A data feature enhancement method and device, electronic equipment and medium

Abstract

The application relates to a data feature enhancement method and device, electronic equipment and a medium, wherein the method comprises the following steps: constructing a feature extraction model, inputting training data into the model for initial training, saving the trained model parameters; inputting the training data into the feature extraction model after initial training, obtaining training features, and constructing a first feature dataset of samples; performing dimensionality increasing processing on the feature vector of each sample in the first feature dataset; projecting the dimensionality-increased feature vector to a pre-shape space to obtain a second feature dataset; constructing a geodesic curve function to fit the second feature dataset; expanding the sample feature data by using the fitted geodesic curve; loading the expanded feature data and original feature data, and training a machine learning model until convergence. Compared with the prior art, in a small sample environment, the application effectively depicts a feature distribution space in combination with a shape space theory, realizes efficient data feature enhancement, and saves time and labor costs.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data enhancement, in particular to a data feature enhancement method and device based on shape space theory, electronic equipment and medium. BACKGROUND

[0002] In data processing using machine learning, especially deep learning, a large number of training samples are required. However, due to equipment cost, labor cost, time cost and other reasons, high-quality training samples with labels are often limited. Therefore, data enhancement is needed to increase the number and diversity of training samples. Data enhancement is a simple and efficient method to improve the performance of machine learning models. There are currently two main ways: one is to perturb the data to a certain extent, such as shifting, rotating, flipping, cropping, adding noise, etc. These processing methods are relatively simple, but also have certain limitations. More complex data processing methods, such as random erasing, Mixup, Cutout, RandAugment, etc. in image processing, can force the model to learn more details of the image data by making more in-depth modifications to the original image data, thereby improving its generalization ability. The other way is to use a generative model to expand the original data. Common generative models include GANs (Generative Adversarial Networks) and VAEs (Autoencoders). Due to their incredible generation capabilities, they have developed rapidly. However, most current generative models still rely on a large amount of data. If the sample data is small, the generated image quality will be poor, and it may generate image samples that do not exist in the real world. SUMMARY

[0003] The purpose of the present application is to provide a data feature enhancement method and device based on shape space theory for small sample training data that may be encountered in the machine learning process. First, the shape space theory is used to extract the features of the training samples as a feature vector projected into a pre-shape space. Then, a geodesic curve is constructed to fit the feature vectors of multiple samples in the pre-shape space. Along the geodesic curve, multiple high-quality feature vectors can be generated to meet the needs of the machine learning model for sample data. This method is easier and faster than current data generation methods, and can also meet the data quantity requirements of model training, thereby improving the data processing capability of the model.

[0004] The purpose of the present application can be achieved by the following technical solutions:

[0005] A data feature enhancement method based on shape space theory, comprising the following steps:

[0006] S1, a feature extraction model is constructed, the training data is input into the model for initial training, and the trained model parameters are saved;

[0007] S2, input the training data into the feature extraction model after the initial training, obtain the training features, and construct a first feature dataset of the samples;

[0008] S3, perform dimensionality increasing processing on the feature vector of each sample in the first feature dataset;

[0009] S4, project the dimensionality-increased feature vector to a pre-shape space to obtain a second feature dataset;

[0010] S5, construct a geodesic curve function to fit the second feature dataset in the pre-shape space;

[0011] S6, use the fitted geodesic curve to expand the sample feature data;

[0012] S7, load the expanded feature data and the original feature data, and train the machine learning model until convergence.

[0013] The step S3 specifically comprises: for each sample feature vector u 0 ={x1, x2, …, x N}, a new element is obtained by using a dimensionality increasing function y=x, and a dimensionality-increased feature vector u N ={x1, y1, x2, y2, …, x N} is obtained, and the vector dimensionality is changed from N to 2N, wherein N is the feature data dimensionality of the image sample.

[0014] The step S4 comprises the following steps:

[0015] S41, subtract the feature mean value of the dimensionality-increased feature vector u from the dimensionality-increased feature vector u to obtain a feature vector u′:

[0016] u′={x′1, y′1, x′2, y′2, …, x′ N ,y′ N |x′ i =x i -X μ ,y′ i =y i -Y μ ,i∈[1,N]}

[0017] wherein,

[0018] S42, divide the feature vector u′ by the length of the vector u′ to obtain a feature vector u″:

[0019] u″={x″1, y″1, x″2, y″2, …, x″ N ,y″ N}

[0020] in,

[0021] S43. Obtain the second feature dataset based on the feature vector u″, denoted as...

[0022] The geodesic curve function is:

[0023]

[0024] Where s represents the radian value; parameters v and w are the parameters to be determined for the geodesic curve, i.e. the characteristic points at the beginning and end of the geodesic curve.

[0025] Step S5 includes the following steps:

[0026] S51. Calculate the second feature dataset. All eigenvectors u″1, u″2, ..., u″ M The distance between them is selected by choosing the feature point with the largest sum of distances as the initialization parameter v. * Sample points;

[0027] S52. Obtain the distance feature vector v * The two furthest known sample feature vectors w 0 ,w 1 And substitute it into the geodesic curve function In the middle, traversing the radians s yields Multiple feature vector sets on the curve parameter v * and By combining them separately, a set of candidate geodesic curves is obtained. Where S represents the number of radian values;

[0028] S53. Calculate the feature datasets separately. All eigenvectors u″1, u″2, ..., u″ M To geodetic curve set The geodesic distance, in the feature vector set Find satisfaction A feature vector Let w * ,in, Representing the feature dataset to geodetic curve set respectively The distance;

[0029] S54, w * Assign to v * Repeat S52-S53 until w * ,v *until the fitting of the second feature dataset is completed.

[0030] The S6 is specifically: using the fitted geodesic curve The number of sample features is expanded, that is, a uniform distribution in a specified range is adopted K radian values are randomly sampled and sequentially substituted into the geodesic line The newly generated feature data {z 1 ,…,z K} are obtained, wherein s represents a radian value.

[0031] The feature extraction model is constructed based on a neural network.

[0032] A data feature enhancement device based on shape space theory, the device comprises:

[0033] A model construction and initial training module: constructing a feature extraction model, inputting training data into the model for initial training, and saving the trained model parameters;

[0034] A data feature enhancement module, configured to perform the following steps: inputting training data into the feature extraction model after initial training, obtaining training features, and constructing a first feature dataset of samples; performing dimensionality increasing processing on the feature vector of each sample in the first feature dataset; projecting the dimensionality-increased feature vector into a pre-shape space to obtain a second feature dataset; constructing a geodesic curve function to fit the second feature dataset in the pre-shape space; and expanding sample feature data using the fitted geodesic curve.

[0035] A retraining module: loading the feature data expanded by the data feature enhancement module and the original feature data, and training a machine learning model until convergence.

[0036] An electronic device comprising a processor and a memory for storing a computer program capable of running on the processor, wherein the processor is configured to run the computer program to perform the data feature enhancement method described above.

[0037] A computer storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the data feature enhancement method described above.

[0038] Compared with the prior art, the present application has the following beneficial effects:

[0039] (1) The core advantage of the present application is that it processes feature data, thereby avoiding the negative influence of data types on the enhancement effect, and is suitable for various data types of data sets.

[0040] (2) Based on the data characteristics, the application constructs a geodesic curve in the pre-shape space, which can improve the richness of the data on the basis of preserving the characteristics of the data itself, thereby providing more information for the training of the machine learning model.

[0041] (3) The application is suitable for the training of the machine learning model in a small sample data environment, and the introduction of the expanded data helps to alleviate the dependence of the model on the data size, improve the training effect of the small sample in the machine learning model, and thus improve the generalization performance of the model, thereby effectively dealing with the problem of insufficient data volume, reducing the time and labor cost, and bringing a positive impact on the research and practice in the field of data science.

[0042] (4) The application directly expands the original data set by generating feature data, which is easier to implement than the current data enhancement method, and the generated data is more rapid. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The method flowchart of the application. DETAILED DESCRIPTION

[0044] The application will be described in detail below in combination with the drawings and specific embodiments. The embodiments are implemented on the premise of the technical solution of the application, and detailed implementation modes and specific operation processes are given, but the protection scope of the application is not limited to the following embodiments.

[0045] The embodiment provides a data feature enhancement method based on shape space theory, which extracts features of an image through a neural network and converts a high-dimensional space, uses the effectiveness and flexibility of a geodesic curve in a shape space, expands new feature data, and improves the classification accuracy of a feature classifier.

[0046] The method of the application is suitable for data sets of various data types, and a feature extraction model is a conventional neural network. The network weight parameters are directly converged and obtained by training the neural network model, then the feature data of the data set is extracted, and the feature data is processed by dimensionality and projection. The shape space theory is used for feature expansion to enhance the performance of the final machine learning model. Specifically, as shown in the figure, the method comprises the following steps: Figure 1

[0047] S1, a feature extraction model based on a neural network is constructed, training data is input into the model for initial training, and the trained model parameters are saved;

[0048] S2, the training data is input into the feature extraction model after initial training, the training features are obtained, and a first feature data set of the sample is constructed Wherein, the number of training set samples is M, and u represents the feature vector of the sample.

[0049] ​S3, dimensionally upgrading each feature vector of the first feature dataset U.

[0050] Specifically, for each feature vector u 0 ={x1,x2,…,x N} of the first feature dataset U, a new element y is obtained by using a dimensionally upgrading function y=x, and a dimensionally upgraded feature vector u N ={x1,y1,x2,y2,…,x N} is obtained, wherein the vector dimension changes from N to 2N, and N is the feature data dimension of the image sample.

[0051] S4, projecting the dimensionally upgraded feature vector u to a pre-shape space to obtain a second feature dataset.

[0052] Specifically, S4 includes the following steps:

[0053] S41, subtracting the feature mean from the dimensionally upgraded feature vector u to obtain a feature vector u′:

[0054] u′={x′1,y′1,x′2,y′2,…,x′ N ,y′ N |x′ i =x i -X μ ,y′ i =y i -Y μ ,i∈[1,N]}

[0055] wherein,

[0056] S42, dividing the feature vector u′ by the length of the vector u′ to obtain a feature vector u″:

[0057] u″={x″1,y″1,x″2,y″2,…,x″ N ,y″ N}

[0058] wherein,

[0059] S43, obtaining the second feature dataset based on the feature vector u″, denoted as

[0060] S5, constructing a geodesic curve function Γ (v,w) (s) to fit the second feature dataset in the pre-shape space

[0061] In this embodiment, the constructed geodesic curve function is represented as:

[0062]

[0063] Wherein, the independent variable s represents the radian value; the parameters v and w are the parameters to be determined for the geodesic curve, namely the characteristic points at the beginning and end of the geodesic curve.

[0064] Specifically, the process for determining the parameters of a geodesic function is as follows:

[0065] S51. Calculate the second feature dataset. All eigenvectors u″1, u″2, ..., u″ M The distance between them is selected by choosing the feature point with the largest sum of distances as the initialization parameter v. * Sample points;

[0066] S52. Obtain the distance feature vector v * The two furthest known sample feature vectors w 0 ,w 1 And substitute it into the geodesic curve function In the middle, traversing the radians s yields Multiple feature vector sets on the curve parameter v * and By combining them separately, a set of candidate geodesic curves is obtained. Where S represents the number of radian values;

[0067] S53. Calculate the feature datasets separately. All eigenvectors u″1, u″2, ..., u″ M To geodetic curve set The geodesic distance, in the feature vector set Find satisfaction A feature vector Let w * ,in, Representing the feature dataset to geodetic curve set respectively The distance;

[0068] S54, w * Assign to v * Repeat S52-S53 until w * ,v * The fitting of the second feature dataset is complete when the values ​​between the two remain unchanged.

[0069] S6. Expand the sample feature data using the fitted geodesic curves.

[0070] Specifically, using the fitted geodetic curves Expand the number of sample features, that is, adopt a uniform distribution within a specified range. K radian values are randomly sampled and substituted into the geodesic line The newly generated feature data {z 1 ,…,z K} is obtained, wherein s represents a radian value.

[0071] S7, load the expanded feature data and the original feature data, and train the machine learning model until convergence.

[0072] The embodiment also provides a data feature enhancement device based on a shape space theory, and the device comprises:

[0073] A model construction and initial training module: a feature extraction model is constructed, training data is input into the model for initial training, and model parameters after training are saved;

[0074] A data feature enhancement module, configured to perform the following steps: inputting the training data into the feature extraction model after initial training, obtaining training features, and constructing a first feature data set of samples; performing dimensionality increasing processing on a feature vector of each sample in the first feature data set; projecting the dimensionality-increased feature vector into a pre-shape space to obtain a second feature data set; constructing a geodesic curve function to fit the second feature data set in the pre-shape space; and expanding the sample feature data by using the fitted geodesic curve.

[0075] A retraining module: load the feature data expanded by the data feature enhancement module and the original feature data, and train the machine learning model until convergence.

[0076] The embodiment also provides an electronic device, comprising a processor and a memory for storing a computer program capable of running on the processor, wherein the processor is configured to run the computer program to perform the data feature enhancement method described above. The specific implementation process of the processor can be referred to the method embodiment, and the implementation principle and technical effects are similar, which will not be described here in detail.

[0077] The embodiment also provides a computer storage medium, which stores a computer program, and the computer program is executed by a processor to implement the data feature enhancement method described above.

[0078] That is, the above functions, if realized in the form of software function units and sold or used as independent products, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in part, or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods according to the embodiments of the present application. The foregoing storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk, and various media that can store program codes.

[0079] The preferred embodiments of the present application are described in detail above. It should be understood that those of ordinary skill in the art can make many modifications and variations without creative labor based on the concept of the present application. Therefore, any technical solutions obtained by logical analysis, reasoning, or limited experiments based on the prior art according to the concept of the present application shall be within the protection scope determined by the claims.

Claims

1. A data feature enhancement method based on shape space theory, characterized in that, Includes the following steps: S1. Construct a feature extraction model, input the training data into the model for initial training, and save the trained model parameters. S2. Input the training data into the initially trained feature extraction model to obtain training features and construct the first feature dataset of the samples. S3. Perform dimensionality increase processing on the feature vector of each sample in the first feature dataset. Specifically, step S3 involves: performing dimensionality increase processing on the feature vector of each sample. Using an upscaling function Find the new element and obtain the eigenvector after dimensionality increase. Vector dimension from Become ,in, The feature data dimension of the image sample; S4. Project the upgraded feature vectors onto the pre-shape space to obtain the second feature dataset; S5. Construct geodesic curve functions to fit the second feature dataset in the pre-shape space; S6. Expand the sample feature data using the fitted geodesic curves. S7. Load the expanded feature data and the original feature data, and train the machine learning model until convergence.

2. The data feature enhancement method based on shape space theory according to claim 1, characterized in that, Step S4 includes the following steps: S41. For the eigenvectors after dimensionality increase Subtracting the mean of its features yields the feature vector. : in, ; S42. Regarding the eigenvectors Divide by vector The modulus length is used to obtain the feature vector. : in, S43, Based on Feature Vectors The second feature dataset is obtained, denoted as .

3. The data feature enhancement method based on shape space theory according to claim 1, characterized in that, The geodesic curve function is: in, Represents radian value; parameter and parameters These are the parameters to be determined for the geodesic curve, i.e., the characteristic points at the beginning and end of the geodesic curve.

4. The data feature enhancement method based on shape space theory according to claim 3, characterized in that, Step S5 includes the following steps: S51. Calculate the second feature dataset. All feature vectors in The distance between them is selected by choosing the feature point with the largest sum of distances as the initialization parameter. Sample points; S52. Obtain the distance feature vector The two furthest known sample feature vectors And substitute it into the geodesic curve function In the middle, traverse the radians get Multiple feature vector sets on the curve , parameters and By combining them separately, a set of candidate geodesic curves is obtained. Where S represents the number of radian values; S53. Calculate the feature datasets separately. All feature vectors in To geodetic curve set The geodesic distance, in the feature vector set Find satisfaction A feature vector , recorded as ,in, Representing the feature dataset to geodetic curve set respectively The distance; S54, will Assigned Repeat S52-S53 until... The fitting of the second feature dataset is complete when the values ​​between the two remain unchanged.

5. The data feature enhancement method based on shape space theory according to claim 1, characterized in that, Specifically, S6 involves: utilizing the fitted geodesic curves... Expand the number of sample features, that is, adopt a uniform distribution within a specified range. Random sampling Substitute the radian values ​​sequentially into the geodesic line. In the process, newly generated feature data is obtained. ,in, This represents the value in radians.

6. The data feature enhancement method based on shape space theory according to claim 1, characterized in that, The feature extraction model is built on a neural network.

7. A data feature enhancement device based on shape space theory, characterized in that, The device includes: Model building and initial training module: Build a feature extraction model, input training data into the model for initial training, and save the trained model parameters; The data feature enhancement module performs the following steps: inputting training data into the initially trained feature extraction model to obtain training features and constructing a first feature dataset for the samples; performing dimensionality upscaling on the feature vectors of each sample in the first feature dataset; projecting the dimensionality upscaled feature vectors onto a pre-shape space to obtain a second feature dataset; constructing a geodesic curve function to fit the second feature dataset in the pre-shape space; and expanding the sample feature data using the fitted geodesic curve. Specifically, performing dimensionality upscaling on the feature vectors of each sample in the first feature dataset involves: processing the feature vectors of each sample... Using an upscaling function Find the new element and obtain the eigenvector after dimensionality increase. Vector dimension from Become ,in, The feature data dimension of the image sample; Retraining module: Loads the feature data expanded by the feature enhancement module and the original feature data, and trains the machine learning model until convergence.

8. An electronic device, characterized in that, It includes a processor and a memory for storing a computer program capable of running on the processor, wherein the processor is used to run the computer program to perform the data feature enhancement method according to any one of claims 1-6.

9. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the data feature enhancement method as described in any one of claims 1-6.

Images