time-series data of the EEG signals for each electrode 200. The EEG signal input unit 111 may remove artifacts based on electrooculography (EOC) and other factors from the input data using independent component analysis.

[0026] In this embodiment, the electroencephalogram (EEG) signal input unit 111 performs a process to calculate the frequency components of the EEG from the input EEG signal data. Specifically, the EEG signal input unit 111 performs a process to calculate the frequency components of the EEG from the EEG signal data y ch egg For this, bandpass filter G bandpass Signal processing is performed using (s) to obtain the frequency component data y of the electroencephalogram signal. ch wave The EEG signal input unit 111 performs the process represented by the following equation (1). bandpass The "s" in (s) represents the Laplace operator.

[0027] <Formula (1)>

number

[0028] Bandpass filter G bandpass(s) is an 8th-order Butterworth filter, and seven types of frequency components (α, α1, α2, Θ, β, β1, β2) are obtained by the process shown in equation (1). The channels correspond to each electrode 200 attached to different positions on the subject's head, and in this embodiment, channels ch = 1, 2, ..., 15. That is, the electroencephalogram (EEG) signal input unit 111 calculates the above seven types of frequency components for each channel corresponding to each electrode 200. The EEG signal input unit 111 also calculates the correlation coefficient (cor) between channels for the same frequency component for each frame interval, which is a time window of a predetermined time length. A frame interval is also called a time interval. In this way, the EEG signal input unit 111 extracts a predetermined number of frequency components from the time-series data of the EEG signal for each channel, and calculates the correlation coefficient for the same frequency component for all combinations of channels consisting of two channels for each frame interval. In other words, the calculated correlation coefficient shows the correlation between channels for the same frequency component in the same frame interval.

[0029] The estimation unit 112 estimates the subject's emotion using the electroencephalogram (EEG) signal data received by the EEG signal input unit 111. In particular, the estimation unit 112 estimates the subject's emotion in the target frame interval by inputting at least the EEG signal data in the target frame interval (target time interval) and the EEG signal data in the immediately preceding frame interval (immediately preceding time interval) into a regression model. Here, the target frame interval is the interval for which emotion estimation is to be performed, and the immediately preceding frame interval is the frame interval immediately preceding the target frame interval. The time lengths of both the target frame interval and the immediately preceding frame interval are predetermined time lengths as described above. Specifically, in this embodiment, the estimation unit 112 estimates the subject's emotion in the target frame interval by inputting the correlation coefficient in the target frame interval calculated by the EEG signal input unit 111 and the correlation coefficient in the immediately preceding frame interval calculated by the EEG signal input unit 111 into a regression model. In this way, the estimation unit 112 considers the current k-th frame interval and the k-1-th past frame interval in order to estimate the current emotion.

[0030] In this embodiment, the estimation unit 112 performs estimation processing using a regression model represented by the following equation (2).

[0031] <Formula (2)>

number

[0032] In equation (2), v represents emotion, i.e., the dependent variable, and is composed of the coordinate values ​​of the horizontal axis (x-axis) and vertical axis (y-axis) of Russell's circle of emotion. A is the feature input to the regression model, i.e., the explanatory variable. That is, A is the electroencephalogram signal data (i.e., measurement data) input to the regression model, and in this embodiment, it is specifically composed of the correlation coefficient of the k-th frame interval and the correlation coefficient of the (k-1)-th frame interval. Also, u represents the weight for the feature.

[0033] Figure 2 schematically shows the regression model used for emotion estimation in this embodiment. In Figure 2, the lower section shows example graphs of the frequency component α of channel 1 and the frequency component α of channel 2, illustrating how the correlation coefficient is calculated from these graphs. As mentioned above, the correlation coefficient is calculated for each frequency component for all pairs of channels.

[0034] As shown in Figure 2, in this embodiment, explanatory variable A includes a group of elements corresponding to the correlation coefficient for the k-th frame interval and a group of elements corresponding to the correlation coefficient for the (k-1)-th frame interval. More specifically, in the example shown in Figure 2, explanatory variable A is represented as a matrix, and the number of dimensions in the row direction is {number of frequency components (7)} × {number of electrode combinations (15 × 14 / 2)} × {number of frame intervals used as recurrence, i.e., features (2)} + {stationary offset (1)} (= 1471). In the example in Figure 2, explanatory variable A has a stationary offset value of 1 in the rightmost column of the matrix. Also, in the example in Figure 2, explanatory variable A has a group of elements corresponding to the correlation coefficient for the k-th frame interval in the left half of the submatrix excluding the stationary offset column, and a group of elements corresponding to the correlation coefficient for the (k-1)-th frame interval in the right half. The number of dimensions in the column direction is the total measurement time divided by the frame interval width. In other words, the number of dimensions in the column direction is the number obtained by dividing the time length of the time-series data of the electroencephalogram signal by the predetermined time length mentioned above, and corresponds to the total number of frame intervals. In the example shown in Figure 2, the number of dimensions in the column direction is specifically 47. Thus, in explanatory variable A, the column direction represents the passage of time. More specifically, the first row of explanatory variable A lists the correlation coefficients of the k-th frame interval (second frame interval) and the (k-1)-th frame interval (first frame interval) for k=2. Similarly, the second row of explanatory variable A lists the correlation coefficients of the k-th frame interval (third frame interval) and the (k-1)-th frame interval (second frame interval) for k=3. In other words, the correlation coefficient of the second frame interval exists as the correlation coefficient of the k-th frame interval in the first row of explanatory variable A, and also as the correlation coefficient of the (k-1)-th frame interval in the second row of explanatory variable A. Thus, the i-th row of explanatory variable A (where i is an integer greater than or equal to 1 and less than or equal to the maximum number of rows) contains the correlation coefficients between the k-th frame interval and the (k-1)-th frame interval for k=i+1.

[0035] Furthermore, as shown in Figure 2, in this embodiment, the dependent variable v includes the x-coordinate and y-coordinate values ​​of the Russell affect circle diagram. More specifically, in the example shown in Figure 2, the dependent variable v is represented as a matrix, with two dimensions in the row direction (x-coordinate and y-coordinate). The number of dimensions in the column direction is the same as the number of dimensions in the column direction of the explanatory variable A. Thus, the dependent variable v represents the emotion for each frame interval. That is, in the example shown in Figure 2, the matrix of the dependent variable v represents the emotion in each of the 47 frame intervals. In other words, the matrix of the dependent variable v represents the change in emotion according to the transition of frame intervals.

[0036] The weights u are also represented by a matrix, with 2 elements in the row direction, the same as the number of elements in the row direction of the dependent variable v. The number of elements in the column direction is 1471, the same as the number of elements in the row direction of the independent variable A.

[0037] The estimation unit 112 estimates the subject's emotion using a regression model in which weights u have been learned. Therefore, the model learning unit 114 performs machine learning processing to learn the values ​​of weights u. To learn the values ​​of weights u, a training signal (training data) is required. In this embodiment, the training signals used are pairs of forest images and their corresponding coordinate values, and pairs of snake images and their corresponding coordinate values. Therefore, in this embodiment, the model learning unit 114 learns the values ​​of weights u using sets of brainwave data from people who viewed the forest image and the true coordinate values ​​of the emotions corresponding to the forest image, and sets of brainwave data from people who viewed the snake image and the true coordinate values ​​of the emotions corresponding to the snake image. The relationship between image stimuli and subjective emotional evaluations was set according to the results shown in a public database (OASIS: see Non-Patent Literature 1). That is, in this embodiment, the coordinate values ​​set based on the public database were set as the training signal values, i.e., the values ​​of the target variable v in the learning phase.

[0038] The model learning unit 114 calculates the weights u using the least squares method with respect to the teacher signal v and the explanatory variable A calculated by the electroencephalogram signal input unit 111. Therefore, the regression model used by the estimation unit 112 in this embodiment can be described as follows: The regression model used by the estimation unit 112 for estimation is a model in which weights are learned using explanatory variables that include the correlation coefficients for each frequency component of multiple electroencephalogram signals of a person in the k-th frame interval (where k is an integer of 2 or more) and the correlation coefficients for each frequency component of multiple electroencephalogram signals of the person in the (k-1)-th frame interval, and the true value of the emotion index value (coordinates in Russell's circle of emotion) corresponding to the image. More specifically, k is an integer between 2 and N, and N is an arbitrary predetermined value corresponding to the total number of frame intervals used for learning.

[0039] In this embodiment, the model learning unit 114 specifically calculates the weights u by ridge regression, as shown in equation (3) below. That is, in this embodiment, the regression model used by the estimation unit 112 is a ridge regression model.

[0040] <Formula (3)>

number

[0041] In equation (3), the u^ on the left side ridge This represents the weights calculated by ridge regression. In equation (3), the term with the hyperparameter λ is the L2 regularization term. Ridge regression can adjust for overfitting by introducing the L2 regularization term.

[0042] Thus, the regression model is trained by the model training unit 114, and the trained regression model is stored in the model storage unit 131. The estimation unit 112 estimates emotion by inputting data from the electroencephalogram signal input unit 111 into the regression model stored in the model storage unit 131, whose weights have been trained in advance. In other words, the estimation unit 112 obtains the coordinates output from this regression model.

[0043] Note that the regression model whose weights are learned by the model learning unit 114, that is, the model used by the estimation unit 112 in the inference phase, may be expressed as follows to clearly distinguish it from the model before learning.

[0044] <Equation (4)>

Number

[0045] In Equation (4), û explicitly indicates that the learned weights are set. Also, v p explicitly indicates that it is the emotion estimated by the estimation unit 112 using the learned regression model.

[0046] The emotion output unit 113 performs a process of outputting the emotion estimated by the estimation unit 112. More specifically, the emotion output unit 113 outputs the emotion of the subject in each frame interval. That is, the emotion output unit 113 outputs the emotion of the subject (coordinates in Russell's affect circumplex) in each frame interval, represented by the value of the target variable v. The emotion output unit 113 may display the estimated emotion on a display or transmit it to another device as the output of the estimated emotion.

[0047] Next, we will describe the results of the emotion estimation experiment using the emotion estimation device 100 described above. The experimental results described below are for when the time length of the frame interval (i.e., the predetermined time length described above) is 4.5 seconds. In other words, the time difference between two consecutive frame intervals is 4.5 seconds. The value of the hyperparameter λ used is 0.02. Figure 3 is a graph showing the emotion estimation results by the emotion estimation device 100. Specifically, Figure 3 shows the experimental results when emotion is estimated using not only the current frame interval but also past frame intervals as explanatory variables. In contrast, Figure 4 shows the experimental results when emotion is estimated using only the current frame interval as explanatory variables. In Figures 3 and 4, the upper graph shows the emotion estimation results for electroencephalogram (EEG) signal data used for model training (hereinafter also referred to as training data), and the lower graph shows the emotion estimation results for EEG signal data not used for model training (hereinafter also referred to as test data). In each figure, graphs of the true values ​​of the x and y coordinates, and graphs of the estimated values ​​of the x and y coordinates are shown. The "cor" value in the upper right of the figure indicates the correlation between the true and estimated values ​​for the x and y coordinates, respectively. For example, 0.6739, shown in the upper right of the lower part of Figure 3, represents the correlation coefficient between the true and estimated values ​​for the x coordinate, and 0.7204, shown in the upper right of the lower part of Figure 3, represents the correlation coefficient between the true and estimated values ​​for the y coordinate. As is clear from Figures 3 and 4, estimating emotions using past frame intervals, as in this embodiment, results in a higher correlation coefficient between the true and estimated values ​​and higher accuracy in estimating emotions compared to estimating emotions using only the current frame interval for the explanatory variable.

[0048] The embodiments have been described above. In this embodiment, emotion estimation using a regression model is performed using not only the electroencephalogram (EEG) signal data of the target frame interval, but also the EEG signal data of the immediately preceding frame interval. This makes it possible to accurately estimate dynamically changing emotions.

[0049] <Example 1> The inventors investigated the relationship between the time length of the frame interval (the predetermined time length mentioned above) and the accuracy of emotion estimation. In other words, the inventors investigated the estimation accuracy when the time difference between two consecutive frame intervals was varied. Figure 5 is a graph showing the relationship between the time length of the frame interval (horizontal axis) and the correlation between the true value and the estimated value (vertical axis). Figure 5 shows the correlation between the true value and the estimated value for emotion estimation for the test data. In Figure 5, the left side is the graph for the x-coordinate, and the right side is the graph for the y-coordinate. As shown in Figure 5, it can be confirmed that there are correlation peaks for both the x-coordinate and the y-coordinate when the time length of the frame interval is 4.5 seconds and 8 seconds. Thus, the inventors discovered that the time length of the frame interval (i.e., the time difference between two consecutive frame intervals) affects the estimation accuracy.

[0050] Therefore, it is desirable that the time length of the frame interval (the predetermined time length mentioned above) is a time length adjusted based on the estimation accuracy of the regression model. In other words, it is preferable for the estimation unit 112 to estimate emotions using a regression model in which the time length of the frame interval has been adjusted based on the estimation accuracy. To put it another way, the estimation unit 112 may estimate emotions by applying a frame interval of a time length selected from among a plurality of time lengths based on the estimation accuracy to the regression model. Specifically, for example, the estimation unit 112 may estimate emotions using one regression model selected from among a plurality of different regression models in which weights have been learned using different predetermined time lengths, based on the estimation accuracy of each regression model. Specifically, for example, the estimation unit 112 may estimate emotions using the regression model with the highest estimation accuracy among a plurality of different regression models. In this way, by using frame intervals with time lengths adjusted based on the estimation accuracy of the regression model, highly accurate estimation can be expected.

[0051] <Modification 2> In the embodiments and modifications described above, the model learning unit 114 calculated the weights using ridge regression. The weights u obtained by ridge regression have a full set of numerical values ​​in the matrix and are not zeroed out. Therefore, in order to examine the sparsification of the weights u, the inventors considered the estimation of emotion using weights calculated by Lasso regression. In this modification, the model learning unit 114 specifically calculates the weights u using Lasso regression as shown in equation (5) below. That is, in this modification, the regression model used by the estimation unit 112 is a Lasso regression model.

[0052] <Formula (5)>

number

[0053] In equation (5), the u^ on the left side lasso This represents the weights calculated by Lasso regression. In equation (5), the term with the hyperparameter λ is the L1 regularization term. Lasso regression has the characteristic that, due to the introduction of the L1 regularization term, the weights u are sparsified, and as a result, the weights for important data are naturally retained. In this modified example, this characteristic is used to attempt to reduce the dimensionality of the weights.

[0054] This report describes the results of an emotion estimation experiment using the emotion estimation device 100 with a regression model whose weights were learned using Lasso regression. The frame interval length was 4.5 seconds. The value of the hyperparameter λ used was 0.02. Figure 6 is a graph showing the emotion estimation results of the emotion estimation device 100 with a regression model whose weights were learned using Lasso regression. Similar to Figure 3, in Figure 6, the upper part is a graph showing the emotion estimation results for the training data, and the lower part is a graph showing the emotion estimation results for the test data. Similar to Figure 3, each figure in Figure 6 shows graphs of the true values ​​of the x and y coordinates, and graphs of the estimated values ​​of the x and y coordinates. The cor value in the upper right of the figure shows the correlation between the true value and the estimated value for the x and y coordinates, respectively. The estimation accuracy of the Lasso regression model is slightly inferior to that of the ridge regression model. However, when weights are learned using Lasso regression, the number of non-zero elements in the weight u is extremely small. Specifically, there were 9 non-zero elements among the weights used to estimate the x-coordinate, and 14 non-zero elements among the weights used to estimate the y-coordinate. In the ridge regression model, the weight u consisted of 1421 × 2 non-zero elements. Therefore, the Lasso regression model allows for a significant reduction in dimensionality while minimizing the degradation of estimation accuracy. Specifically, the accuracy degradation was 7.15% for the x-coordinate and 9.97% for the y-coordinate. In contrast, the dimensionality reduction was 99.4% for the x-coordinate and 99.0% for the y-coordinate.

[0055] The above describes the second modification. As mentioned above, the number of dimensions of the weights can be significantly reduced by learning the weights of the regression model using Lasso regression. This contributes to reducing the computational load of the emotion estimation device 100. Furthermore, the values ​​of the elements of explanatory variable A, which are calculated with elements whose weights are zero, do not need to be based on actual electroencephalogram (EEG) signal data. In other words, dummy values ​​may be used as the values ​​of the elements of explanatory variable A. Therefore, the electrodes 200 required to calculate the values ​​(correlation coefficients) of the elements of explanatory variable A can be omitted. In other words, the emotion estimation system 10 can estimate the emotions of a subject using only the electrodes 200 that contribute to calculating the values ​​(correlation coefficients) of the elements of explanatory variable A that are calculated with non-zero weight values. It should be noted that there are also advantages to learning the weights of the regression model using Ridge regression, as in the embodiment described above. That is, in this case, emotions can be estimated with higher accuracy compared to when the weights of the regression model are learned using Lasso regression.

[0056] It should be noted that the present invention is not limited to the embodiments described above, and can be modified as appropriate without departing from the spirit of the invention. For example, in the embodiments described above and their modifications, emotion was estimated by a regression model using data in the target frame interval and data in the frame interval immediately preceding the target frame interval. However, emotion may be estimated by a regression model using data in the target frame interval and data in two or more frame intervals prior to the target frame interval. [Explanation of Symbols]

[0057] 10. Emotion Estimation System 100 Emotion Estimation Device 110 processors 111 EEG signal input section 112 Estimation Department 113 Emotional Output Unit 114 Model Learning Section 130 memory 131 Model Memory Unit 150 Input / Output Interfaces 200 electrodes 300 electroencephalographs

Claims

1. An electroencephalogram (EEG) signal input unit receives time-series data of the subject's EEG signals, An estimation unit that estimates the subject's emotions during the target time interval by inputting at least the electroencephalogram (EEG) signal data for the target time interval, which is a predetermined time interval, and the EEG signal data for the preceding time interval, which is a predetermined time interval, immediately preceding the target time interval, into a regression model; An emotion output unit that outputs the estimated emotion and An emotion estimation device having the following features.

2. The regression model is a model whose weights have been learned using explanatory variables that include the correlation coefficients for each frequency component of multiple brainwave signals of a person in the k-th time interval of a predetermined time length (where k is an integer of 2 or more) when the person viewed the image, and the correlation coefficients for each frequency component of multiple brainwave signals of the person in the (k-1) time interval of the predetermined time length prior to the k-th time interval, and the true value of the emotion index corresponding to the image. The emotion estimation device according to claim 1.

3. The predetermined time length is a time length adjusted based on the estimation accuracy of the regression model. The emotion estimation device according to claim 2.

4. The regression model is a ridge regression model. The emotion estimation device according to any one of claims 1 to 3.

5. The regression model is a Lasso regression model. The emotion estimation device according to any one of claims 1 to 3.

Images