WO2021098618A1

WO2021098618A1 - Data classification method and apparatus, terminal device and readable storage medium

Info

Publication number: WO2021098618A1
Application number: PCT/CN2020/128856
Authority: WO
Inventors: 董师周; 乔宇; 王亚立
Original assignee: 中国科学院深圳先进技术研究院
Priority date: 2019-11-21
Filing date: 2020-11-13
Publication date: 2021-05-27
Also published as: CN110929785B; CN110929785A

Abstract

Disclosed are a data classification method and apparatus, a terminal device and a readable storage medium, wherein same are applicable to the technical field of machine learning. The data classification method comprises: first, receiving data to be classified (S21); and then, inputting said data into a trained classification model to obtain at least one data label of said data (S22), wherein the classification model is obtained by means of training according to a distance focal loss function, and the distance focal loss function is used for representing the difference between a predicted label of sample data and a preset label of the sample data according to the distance between the predicted label of the sample data and the preset label of the sample data. When classification training is carried out by means of sample data, there is no need to take the number of samples into consideration, and it is only necessary to determine a classification boundary according to the difference between a predicted label of the sample data and a preset label of the sample data, such that data imbalance can be effectively reduced, and the trained classification model can more accurately perform label classification on data to be classified.

Description

Data classification method, device, terminal equipment and readable storage medium

Technical field

This application belongs to the field of machine learning technology, and in particular relates to a data classification method, device, terminal device, and readable storage medium.

Background technique

Multi-label classification is one of the basic research tasks of machine learning. Its purpose is to predict multiple category labels that appear in each sample data. When training a classification model, because the type and number of category labels in each sample data are not fixed Therefore, in the predicted labels of the sample data, the data volume of individual category labels is very different from the data volume of other category labels, and data imbalance occurs, which in turn reduces the accuracy of the machine learning model.

In the prior art, when improving the data imbalance, a convolutional neural network can be used to extract the features of the sample data, and then the features are linearly combined through a fully connected layer, and then the classification probability is generated through the sigmoid function, and then the focus loss function Perform back propagation as a loss function, train a classification model, and then perform classification through the trained classification model.

However, when training the classification model according to the prior art, due to the different number of each category in the sample data, for a small number of categories, sufficient sample data features cannot be extracted, and adjustments are made only by relying on the weights in the focus loss function. As a result, the classification boundary cannot be accurately obtained, and therefore, a certain degree of data imbalance is still caused, which affects the accuracy of label classification.

Summary of the invention

The embodiments of the present application provide a data classification method, device, terminal device, and readable storage medium to improve the situation in which the prior art only relies on the weight in the focus loss function for adjustment, resulting in the inability to accurately obtain the classification boundary, resulting in a certain degree The data is not balanced, which affects the accuracy of label classification.

In the first aspect, an embodiment of the present application provides a data classification method, including:

First receive the data to be classified, and then input the data to be classified into the trained classification model to obtain at least one data label of the data to be classified. Among them, the classification model is trained based on the distance focus loss function. The distance focus loss function is used to represent the predicted label of the sample data and the preset label of the sample data according to the predicted label of the sample data and the separation distance of the preset label of the sample data. The gap between.

In some implementation manners, the training method of the trained classification model is: first obtain at least one sample data from a preset database, where each sample data includes at least one preset label. Then through the preset classification model, the predicted label of each sample data is obtained. Then obtain the separation distance between the predicted label of the sample data and the preset label of the sample data. Then according to the separation distance, the maximum distance focus loss value is calculated through the distance focus loss function, where the maximum distance focus loss value is used to indicate the maximum value of the gap between the predicted label of the sample data and the preset label of the sample data. Finally, the preset classification model is trained according to the maximum distance focus loss value, and the trained classification model is obtained.

It should be noted that the predicted label includes N categories, where N is an integer greater than 1.

Correspondingly, obtaining the separation distance between the predicted label of the sample data and the preset label of the sample data includes: obtaining the sample data according to the value of the i-th type predicted label of the sample data and the value of the i-th type preset label of the sample data The separation distance between the i-th type predicted label and the i-th type preset label, where i is an integer greater than or equal to 1 and less than or equal to N.

In some other implementations, according to the value of the i-th type predicted label of the sample data and the value of the i-th type preset label of the sample data, the separation distance between the i-th type predicted label and the i-th type preset label of the sample data is obtained , You can first subtract the value of the i-th type predicted label from the value of the i-th type preset label to obtain the absolute distance between the i-th type predicted label and the i-th type preset label of the sample data. Then, the absolute distance is multiplied by the preset scaling factor to obtain the separation distance between the i-th type predicted label and the i-th type preset label of the sample data.

Optionally, according to the separation distance, the maximum separation focus loss value is calculated through the separation focus loss function, and the i-th type prediction can be adjusted according to the separation distance, the value of the i-th type predicted label, and the value of the i-th type preset label. The value range of the label value is used to obtain the predicted label value of the i-th type after the range adjustment. Then, according to the predicted label value of the i-th type and the distance focus loss function after the range adjustment, the maximum distance focus loss value is obtained.

In some implementations, according to the separation distance, the value of the i-th predicted label, and the value of the i-th preset label, the value range of the i-th predicted label value is adjusted to obtain the i-th predicted label after the range adjustment. The value includes: first multiplying the i-th type preset label value by two and then subtracting one to obtain the i-th type preset label value after mapping. Then subtract the product of the separation distance and the mapped preset label value from the i-th type predicted label value to obtain the mapped i-th type predicted label value. Finally, the mapped i-th type predicted label value is multiplied by the preset range scaling factor to obtain the i-th type predicted label value after the range adjustment.

Optionally, obtaining the maximum distance focus loss value according to the predicted label value of the i-th type after the range adjustment and the distance focus loss function includes: first classifying the predicted label value of the i-th type after the range adjustment, and after obtaining the second classification The i-th class predicted label value. Then, the maximum distance focus loss value is obtained according to the predicted label value and distance focus loss function of the i-th class after the two classifications.

In some implementations, the execution subject of the data classification method is a terminal with image processing capabilities. Exemplarily, the terminal may be a physical terminal, such as a desktop computer, a server, a notebook computer, a tablet computer, etc., or a virtual terminal, such as a cloud server, cloud computing, etc. It should be understood that the above execution subject is only an example, and it is not limited to the above terminal.

In a second aspect, an embodiment of the present application provides a data classification device, including a receiving module, configured to receive data to be classified. The classification module is used to input the data to be classified into the trained classification model to obtain at least one data label of the data to be classified, where the classification model is trained according to the distance focus loss function, and the distance focus loss function is used to calculate the data according to the sample data. The separation distance between the predicted label and the preset label of the sample data represents the gap between the predicted label of the sample data and the preset label of the sample data.

In some implementations, the device further includes a training module for obtaining a trained classification model according to the following steps: first obtain at least one sample data from a preset database, wherein each sample data includes at least one preset label. Then, through the preset classification model, the predicted label of each sample data is obtained. Then obtain the separation distance between the predicted label of the sample data and the preset label of the sample data. Then according to the separation distance, the maximum distance focus loss value is calculated through the distance focus loss function, where the maximum distance focus loss value is used to indicate the maximum value of the gap between the predicted label of the sample data and the preset label of the sample data. Finally, the preset classification model is trained according to the maximum distance focus loss value, and the trained classification model is obtained.

Correspondingly, the training module is specifically used to obtain the value of the i-th type predicted label and the i-th type preset label of the sample data according to the value of the i-th type predicted label of the sample data and the value of the i-th type preset label of the sample data. The separation distance, where i is an integer greater than or equal to 1 and less than or equal to N.

In some other implementations, the training module is specifically used to first subtract the value of the i-th type predicted label from the value of the i-th type preset label to obtain one of the i-th type predicted label and the i-th type preset label of the sample data. The absolute distance between. Then, the absolute distance is multiplied by the preset scaling factor to obtain the separation distance between the i-th type predicted label and the i-th type preset label of the sample data.

Optionally, the training module is specifically configured to first adjust the value range of the i-th type predicted label value according to the separation distance, the value of the i-th type predicted label, and the value of the i-th type preset label, to obtain the adjusted range Type i predicted label value. Then, according to the predicted label value of the i-th type and the distance focus loss function after the range adjustment, the maximum distance focus loss value is obtained.

In some other implementations, the training module is specifically configured to first multiply the preset label value of the i-th type by two and then subtract one to obtain the mapped preset label value of the i-th type. Then subtract the product of the separation distance and the mapped preset label value from the i-th type predicted label value to obtain the mapped i-th type predicted label value. Finally, the mapped i-th type predicted label value is multiplied by the preset range scaling factor to obtain the i-th type predicted label value after the range adjustment.

Optionally, the training module is specifically configured to first classify the predicted label value of the i-th type after the range adjustment, and obtain the predicted label value of the i-th type after the two classification. Then, the maximum distance focus loss value is obtained according to the predicted label value and distance focus loss function of the i-th class after the two classifications.

In the third aspect, the embodiments of the present application provide a terminal device, including: a memory, a processor, and a computer program stored in the memory and running on the processor. Methods.

In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, and the computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as provided in the first aspect.

In the fifth aspect, the embodiments of the present application provide a computer program product, which when the computer program product runs on a terminal device, causes the terminal device to execute the method provided in the above-mentioned first aspect.

It can be understood that, for the beneficial effects of the second aspect to the fifth aspect described above, reference may be made to the related description in the first aspect described above, and details are not repeated here.

Compared with the prior art, the embodiment of the present application has the beneficial effect that the received data to be classified is classified through the trained classification model to obtain at least one data label of the data to be classified. Among them, the trained classification model is obtained by training the preset classification model according to the distance focus loss function. Since the distance focus loss function can represent the gap between the predicted label of the sample data and the preset label of the sample data according to the distance between the predicted label of the sample data and the preset label of the sample data. Therefore, the preset classification model is trained by the distance focus loss function. When the sample data is classified and trained, the number of samples does not need to be considered, and the classification is determined based on the difference between the predicted label of the sample data and the preset label of the sample data. The boundary can effectively reduce the data imbalance, and the trained classification model can more accurately classify the classified data.

Description of the drawings

In order to more clearly describe the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only of the present application. For some embodiments, those of ordinary skill in the art can obtain other drawings based on these drawings without creative labor.

FIG. 1 is a schematic diagram of an application scenario of a data classification method provided by an embodiment of the present application;

2 is a schematic flowchart of a data classification method provided by an embodiment of the present application;

FIG. 3 is a schematic flowchart of a data classification method provided by another embodiment of the present application;

4 is a schematic diagram of sample data labels in a data classification method provided by an embodiment of the present application;

FIG. 5 is a schematic flowchart of a data classification method provided by another embodiment of the present application;

FIG. 6 is a schematic flowchart of a data classification method provided by another embodiment of the present application;

FIG. 7 is a schematic flowchart of a data classification method provided by another embodiment of the present application;

FIG. 8 is a schematic flowchart of a data classification method provided by another embodiment of the present application;

FIG. 9 is a schematic structural diagram of a data classification device provided by an embodiment of the present application;

FIG. 10 is a schematic structural diagram of a data classification device provided by another embodiment of the present application;

FIG. 11 is a schematic structural diagram of a terminal device provided by an embodiment of the present application.

Detailed ways

In the following description, for the purpose of illustration rather than limitation, specific details such as a specific system structure and technology are proposed for a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid unnecessary details from obstructing the description of this application.

Reference to "one embodiment" or "some embodiments" described in the specification of this application means that one or more embodiments of this application include a specific feature, structure, or characteristic described in combination with the embodiment. Therefore, the sentences "in one implementation", "in some implementations", "in some other implementations", "in other implementations", etc. appearing in the differences in this specification are not necessarily All refer to the same embodiment, but mean "one or more but not all embodiments" unless it is specifically emphasized otherwise. The terms "including", "including", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized.

The data classification method provided by the embodiments of this application can be applied to mobile phones, tablet computers, wearable devices, in-vehicle devices, augmented reality (AR)/virtual reality (VR) devices, notebook computers, and super mobile personal computers For terminal devices such as ultra-mobile personal computer (UMPC), netbooks, personal digital assistants (personal digital assistants, PDAs), security cameras, surveillance cameras, etc., the embodiments of this application do not impose any restrictions on the specific types of terminal devices.

Figure 1 shows a schematic diagram of an application scenario of the data classification method provided by the present application. Referring to Figure 1, in this scenario, an image acquisition device 11, a server 12, and a database 13 are included. The image acquisition device 11 and the server 12 communicate with each other, and the server 12 and the database 13 communicate with each other. The communication connection can be wired Network or wireless network, where wireless network can include wireless local area network (Wireless Localarea Networks, WLAN) (such as Wi-Fi network), Bluetooth, Zigbee, mobile communication network, Near Field Communication (NFC), infrared technology (Infrared, IR) and other communication solutions. Wired networks can include optical fiber networks, telecommunication networks, intranets, etc., such as Local Area Network (LAN), Wide Area Network (WAN), Metropolitan Area Network (MAN), and public switched telephone network ( Public Switched Telephone Network, PSTN), etc. There are no restrictions on the types of wireless networks and wired networks.

As an example and not a limitation, the image acquisition device 11 may include a tablet computer 111, a notebook computer 112, a desktop computer 113, a smart phone 114, a digital camera 115, a surveillance camera 116, etc. The method of acquiring an image may be a real-time image captured by a camera. It may also be to call the image stored in the image acquisition device 11, or to access a server or database storing the image through the image acquisition device 11, and forward the image to the server 12.

For example, when the image acquisition device 11 is a device with a photographing function, such as a smart phone 114, a digital camera 115, or a surveillance camera 116, a real-time image can be captured by the camera and sent to the server 12.

When the image acquisition device 11 is a tablet computer 111, a notebook computer 112, a desktop computer 113, etc., the image stored therein can be sent to the server 12. At this time, the image acquisition device 11 and the server 12 can be two separate devices. That is, the server 12 is a cloud server, a rack server, a cabinet server, a blade server, etc.; or the image acquisition device 11 and the server 12 may also be the same device, for example, the server 12 may be a virtual server running on a desktop computer 113 , There is no restriction here.

In the same way, the database 13 and the server 12 can be implemented on the same device, or on different devices. The implementation is a common method used by those skilled in the art and will not be repeated here.

FIG. 2 shows a schematic flowchart of a data classification method provided by an embodiment of the present application. As an example and not a limitation, the method can be applied to terminal devices in the above-mentioned scenarios, such as a tablet computer 111, a notebook computer 112, a desktop computer 113, Smart phone 114, digital camera 115 or surveillance camera 116, etc.

Please refer to Figure 2. The data classification methods include:

S21. Receive data to be classified.

It should be noted that in this application, pictures are used as the data to be classified for description, but the type of data to be classified is not limited to this. For example, the data to be classified can also be data in other forms such as video, text, and audio. At this time, it is necessary to make corresponding adjustments according to the type of data, and the adjustment method is a common method of those skilled in the art, and will not be repeated here.

S22. Input the data to be classified into the trained classification model to obtain at least one data label of the data to be classified.

Among them, the classification model is trained based on the distance focus loss function. The distance focus loss function is used to represent the predicted label of the sample data and the preset label of the sample data according to the predicted label of the sample data and the separation distance of the preset label of the sample data. The gap between.

It should be noted that if the gap between the predicted label of the sample data and the preset label of the sample data is small, it means that the category is easy to distinguish, and the boundary of the classification can be closer (that is, the separation distance); otherwise, it means that the category It is difficult to distinguish, and the classification boundary needs to be set farther to reduce the difficulty of distinguishing.

In this implementation manner, the received data to be classified is classified through the trained classification model to obtain at least one data label of the data to be classified. Among them, the trained classification model is obtained by training the preset classification model according to the distance focus loss function. Since the distance focus loss function can represent the gap between the predicted label of the sample data and the preset label of the sample data according to the distance between the predicted label of the sample data and the preset label of the sample data. Therefore, the preset classification model is trained by the distance focus loss function. When the sample data is classified and trained, the number of samples does not need to be considered, and the classification is determined based on the difference between the predicted label of the sample data and the preset label of the sample data. The boundary can effectively reduce the data imbalance, and the trained classification model can more accurately classify the classified data.

Please refer to Fig. 2. In another embodiment of the data classification method, the training method of the trained classification model is:

S31. Obtain at least one sample data from a preset database.

Wherein, each sample data includes at least one preset label.

In some implementations, multiple sample data and at least one preset label corresponding to each sample data are stored in the preset database. For example, when the sample data is a picture, if there are both cats and dogs in the picture, Then there are two preset labels corresponding to the sample data, namely "cat" and "dog".

The preset label can be represented by a vector containing N elements, where N is the number of label categories, N is an integer greater than 1, and the value range of each element is [0,1].

As an example and not a limitation, please refer to Figure 4. Figure 4 shows a schematic diagram of sample data labels. Assuming there are 4 types of labels in sample data, they are square, circle, triangle, and diamond. In Figure 4 , There are square 15, circle 16, triangle 17, the preset label vector y of the sample data can be expressed as y∈[1,1,1,0].

Among them, the value of the first element is 1, which means that in the sample data, the probability of the label of the first category (ie the label of square 15) is 100%; the value of the second element is 1, which means that In the sample data, the probability of the label of the second category (that is, the label of the circle 16) is 100%, and the value of the third element is 1, which means that there is a label of the third category in the sample data ( That is, the probability of the label of the triangle 17) is 100%; the value of the fourth element is 0, which means that the probability of the label of the fourth category (that is, the label of the diamond) is 0% in the sample data.

S32. Obtain the predicted label of each sample data through a preset classification model.

In some embodiments, the predicted label can also be represented by a vector containing N elements. Generally speaking, for the same batch of sample data, the number of label categories is certain, that is, the predicted label also contains labels of N categories. Refer to the example in S31 and Figure 4, the predicted label vector of the sample data can be used

Means that the value of the first element is 0.9, which means that in the sample data, the probability of the label of the first category (ie the label of square 15) is 90%; the value of the second element is 0.7, then Indicates that in the sample data, there is a probability of 70% of the label of the second category (that is, the label of the circle 16); the value of the third element is 0.6, which means that there is a label of the third category in the sample data The probability of the label (that is, the label of the triangle 17) is 60%; the value of the fourth element is 0, which means that in the sample data, the probability of the label of the fourth category (that is, the label of the diamond) is 80%.

S33. Obtain the separation distance between the predicted label of the sample data and the preset label of the sample data.

Refer to the predicted label vector of the sample data given in S32 and S31 and the preset label vector of the sample data, because

It is the prediction result obtained through the classification of the preset classification model, and there is a certain difference between y, and this difference is the separation distance between the predicted label of the sample data and the preset label of the sample data.

In some embodiments, the value of the label can be predicted based on the i-th type of sample data

_{The value y i of} the i-th type preset label of the sample data to obtain the interval distance between the i-th type predicted label and the i-th type preset label of the sample data, where i is an integer greater than or equal to 1 and less than or equal to N .

Referring to FIG. 5, the method of obtaining the separation distance between the i-th type predicted label and the i-th type preset label of the sample data may include:

S331. Subtract the value of the i-th type of predicted label from the value of the i-th type of preset label to obtain the absolute distance between the i-th type of predicted label and the i-th type of preset label of the sample data.

As an example and not a limitation, refer to S31 and S32 in

And y, the value y ₁ of the preset label of the first type is 1, and the value of the predicted label of the first type

Is 0.9, then the absolute distance between the first type of predicted label and the first type of preset label of the sample data is

S332. Multiply the absolute distance by the preset scaling factor to obtain the separation distance between the i-th type predicted label and the i-th type preset label of the sample data.

In some embodiments, the preset zoom factor can be expressed by λ, then the separation distance

due to

The value range of the elements in and y is [0,1], so the absolute distance

The range is also between [0,1]. It should be noted that the smaller the absolute distance of the i-th category, the easier it is to distinguish the i-th category, and the larger the absolute distance of the i-th category, the more difficult it is to distinguish the i-th category.

However, when the absolute distance is between [0,1], because the value range is narrow, it is difficult to effectively reflect the degree of easy distinction of the category. Therefore, it is easier to judge the absolute distance by magnifying the absolute distance by λ times and expanding the absolute distance. Whether the category is easy to distinguish, for example, refer to the examples in S31 and S32, y∈[1,1,1,0],

When it is not zoomed, the absolute distance of the second category is 0.3, and the absolute distance of the third category is 0.4. The two are close to the midpoint of the value range 0.5, indicating that the two are between easy to distinguish and difficult to distinguish. Set λ to 4, and then scale the absolute distance above. The absolute distance of the second category is enlarged from 0.3 to 1.2, and the absolute distance of the third category is enlarged from 0.4 to 1.6. The distance has also been enlarged by four times, making the two farther away from the midpoint of the value range, making it easier to judge whether the category is easy to distinguish.

In the above embodiment, the absolute distance between the i-th type prediction label and the i-th type preset label of the sample data is enlarged, and the enlarged absolute distance is used as the i-th type prediction label and the i-th type prediction label of the sample data. Setting the separation distance between the labels enlarges the gap between the predicted label of the sample data and the preset label of the sample data, making the decision boundary clearer, and obtaining the predicted label of the sample data more accurately. At the same time, since the absolute distance is obtained by subtracting the value of the i-th type predicted label from the value of the i-th preset label, for each training, the absolute distance of the i-th type will adaptively change according to the value of the i-th type predicted label. , So that the obtained separation distance is more accurate, and then the predicted label prediction of the category is more accurate, and the prediction effect of the classification model is improved.

S34: According to the separation distance, calculate the maximum separation focus loss value through the separation focus loss function.

Among them, the maximum distance focus loss value is used to indicate the maximum value of the gap between the predicted label of the sample data and the preset label of the sample data.

Referring to Figure 6, the maximum pitch focus loss value can be calculated in the following way.

S341: According to the separation distance, the value of the i-th type of predicted label and the value of the i-th type of preset label, adjust the value range of the i-th type of predicted label value to obtain the i-th type of predicted label value after the range adjustment.

In some embodiments, the value range of the predicted label value of the i-th type can be adjusted by using the spacing distance, the value of the i-th type predicted label and the value of the i-th type preset label, and the output of the predicted label value can be adjusted on the basis of S33. The changing curvature of the curve makes the decision boundary of the value of the i-th prediction label after the range adjustment clearer, and improves the prediction effect of the classification model.

Referring to Figure 7, according to the separation distance, the value of the i-th type predicted label and the value of the i-th type preset label, adjust the value range of the i-th type predicted label value to obtain the i-th type predicted label value after the range adjustment. include:

S3411, multiply the i-th type preset label value by two and then subtract one to obtain the i-th type preset label value after mapping.

S3412, subtract the product of the separation distance and the mapped preset label value of the i-th type from the predicted label value of the i-th type to obtain the predicted label value of the i-th type after the mapping.

S3413: Multiply the mapped predicted label value of the i-th type by a preset range scaling factor to obtain the predicted label value of the i-th type after the range adjustment.

In some implementations, the steps in S3411, S3412, and S3413 can be expressed by a formula, that is, the predicted label value of the i-th category after the range adjustment

Is calculated as:

Among them, s is the scale scaling factor.

As an example and not a limitation, refer to the examples in S31 and S32,

The value range of and y _i are both [0,1], the value of s can be set to 10, then

The value range of is [-10m _i ,10+10m _i ].

Relative to

When the number of types of predicted labels is the same, the curvature of the predicted label value output curve is greater, and the difference between the predicted label values of different categories is larger, making the decision boundary of the i-th predicted label clearer.

S342: Obtain a maximum distance focus loss value according to the i-th type predicted label value and the distance focus loss function after the range adjustment.

Among them, the method of obtaining the focal loss value of the maximum distance can be through the following steps:

S3421 Perform two classifications on the predicted label value of the i-th type after the range adjustment, and obtain the predicted label value of the i-th type after the two classification.

Since the decision boundary of the predicted label value of the i-th type after the range adjustment is very clear, it is necessary to classify the predicted label value of the i-th type after the range adjustment to determine whether each type of label exists in the sample data.

There are many ways of binary classification, such as using Sigmoid function, Logistic regression and so on.

As an example and not a limitation, when using the Sigmoid function to calculate, the predicted label value of the i-th class after two classifications

It can be expressed by the following formula:

S3422, according to the predicted label value and the distance focus loss function of the i-th category after the two classifications, obtain the maximum distance focus loss value.

In some embodiments, the pitch focus loss function is:

will

As

(which is

) Substitute into the formula to get the focal loss function of the maximum distance:

Among them, w _i ⁰ represents the weight of the corresponding loss function when there is no predicted label of the i-th category in the _{sample data; w i} ¹ represents the weight of the corresponding loss function when the predicted label of the i-th category exists in a sample data, and its calculation The way is:

Both α and β are preset parameters. In some embodiments, α=0.5 and β=2, but not limited to this.

Finally, through the maximum distance focus loss function, the maximum focus loss value of each category is calculated.

S35: Train a preset classification model according to the maximum distance focus loss value, and obtain a trained classification model.

It should be noted that the maximum distance focus loss value can be used for back propagation, multiple iterations, and repeated training of the preset classification model, and finally the trained classification model is obtained. The specific training method is not limited here.

Here, the training of the automatic picture classification model is taken as an example to illustrate the application scenarios of the data classification method provided in this application.

First, a large number of image samples are collected first, as sample data, which can be represented by D={<x _i ,y _i >|i=1, 2, 3...N}. Where x _i is an image sample, and y _i are multiple category labels corresponding to the image sample.

Then, the machine learning classification model is determined. Here, a convolutional neural network f _θ can be used, where θ is a parameter of the model.

Next, input B image samples into the convolutional neural network f _θ , and update the parameters θ of the convolutional neural network according to the following formula:

Among them, L is the maximum focus loss value calculated by the maximum focus loss function provided in this application.

Then, iterate the previous step T times until the model converges or L is less than the preset threshold, and the trained classification model f _θ* can be obtained.

Finally, the image x to be predicted is input to the trained classification model f _θ* , and the multi-class label vector f _θ* (x) of the image to be predicted is output.

It should be understood that the size of the sequence number of each step in the foregoing embodiment does not mean the order of execution. The execution sequence of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present application.

Corresponding to the data classification method described in the above embodiment, FIG. 9 shows a schematic structural diagram of a data classification device provided in an embodiment of the present application. For ease of description, only the parts related to the embodiment of the present application are shown.

Referring to FIG. 9, the device includes: a receiving module 51 for receiving data to be classified. The classification module 52 is used to input the data to be classified into the trained classification model to obtain at least one data label of the data to be classified, wherein the classification model is obtained by training according to the distance focus loss function, and the distance focus loss function is used according to the sample data The separation distance between the predicted label of the sample data and the preset label of the sample data represents the gap between the predicted label of the sample data and the preset label of the sample data.

In some implementations, referring to FIG. 10, the device further includes a training module 53 for obtaining a trained classification model according to the following steps: first obtain at least one sample data from a preset database, wherein each sample data includes at least one The preset label. Then, the predicted label of each sample data is obtained through the preset classification model. Then obtain the separation distance between the predicted label of the sample data and the preset label of the sample data. Then according to the separation distance, the maximum distance focus loss value is calculated through the distance focus loss function, where the maximum distance focus loss value is used to indicate the maximum value of the gap between the predicted label of the sample data and the preset label of the sample data. Finally, the preset classification model is trained according to the maximum distance focus loss value, and the trained classification model is obtained.

Correspondingly, the training module 53 is specifically configured to obtain the value of the i-th type predicted label and the i-th type preset label of the sample data according to the value of the i-th type predicted label of the sample data and the value of the i-th type preset label of the sample data The separation distance of, where i is an integer greater than or equal to 1 and less than or equal to N.

In some other implementations, the training module 53 is specifically configured to first subtract the value of the i-th type predicted label from the value of the i-th type preset label to obtain the i-th type predicted label and the i-th type preset label of the sample data. The absolute distance between. Then, the absolute distance is multiplied by the preset scaling factor to obtain the separation distance between the i-th type predicted label and the i-th type preset label of the sample data.

Optionally, the training module 53 is specifically configured to first adjust the value range of the i-th type predicted label value according to the separation distance, the value of the i-th type predicted label, and the value of the i-th type preset label, to obtain the adjusted range The i-th type predicts the label value. Then, according to the predicted label value of the i-th type and the distance focus loss function after the range adjustment, the maximum distance focus loss value is obtained.

In some other implementations, the training module 53 is specifically configured to first multiply the preset label value of the i-th type by two and then subtract one to obtain the mapped preset label value of the i-th type. Then subtract the product of the separation distance and the mapped preset label value from the i-th type predicted label value to obtain the mapped i-th type predicted label value. Finally, the mapped i-th type predicted label value is multiplied by the preset range scaling factor to obtain the i-th type predicted label value after the range adjustment.

Optionally, the training module 53 is specifically configured to first classify the predicted label value of the i-th type after the range adjustment, and obtain the predicted label value of the i-th type after the two classification. Then, the maximum distance focus loss value is obtained according to the predicted label value and distance focus loss function of the i-th class after the two classifications.

It should be noted that the information exchange and execution process between the above-mentioned devices are based on the same concept as the method embodiment of this application, and its specific functions and technical effects can be found in the method embodiment section. Go into details again.

Those skilled in the art can clearly understand that, for the convenience and conciseness of description, only the division of the above functional units and modules is used as an example. In practical applications, the above functions can be allocated to different functional units and modules as needed. Module completion, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist alone physically, or two or more units can be integrated into one unit. The above-mentioned integrated units can be hardware-based Formal realization can also be realized in the form of software functional units. In addition, the specific names of the functional units and modules are only used to facilitate distinguishing from each other, and are not used to limit the protection scope of the present application. For the specific working process of the units and modules in the foregoing system, reference may be made to the corresponding process in the foregoing method embodiment, which will not be repeated here.

FIG. 11 shows a schematic structural diagram of a terminal device provided by an embodiment of the present application. Referring to FIG. 11, the terminal device 6 includes:

The memory 62, the processor 61, and a computer program 63 that is stored in the memory 62 and can run on the processor 61, and the processor 61 implements the steps in the foregoing method embodiments when the computer program 63 is executed.

The embodiments of the present application also provide a computer-readable storage medium, where the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in each of the foregoing method embodiments can be realized.

The embodiments of the present application provide a computer program product. When the computer program product runs on a mobile terminal, the steps in the foregoing method embodiments can be realized when the mobile terminal is executed.

If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on this understanding, the implementation of all or part of the processes in the above-mentioned embodiment methods in the present application can be accomplished by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. The computer program can be stored in a computer-readable storage medium. When executed by the processor, the steps of the foregoing method embodiments can be implemented. Wherein, the computer program includes computer program code, and the computer program code may be in the form of source code, object code, executable file, or some intermediate forms. The computer-readable medium may at least include: any entity or device capable of carrying the computer program code to the photographing device/terminal device, recording medium, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunications signal and software distribution medium. For example, U disk, mobile hard disk, floppy disk or CD-ROM, etc. In some jurisdictions, according to legislation and patent practices, computer-readable media cannot be electrical carrier signals and telecommunication signals.

In the above-mentioned embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail or recorded in an embodiment, reference may be made to related descriptions of other embodiments.

A person of ordinary skill in the art may realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented by electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software depends on the specific application and design constraint conditions of the technical solution. Professionals and technicians can use different methods for each specific application to implement the described functions, but such implementation should not be considered beyond the scope of this application.

In the embodiments provided in this application, it should be understood that the disclosed apparatus/network equipment and method may be implemented in other ways. For example, the device/network device embodiments described above are only illustrative. For example, the division of the modules or units is only a logical function division, and there may be other divisions in actual implementation, such as multiple units. Or components can be combined or integrated into another system, or some features can be omitted or not implemented. In addition, the displayed or discussed mutual coupling or direct coupling or communication connection may be indirect coupling or communication connection through some interfaces, devices or units, and may be in electrical, mechanical or other forms.

The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solutions of the embodiments.

The above-mentioned embodiments are only used to illustrate the technical solutions of the present application, not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, a person of ordinary skill in the art should understand that it can still implement the foregoing The technical solutions recorded in the examples are modified, or some of the technical features are equivalently replaced; these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the application, and should be included in Within the scope of protection of this application.

Claims

A data classification method, characterized in that it includes:

Receive data to be classified;

Input the to-be-classified data into the trained classification model to obtain at least one data label of the to-be-classified data, wherein the classification model is obtained by training according to the distance focus loss function, and the distance focus loss function is The distance between the predicted label of the sample data and the preset label of the sample data indicates the gap between the predicted label of the sample data and the preset label of the sample data.
The method according to claim 1, wherein the training method of the trained classification model is:

Acquiring at least one of the sample data from a preset database, wherein each of the sample data includes at least one preset label;

Obtain the predicted label of each of the sample data through a preset classification model;

Obtaining the separation distance between the predicted label of the sample data and the preset label of the sample data;

According to the spacing distance, the maximum spacing focus loss value is calculated by the spacing focus loss function, wherein the maximum spacing focus loss value is used to indicate one of the predicted label of the sample data and the preset label of the sample data The maximum value of the gap between;

The preset classification model is trained according to the maximum distance focus loss value, and the trained classification model is obtained.
The method according to claim 2, wherein the predicted label includes N categories, wherein N is an integer greater than 1;

Correspondingly, the obtaining the separation distance between the predicted label of the sample data and the preset label of the sample data includes:

According to the value of the i-th type of predicted label of the sample data and the value of the i-th type of preset label of the sample data, the separation distance between the i-th type of predicted label and the i-th type of preset label of the sample data is obtained, Wherein, i is an integer greater than or equal to 1 and less than or equal to N.
The method according to claim 3, characterized in that, according to the value of the predicted label of the i-th type of the sample data and the value of the preset label of the i-th type of the sample data, the i-th type of the sample data is obtained The distance between the predicted label and the i-th preset label includes:

Subtracting the value of the i-th type prediction label from the value of the i-th type preset label to obtain the absolute distance between the i-th type prediction label and the i-th type preset label of the sample data;

The absolute distance is multiplied by a preset scaling factor to obtain the separation distance between the i-th type predicted label and the i-th type preset label of the sample data.
The method according to claim 3 or 4, wherein the calculating and obtaining a maximum pitch focus loss value through the pitch focus loss function according to the pitch distance comprises:

According to the separation distance, the value of the i-th type prediction label, and the value of the i-th type preset label, adjust the value range of the i-th type prediction label value to obtain the i-th type prediction after the range adjustment Label value

According to the predicted tag value of the i-th type after the range adjustment and the distance focus loss function, the maximum distance focus loss value is obtained.
The method according to claim 5, wherein the adjustment of the i-th type predicted label is based on the separation distance, the value of the i-th type predicted label, and the value of the i-th type preset label The value range of the value to obtain the i-th type predicted label value after the range adjustment, including:

Multiplying the i-th type preset label value by two and then subtracting one to obtain the i-th type preset label value after mapping;

Subtracting the product of the separation distance and the mapped i-th preset label value from the i-th type predicted label value to obtain a mapped i-th type predicted label value;

Multiplying the mapped i-th type predicted label value by a preset range scaling factor to obtain the i-th type predicted label value after the range adjustment.
The method according to claim 5, wherein the obtaining the maximum distance focus loss value according to the i-th type predicted label value after the range adjustment and the distance focus loss function comprises:

Perform two classifications on the predicted label value of the i-th category after the range adjustment, and obtain the predicted label value of the i-th category after the two classification;

Obtain the maximum distance focus loss value according to the predicted label value of the i-th type after the two classifications and the distance focus loss function.
A data classification device is characterized in that it comprises:

The receiving module is used to receive the data to be classified;

The classification module is configured to input the data to be classified into the trained classification model to obtain at least one data label of the data to be classified, wherein the classification model is trained according to the distance focus loss function, and the The pitch focus loss function is used to represent the gap between the predicted label of the sample data and the preset label of the sample data according to the separation distance between the predicted label of the sample data and the preset label of the sample data.
A terminal device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program as claimed in claims 1 to 7. The method of any one.
A computer-readable storage medium storing a computer program, wherein the computer program implements the method according to any one of claims 1 to 7 when the computer program is executed by a processor.