Method and device for determining feature space

Abstract

The present disclosure provides a method and device for determining a feature space, and belongs to the field of computer technology. The method includes: obtaining features of multiple training samples in different dimensions, wherein the multiple training samples include multiple groups of training samples of different categories; determining the inter-class distance and intra-class distance corresponding to the features of each dimension according to the features of the multiple training samples in different dimensions, and dividing the features of different dimensions into a first feature set and a second feature set according to the inter-class distance and intra-class distance corresponding to the features of each dimension, wherein the features in the first feature set contribute more to determining the category than the features in the second feature set, and determining the target feature space after the features of different dimensions are reduced in dimension according to the first feature set and the second feature set. By adopting the present disclosure, an accurate feature space can be obtained.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and in particular to a method and device for determining a feature space. Background Art

[0002] With the development of computer technology, more and more classification and recognition tasks will use multi-dimensional features. For example, in speech emotion recognition tasks, fundamental frequency features, formant features, Mel-frequency cepstral coefficient features, zero-crossing rate features, etc. are used. When the feature dimension used is too large, some redundant features may be introduced, which will reduce the classification and recognition performance and increase the complexity of calculation. Therefore, it is necessary to provide a method to reduce the feature dimension, improve the classification and recognition performance, and reduce the complexity of calculation by reducing the feature dimension.

[0003] In the related art, when reducing the dimension of features, r features are extracted from each training sample, and the variance of each feature is determined. The variances are sorted from large to small, and the first n features with the largest variances are selected as important features, and the remaining rn features are selected as unimportant features. The feature space after dimensionality reduction is determined by retaining n important features with large variance changes and removing rn unimportant features with small variance changes. The feature space obtained by the related art is not accurate enough, which in turn affects the accuracy of the classification and recognition results. Summary of the invention

[0004] In order to solve the problems of the related art, the embodiments of the present disclosure provide a method and device for determining a feature space. The technical solution is as follows:

[0005] In a first aspect, the present disclosure provides a method for determining a feature space, the method comprising:

[0006] Acquire features of multiple training samples in different dimensions, wherein the multiple training samples include multiple groups of training samples of different categories;

[0007] Determine the inter-class distance and intra-class distance corresponding to each dimension of the feature according to the features of the multiple training samples in different dimensions;

[0008] According to the inter-class distance and intra-class distance corresponding to each dimension of the feature, the features of the different dimensions are divided into a first feature set and a second feature set, wherein the contribution of the features in the first feature set to determining the category is greater than that of the features in the second feature set;

[0009] A target feature space after dimensionality reduction of the features of different dimensions is determined according to the first feature set and the second feature set.

[0010] In a possible implementation, determining the inter-class distance and intra-class distance corresponding to each dimension of the feature according to the features of the multiple training samples in different dimensions includes:

[0011] Determine the inter-class distance corresponding to each dimensional feature according to the number of training samples of each category in the multiple training samples, the number of categories of the multiple training samples, the mean of each dimensional feature in the training samples of each category, and the mean of each dimensional feature in the multiple training samples;

[0012] The intra-class distance corresponding to each dimensional feature is determined according to the number of categories of the multiple training samples, the mean value of each dimensional feature in each category of training samples, and the feature value of each training sample in each category of training samples.

[0013] In this way, the inter-class distance and intra-class distance can be accurately determined.

[0014] In a possible implementation, the features of different dimensions are divided into a first feature set and a second feature set according to the inter-class distance and the intra-class distance corresponding to each dimension of the feature, including:

[0015] Determine the ratio of the inter-class distance to the intra-class distance corresponding to each dimension of the feature, and obtain the ratio corresponding to each dimension of the feature;

[0016] According to the ratio corresponding to each dimensional feature, the features of different dimensions are divided into a first feature set and a second feature set.

[0017] In a possible implementation, the features of different dimensions are divided into a first feature set and a second feature set according to the ratio corresponding to each dimensional feature, including:

[0018] According to the order of the ratios corresponding to the features of each dimension from large to small, m features with the largest ratios are selected to form a first type of feature set, and the remaining features of the features of different dimensions except the features in the first type of feature set are used to form a second type of feature set, where m is greater than or equal to 1 and less than or equal to the total dimension of the features.

[0019] In this way, important features and unimportant features can be accurately distinguished.

[0020] In a possible implementation, determining, according to the first feature set and the second feature set, a target feature space after reducing the dimensions of the features of different dimensions includes:

[0021] according to Based on the principle of taking the maximum value, the target feature space after the features of different dimensions are reduced in dimension is determined;

[0022] Wherein, W represents the target feature space, is the covariance matrix of the features in the first type of feature set in the first dimensional space, is the covariance matrix of the features in the second feature set in the first dimensional space, express traces, express The dimension of the first dimensional space is higher than the dimension of the target feature space.

[0023] In this way, the problem of linearly inseparable features can be solved.

[0024] In a second aspect, the present disclosure provides a device for determining a feature space, the device comprising:

[0025] An acquisition module, used to acquire features of a plurality of training samples in different dimensions, wherein the plurality of training samples include a plurality of groups of training samples of different categories;

[0026] Identify modules for:

[0027] Determine the inter-class distance and intra-class distance corresponding to each dimension of the feature according to the features of the multiple training samples in different dimensions;

[0028] According to the inter-class distance and intra-class distance corresponding to each dimension of the feature, the features of the different dimensions are divided into a first feature set and a second feature set, wherein the contribution of the features in the first feature set to determining the category is greater than that of the features in the second feature set;

[0029] A target feature space after dimensionality reduction of the features of different dimensions is determined according to the first feature set and the second feature set.

[0030] In a possible implementation manner, the determining module is used to:

[0031] Determine the inter-class distance corresponding to each dimensional feature according to the number of training samples of each category in the multiple training samples, the number of categories of the multiple training samples, the mean of each dimensional feature in the training samples of each category, and the mean of each dimensional feature in the multiple training samples;

[0032] The intra-class distance corresponding to each dimensional feature is determined according to the number of categories of the multiple training samples, the mean value of each dimensional feature in each category of training samples, and the feature value of each training sample in each category of training samples.

[0033] In a possible implementation manner, the determining module is used to:

[0034] Determine the ratio of the inter-class distance to the intra-class distance corresponding to each dimension of the feature, and obtain the ratio corresponding to each dimension of the feature;

[0035] According to the ratio corresponding to each dimensional feature, the features of different dimensions are divided into a first feature set and a second feature set.

[0036] In a possible implementation manner, the determining module is used to:

[0037] According to the order of the ratios corresponding to the features of each dimension from large to small, m features with the largest ratios are selected to form a first type of feature set, and the remaining features of the features of different dimensions except the features in the first type of feature set are used to form a second type of feature set, where m is greater than or equal to 1 and less than or equal to the total dimension of the features.

[0038] In a possible implementation manner, the determining module is used to:

[0039] according to Based on the principle of taking the maximum value, the target feature space after the features of different dimensions are reduced in dimension is determined;

[0040] Wherein, w represents the target feature space, is the covariance matrix of the features in the first type of feature set in the first dimensional space, is the covariance matrix of the features in the second feature set in the first dimensional space, express traces, express The dimension of the first dimensional space is higher than the dimension of the target feature space.

[0041] The beneficial effects brought by the technical solution provided by the embodiments of the present disclosure include at least:

[0042] In the disclosed embodiment, the recognition device can obtain features of multiple training samples in different dimensions. The recognition device determines the inter-class distance and intra-class distance corresponding to each dimension of the features based on the features of the multiple training samples in different dimensions, and uses the inter-class distance and intra-class distance corresponding to each dimension of the features to divide the features of different dimensions into a first feature set and a second feature set, wherein the contribution of the features in the first feature set to determining the category is higher than that of the features in the second feature set, and the recognition device determines the feature space after the dimensionality is reduced based on the first feature set and the second feature set. In this way, the inter-class distance and intra-class distance corresponding to each dimension of the features can be determined, and the first feature set and the second feature set can be determined using the inter-class distance and intra-class distance corresponding to each dimension of the features, wherein the first feature set is an important feature and the second feature set is an unimportant feature, and then the first feature set and the second feature set are used to determine the target feature space after the dimensionality of the features of different dimensions is reduced. Since not only the variance of the features in each category of training samples is considered when determining important features and unimportant features, but also the distinguishability of the features between different categories of training samples (i.e., inter-class distance) is considered, the important features determined can be more accurate, and thus the features determined after dimensionality reduction can be more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flow chart of a method for determining a feature space provided by an embodiment of the present disclosure;

[0044] Figure 2 It is a schematic diagram of the principle of KPCA (kernel principal component analysis) provided by an embodiment of the present disclosure;

[0045] Figure 3 is a structural schematic diagram of a device for determining a feature space provided by an embodiment of the present disclosure;

[0046] Figure 4 It is a structural diagram of an identification device provided by an embodiment of the present disclosure. DETAILED DESCRIPTION

[0047] In order to make the objectives, technical solutions and advantages of the present disclosure more clear, the embodiments of the present disclosure will be further described in detail below with reference to the accompanying drawings.

[0048] In the embodiments of the present disclosure, the method for determining the feature space can be applied to a variety of classification and recognition tasks. For example, speech emotion recognition tasks, image classification and recognition tasks, etc. In the embodiments of the present disclosure, taking the speech emotion recognition task as an example, various features are used, and the features may include fundamental frequency, resonance peak, Mel-frequency cepstral coefficient, zero-crossing rate, amplitude, and statistical characteristics of the features on the sentence (for example, mean, variance, maximum value, minimum value, etc.). In this way, the dimensions of the features used for the speech emotion recognition task are relatively large, which will lead to the introduction of some redundant features, reduce the accuracy of speech emotion recognition, and increase the complexity of calculation, so it is necessary to provide a method for determining the feature space.

[0049] The embodiment of the present disclosure provides a method for determining a feature space, and the execution subject of the method can be an identification device, and the identification device can be a terminal or a server. The identification device can be provided with a processor, a memory, a transceiver, etc. The processor can be used for processing in the process of determining the feature space, such as determining the inter-class distance and intra-class distance of each dimension of the feature, etc. The memory can be used to store the data required and the data generated in the process of determining the feature space, such as storing the features of different dimensions of multiple training samples, etc. The transceiver can be used to receive and send data, such as receiving the features of multiple training samples in different dimensions or receiving multiple training samples, etc.

[0050] like Figure 1 As shown, a flowchart of the method for determining the feature space is given:

[0051] Step 101: A recognition device obtains features of multiple training samples in different dimensions, wherein the multiple training samples include multiple groups of training samples of different categories.

[0052] In this embodiment, the user can input multiple training samples into the recognition device through the interface provided by the recognition device. The multiple training samples may include multiple groups of training samples of different categories. For example, the multiple training samples are samples calibrated with labels. In the speech emotion classification and recognition task, the multiple training samples can be divided into training samples of the anger category, training samples of the happiness category, training samples of the sadness category, etc. In the image classification and recognition task, the multiple training samples can be divided into training samples of the cat category, training samples of the dog category, training samples of the cow category, etc. After the recognition device obtains the multiple training samples, it can use the stored feature extraction model to extract features of the multiple training samples in different dimensions, and the dimensions of the features extracted from each training sample are the same.

[0053] Alternatively, the recognition device may send a request for acquiring features to other devices, and after receiving the request, the other devices may provide the recognition device with features of multiple training samples in different dimensions.

[0054] Step 102: The recognition device determines the inter-class distance and intra-class distance corresponding to each dimension of the feature according to the features of the multiple training samples in different dimensions.

[0055] In this embodiment, the recognition device can use the features of multiple training samples in different dimensions to determine the inter-class distance and intra-class distance corresponding to each dimension feature. For a certain dimension feature, the inter-class distance corresponding to the dimension feature is used to measure the distance between the dimension feature in different categories of training samples. The intra-class distance corresponding to the dimension feature is used to measure the mean square distance of the dimension feature in the same category of training samples.

[0056] In a possible implementation, the processing of step 102 may be:

[0057] According to the number of training samples in each category among multiple training samples, the number of categories of multiple training samples, the mean of each dimensional feature in the training samples in each category and the mean of each dimensional feature in multiple training samples, the inter-class distance corresponding to each dimensional feature is determined; according to the number of categories of multiple training samples, the mean of each dimensional feature in the training samples in each category and the feature value of the feature of each training sample in the training samples in each category, the intra-class distance corresponding to each dimensional feature is determined.

[0058] In this embodiment, the recognition device can use the category label of each training sample to divide multiple training samples into multiple categories of training samples, determine the number of training samples in each category, and the number of categories of multiple training samples. Then for any category of training samples, the recognition device can add the i-th dimension features of each training sample in the category and divide it by the number of training samples in the category to obtain the mean of the i-th dimension features in the training samples of the category. And the recognition device can add the i-th dimension features of each training sample in the multiple training samples and divide it by the number of training samples to obtain the mean of the i-th dimension features in the multiple training samples (the mean can be called the overall mean). Then the recognition device determines the inter-class distance corresponding to each dimensional feature based on these calculation results. And the recognition device determines the intra-class distance corresponding to each dimensional feature according to the number of categories of multiple training samples, the mean of each dimensional feature in each category of training samples, and the feature value of the feature of each training sample in each category of training samples.

[0059] Optionally, the processing method for determining the inter-class distance and the intra-class distance may be:

[0060] For the i-th dimension feature in the feature, the inter-class distance corresponding to the i-th dimension feature is expressed by formula (1):

[0061]

[0062] The intra-class distance corresponding to the i-th dimension feature is expressed by formula (2):

[0063]

[0064] In formula (1), b i represents the inter-class distance corresponding to the i-th dimension feature, c is the number of categories of training samples, n t represents the number of training samples of the tth class, u t,i represents the mean of the i-th dimension feature in the t-th class training sample, Represents the mean of the i-th dimension feature in multiple training samples (i.e. the overall mean mentioned above), and each item is multiplied by n t It can represent the contribution of the t-th class training sample to the inter-class distance. In formula (2), w i represents the intra-class distance corresponding to the i-th dimension feature, x r represents the rth training sample, x r ∈t represents x r Belongs to the tth class of training samples, Represents the eigenvalue of the i-th dimension feature of the r-th training sample in the t-th class training sample.

[0065] Optionally, the inter-class distance corresponding to the i-th dimension feature can also be expressed by formula (3):

[0066]

[0067] In formula (3), b i ,c,n t 、u t,i , Same as the explanation in formula (1), N represents the total number of training samples. It is also used to measure the contribution of the t-th class training sample to the inter-class distance.

[0068] Step 103: The recognition device divides the features of different dimensions into a first feature set and a second feature set according to the inter-class distance and intra-class distance corresponding to each dimension of the feature.

[0069] Among them, the features in the first feature set contribute more to the determination of the category than the features in the second feature set. That is, for the classification task, the features in the first feature set are important features, and the features in the second feature set are unimportant features.

[0070] In this embodiment, the recognition device can divide the features of different dimensions of multiple training samples into a first feature set and a second feature set according to the inter-class distance and intra-class distance corresponding to each dimension of the feature. In this way, since good features have the characteristics of separating samples of different categories as much as possible and bringing samples of the same category closer together, the features can be divided into a first feature set and a second feature set according to the inter-class distance and intra-class distance of each dimension of the feature.

[0071] Optionally, the processing of step 103 may be:

[0072] The recognition device determines the ratio of the inter-class distance and the intra-class distance corresponding to each dimension of the feature, and obtains the ratio corresponding to each dimension of the feature; according to the ratio corresponding to each dimension of the feature, the features of different dimensions are divided into a first feature set and a second feature set.

[0073] In this embodiment, the recognition device can calculate the ratio of the inter-class distance and the intra-class distance corresponding to the i-th dimension feature, and the ratio is calculated using σ i Indicates that:

[0074]

[0075] In formula (4), b i is the inter-class distance of the i-th dimension feature, w i is the intra-class distance of the i-th dimension feature, and d is the total dimension of the features extracted from each training sample. If each training sample includes d-dimensional features, the ratios corresponding to the d-dimensional features are calculated to be σ1,…,σ d .

[0076] The recognition device can then use the ratio corresponding to each dimension of the feature to divide the features of different dimensions into a first feature set and a second feature set.

[0077] Optionally, the recognition device may divide the features of different dimensions into a first feature set and a second feature set in the following manner:

[0078] The recognition device selects m features with the largest ratios in descending order according to the ratios corresponding to each dimensional feature, forming a first feature set, and combines the remaining features of features of different dimensions except those in the first feature set into a second feature set.

[0079] Among them, m is greater than or equal to 1, and less than or equal to the total dimension of the feature. That is, m is greater than or equal to 1, and less than or equal to d. m = d*a, a is a preset value, indicating the proportion of the first type of feature dimension to the total feature dimension, a is greater than 0, and less than or equal to 1.

[0080] In this embodiment, the recognition device obtains a pre-stored, which is preset by a technician, and multiplies a by d to obtain m. The recognition device can sort the ratios corresponding to the features of each dimension in descending order, select the m features with the largest ratios, and form the m features into a first-category feature set, and can form the features other than the m features among the d features into a second-category feature set.

[0081] The first type of feature set mentioned above can be obtained using X H Identification, the second type of feature set can use X L express.

[0082] It should be noted here that the ratio of the above-calculated inter-class distance to the intra-class distance, the larger the ratio, the better the discrimination of training samples of different categories, the more important the feature, on the contrary, the smaller the ratio, the worse the discrimination of training samples of different categories, the less important the feature. This is because in the classification and recognition task, an important feature should have the characteristics of making the same type of training samples as close as possible (that is, the intra-class distance is as small as possible), and the different types of training samples as classified as much as possible (that is, the inter-class distance is as large as possible).

[0083] Step 104: The recognition device determines a target feature space after reducing the dimensions of features of different dimensions according to the first feature set and the second feature set.

[0084] Among them, the feature space after the dimensions of features of different dimensions are reduced (i.e., the target feature space) is represented by a matrix.

[0085] In this embodiment, the recognition device may use the first type of feature set and the second type of feature set to determine the target feature space under the principle of maximizing the variance of the first type of feature set and minimizing the variance of the second type of feature set.

[0086] Optionally, the identification device can be The principle of taking the maximum value is used to determine the feature space after reducing the dimension.

[0087] In this embodiment, the recognition device can calculate formula (5): and maximize the calculated result.

[0088] When the result of formula (5) takes the maximum value, the variance of the first type of feature set after dimensionality reduction is maximized, and the variance of the second type of feature set after dimensionality reduction is minimized, and the target feature space W after dimensionality reduction is determined. In formula (5), is the covariance matrix of the features in the first-dimensional space of the first-category feature set, Φ(·) is a nonlinear mapping function to map the original features of the training sample to the first dimensional space, that is, the features of the first type of feature set of the training sample are represented in the first dimensional space as Φ(X H ),but is equivalent to W T Φ(X H )Φ(X H ) T W, if W T Φ(X H ) is represented by Y, Y represents Φ(X H ) The features obtained in the feature space W, W T Φ(X H )Φ(XH ) T W is equivalent to YY T ,YY T Represents X H The covariance matrix obtained in the feature space W after dimensionality reduction. The "T" in the upper right corner here represents transposition.

[0089] Similarly, is the covariance matrix of the features in the second feature set in the first dimensional space, Φ(·) is a nonlinear mapping function to map the original features of the training sample to the first dimensional space, that is, the features of the second type of feature set of the training sample are represented in the first dimensional space as Φ(X L ), It is also equivalent to W T Φ(X L )Φ(X L ) T W, Represents X L The covariance matrix obtained in the feature space W after dimensionality reduction. The "T" in the upper right corner here represents transposition.

[0090] express The trace of The sum of the diagonal elements of the covariance matrix. Since the diagonal elements of the covariance matrix represent the variance of the corresponding dimensional features, Represents the sum of the variances of each feature in the first feature set in the target feature space W. express The trace of The sum of the diagonal elements of the covariance matrix. Since the diagonal elements of the covariance matrix represent the variance of the corresponding dimensional features, represents the sum of the variances of the features in the second feature set in the target feature space W. Therefore, in order to make Maximization can make the ratio of the sum of the variances of the features after dimensionality reduction in the first type of feature set to the sum of the variances of the features after dimensionality reduction in the second type of feature set as large as possible.

[0091] In addition, if Figure 2As shown, the principle of KPCA to reduce feature dimensions is: map the features of the training samples (i.e., the features in the original feature space) to a high-dimensional space (i.e., the first-dimensional space) through a nonlinear mapping function Φ(·) (which can be called a kernel function), and then use the PCA (principal component analysis) algorithm to reduce the feature dimensions. Here, the features of the training samples are not linearly separable in the original feature space, but after being processed by the kernel function mapping, they become linearly separable in the high-dimensional space (i.e., the first-dimensional space). The PCA algorithm can be used to reduce the feature dimensions of the features in the high-dimensional space. The KPCA method is used here in the embodiment of the present disclosure, which can also effectively solve the problem of linear inseparability of features.

[0092] Subsequently, the recognition device can perform dimensionality reduction processing on the features of the object to be recognized based on the feature space after the dimensionality reduction, and then perform recognition processing on the object to be recognized based on the features of the object to be recognized after the dimensionality reduction. The object to be recognized is an object whose category has not been recognized, such as speech data whose emotion category has not been recognized, images whose image category has not been recognized, etc.

[0093] In this embodiment, when the recognition device recognizes the object to be recognized, it can use the feature space after the dimension reduction to multiply the feature of the object to be recognized to obtain the low-dimensional feature of the object to be recognized, thereby achieving the purpose of reducing the dimension. The subsequent recognition device can use the low-dimensional feature of the object to be recognized to determine the corresponding category to be recognized. Of course, the subsequent recognition device can also provide the feature space after the dimension reduction to other devices for other devices to perform classification and recognition processing on the object to be recognized.

[0094] Of course, in the embodiment of the present application, after the recognition device determines the feature space with reduced dimensionality, it can also provide it to other devices for the other devices to perform classification and recognition processing on the object to be recognized.

[0095] In the embodiment of the present disclosure, the recognition device can obtain the features of multiple training samples in different dimensions, wherein the multiple training samples include multiple groups of training samples of different categories. The recognition device determines the inter-class distance and intra-class distance corresponding to the features of each dimension according to the features of the multiple training samples in different dimensions, and uses the inter-class distance and intra-class distance corresponding to the features of each dimension to divide the features of different dimensions into a first feature set and a second feature set, wherein the contribution of the features in the first feature set to determining the category is higher than that of the features in the second feature set, and the recognition device determines the feature space after the dimension is reduced according to the first feature set and the second feature set. In this way, the inter-class distance and intra-class distance corresponding to the features of each dimension can be determined, and the first feature set and the second feature set can be determined using the inter-class distance and intra-class distance corresponding to the features of each dimension, wherein the first feature set is an important feature and the second feature set is an unimportant feature, and then the first feature set and the second feature set are used to determine the target feature space after the dimensions of the features of different dimensions are reduced. Since not only the variance of the features in each category of training samples is considered when determining important features and unimportant features, but also the distinguishability of the features between different categories of training samples is considered, the important features determined can be more accurate, and thus the features determined after dimensionality reduction can be more accurate.

[0096] Based on the same technical concept, the embodiment of the present disclosure also provides a device for determining a feature space, such as Figure 3 As shown, the device comprises:

[0097] An acquisition module 310 is used to acquire features of multiple training samples in different dimensions, wherein the multiple training samples include multiple groups of training samples of different categories;

[0098] The determination module 320 is used to:

[0099] Determine the inter-class distance and intra-class distance corresponding to each dimension of the feature according to the features of the multiple training samples in different dimensions;

[0100] According to the inter-class distance and intra-class distance corresponding to each dimension of the feature, the features of the different dimensions are divided into a first feature set and a second feature set, wherein the contribution of the features in the first feature set to determining the category is greater than that of the features in the second feature set;

[0101] A target feature space after the features of different dimensions are reduced in dimension according to the first feature set and the second feature set.

[0102] In a possible implementation, the determining module 320 is configured to:

[0103] Determine the inter-class distance corresponding to each dimensional feature according to the number of training samples of each category in the multiple training samples, the number of categories of the multiple training samples, the mean of each dimensional feature in the training samples of each category, and the mean of each dimensional feature in the multiple training samples;

[0104] The intra-class distance corresponding to each dimensional feature is determined according to the number of categories of the multiple training samples, the mean value of each dimensional feature in each category of training samples, and the feature value of each training sample in each category of training samples.

[0105] In a possible implementation, the determining module 320 is configured to:

[0106] Determine the ratio of the inter-class distance to the intra-class distance corresponding to each dimension of the feature, and obtain the ratio corresponding to each dimension of the feature;

[0107] According to the ratio corresponding to each dimensional feature, the features of different dimensions are divided into a first feature set and a second feature set.

[0108] In a possible implementation, the determining module 320 is configured to:

[0109] According to the order of the ratios corresponding to the features of each dimension from large to small, m features with the largest ratios are selected to form a first type of feature set, and the remaining features of the features of different dimensions except the features in the first type of feature set are used to form a second type of feature set, where m is greater than or equal to 1 and less than or equal to the total dimension of the features.

[0110] In a possible implementation, the determining module 320 is configured to:

[0111] according to Based on the principle of taking the maximum value, the target feature space after the features of different dimensions are reduced in dimension is determined;

[0112] Wherein, W represents the target feature space, is the covariance matrix of the features in the first type of feature set in the first dimensional space, is the covariance matrix of the features in the second feature set in the first dimensional space, express traces, express The dimension of the first dimensional space is higher than the dimension of the target feature space.

[0113] In the embodiments of the present disclosure, it should be noted that: when determining the feature space, the device for determining the feature space provided in the above embodiments only uses the division of the above functional modules as an example. In actual applications, the above functional distribution can be completed by different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the device for determining the feature space provided in the above embodiments and the method embodiment for determining the feature space belong to the same concept. The specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0114] Figure 4 It is a structural diagram of an identification device provided by an embodiment of the present invention. The identification device 400 may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 401 and one or more memories 402, wherein the memory 402 stores at least one instruction, and the at least one instruction is loaded and executed by the processor 401 to implement the above-mentioned method steps for determining the feature space.

[0115] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware or by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.

[0116] The above description is only a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present disclosure should be included in the protection scope of the present disclosure.

Claims

1. A method for determining a feature space, characterized in that: The method comprises: Acquire features of multiple training samples of a speech emotion recognition task in different dimensions, wherein the multiple training samples include multiple groups of training samples of different categories, the features of different dimensions include fundamental frequency, formant, Mel cepstral coefficient, zero crossing rate, amplitude or multiple statistical characteristics of features on sentences, and the categories of the multiple training samples include anger category, happiness category and sadness category; Determine the inter-class distance and intra-class distance corresponding to each dimension of the feature according to the features of the multiple training samples in different dimensions; According to the inter-class distance and intra-class distance corresponding to each dimension of the feature, the features of the different dimensions are divided into a first feature set and a second feature set, wherein the contribution of the features in the first feature set to determining the category is greater than that of the features in the second feature set; A target feature space after dimensionality reduction of the features of different dimensions is determined according to the first feature set and the second feature set.

2. The method according to claim 1, characterized in that: The determining, according to the features of the multiple training samples in different dimensions, the inter-class distance and the intra-class distance corresponding to each dimension of the feature includes: Determine the inter-class distance corresponding to each dimensional feature according to the number of training samples of each category in the multiple training samples, the number of categories of the multiple training samples, the mean of each dimensional feature in the training samples of each category, and the mean of each dimensional feature in the multiple training samples; The intra-class distance corresponding to each dimensional feature is determined according to the number of categories of the multiple training samples, the mean value of each dimensional feature in each category of training samples, and the feature value of each training sample in each category of training samples.

3. The method according to claim 1 or 2, characterized in that: The method of dividing the features of different dimensions into a first feature set and a second feature set according to the inter-class distance and intra-class distance corresponding to each dimension of the features includes: Determine the ratio of the inter-class distance to the intra-class distance corresponding to each dimension of the feature, and obtain the ratio corresponding to each dimension of the feature; According to the ratio corresponding to each dimensional feature, the features of different dimensions are divided into a first feature set and a second feature set.

4. The method according to claim 3, characterized in that The method of dividing the features of different dimensions into a first feature set and a second feature set according to the ratio corresponding to each dimensional feature includes: According to the order of the ratios corresponding to the features of each dimension from large to small, m features with the largest ratios are selected to form a first type of feature set, and the remaining features of the features of different dimensions except the features in the first type of feature set are used to form a second type of feature set, where m is greater than or equal to 1 and less than or equal to the total dimension of the features.

5. The method according to claim 1 or 2, characterized in that: The step of determining, according to the first feature set and the second feature set, a target feature space after reducing the dimensions of the features of different dimensions comprises: according to Based on the principle of taking the maximum value, the target feature space after the features of different dimensions are reduced in dimension is determined; Wherein, W represents the target feature space, is the covariance matrix of the features in the first type of feature set in the first dimensional space, is the covariance matrix of the features in the second feature set in the first dimensional space, express traces, express The dimension of the first dimensional space is higher than the dimension of the target feature space.

6. A device for determining a feature space, characterized in that: The device comprises: An acquisition module, used for acquiring features of multiple training samples of a speech emotion recognition task in different dimensions, wherein the multiple training samples include multiple groups of training samples of different categories, the features of different dimensions include multiple of fundamental frequency, formant, Mel cepstral coefficient, zero crossing rate, amplitude or statistical characteristics of features on sentences, and the categories of the multiple training samples include anger category, happiness category and sadness category; Identify modules for: Determine the inter-class distance and intra-class distance corresponding to each dimension of the feature according to the features of the multiple training samples in different dimensions; According to the inter-class distance and intra-class distance corresponding to each dimension of the feature, the features of the different dimensions are divided into a first feature set and a second feature set, wherein the contribution of the features in the first feature set to determining the category is greater than that of the features in the second feature set; A target feature space after dimensionality reduction of the features of different dimensions is determined according to the first feature set and the second feature set.

7. The device according to claim 6, characterized in that The determining module is used to: Determine the inter-class distance corresponding to each dimensional feature according to the number of training samples of each category in the multiple training samples, the number of categories of the multiple training samples, the mean of each dimensional feature in the training samples of each category, and the mean of each dimensional feature in the multiple training samples; The intra-class distance corresponding to each dimensional feature is determined according to the number of categories of the multiple training samples, the mean value of each dimensional feature in each category of training samples, and the feature value of each training sample in each category of training samples.

8. The device according to claim 6 or 7, characterized in that The determining module is used to: Determine the ratio of the inter-class distance to the intra-class distance corresponding to each dimension of the feature, and obtain the ratio corresponding to each dimension of the feature; According to the ratio corresponding to each dimensional feature, the features of different dimensions are divided into a first feature set and a second feature set.

9. The device according to claim 8, characterized in that The determining module is used to: According to the order of the ratios corresponding to the features of each dimension from large to small, m features with the largest ratios are selected to form a first type of feature set, and the remaining features of the features of different dimensions except the features in the first type of feature set are used to form a second type of feature set, where m is greater than or equal to 1 and less than or equal to the total dimension of the features.

10. The device according to claim 6 or 7, characterized in that The determining module is used to: according to Based on the principle of taking the maximum value, the target feature space after the features of different dimensions are reduced in dimension is determined; Wherein, w represents the target feature space, is the covariance matrix of the features in the first type of feature set in the first dimensional space, is the covariance matrix of the features in the second feature set in the first dimensional space, express traces, express The dimension of the first dimensional space is higher than the dimension of the target feature space.

Images