WO2020082595A1

WO2020082595A1 - Image classification method, terminal device and non-volatile computer readable storage medium

Info

Publication number: WO2020082595A1
Application number: PCT/CN2018/124630
Authority: WO
Inventors: 金戈; 徐亮; 肖京
Original assignee: 平安科技（深圳）有限公司
Priority date: 2018-10-26
Filing date: 2018-12-28
Publication date: 2020-04-30
Also published as: CN109522939A; CN109522939B

Abstract

Disclosed are an image classification method, a terminal device, and a non-volatile computer readable storage medium, applicable to the technical field of computers. The method comprises: obtaining a target image to be classified (S101); on the basis of optimal parameters in an image classification model, performing feature extraction on the target image to obtain image features, and performing classification prediction process on the image features to obtain an image classification result (S102); the optimal parameters are obtained on the basis of a preset noise value when the 2-norm of the loss function of the image classification model is less than a first preset value, and the preset noise value is used to enable model parameters determined by the trained image classification model to avoid saddle points during iterative optimization; and outputting the image classification result (S103). The described image classification method can analyze the image features of an input image on the basis of the optimal parameters in the model, thereby improving the classification accuracy of the image classification model.

Description

Image classification method, terminal device and computer non-volatile readable storage medium

This application requires the priority of the Chinese patent application filed on October 26, 2018 in the China Patent Office, with the application number 201811255779.6 and the invention titled "Image Classification Method, Terminal Equipment, and Computer-readable Storage Media" Incorporated in this application.

Technical field

The present application belongs to the field of computer technology, and particularly relates to an image classification method, a terminal device, and a computer non-volatile readable storage medium.

Background technique

Image classification models based on deep learning or partial machine learning require training before they can be used to perform specific image classification functions, such as ethnic classification functions. The process of training the image classification model is actually the process of optimizing the parameters in the image classification model, that is, to find the optimal parameters of the image classification model. After the image classification model training is completed, the image classification model can be used To perform the corresponding image classification function.

When optimizing the parameters in the model, common momentum optimization algorithms such as stochastic gradient descent algorithm can generally be used to update the parameters in the image classification model to find the optimal parameters. The stochastic gradient descent algorithm specifically needs to determine whether the model finds the optimal parameter by whether the loss function in the image classification model reaches the global minimum. However, when using the stochastic gradient descent algorithm, the loss function may be caused by the saddle point in the loss function It will not be able to converge to the global extremum point, and the optimal parameters of the image classification model cannot be determined. The image classification model needs to analyze the image characteristics of the input image based on the optimal parameters in the model. For the image classification model that cannot determine the optimal parameters, the classification accuracy of the corresponding image classification model decreases.

technical problem

An embodiment of the present application provides an image classification method, terminal device, and computer non-volatile readable storage medium to solve the problem of low classification accuracy of the image classification model in the prior art.

Technical solution

A first aspect of the embodiments of the present application provides that the first aspect provides an image classification method, including:

Obtain the target image to be classified;

Based on the optimal parameters in the image classification model, the target image is subjected to feature extraction to obtain image features, and the image features are subjected to classification prediction processing to obtain image classification results, wherein the optimal parameters are in the image classification model When the second norm of the loss function is less than the first preset value, it is obtained based on a preset noise value, which is used to make the model parameters determined by the trained image classification model avoid the saddle point during iterative optimization;

The image classification result is output.

A second aspect of the embodiments of the present application provides a terminal device. The terminal device includes a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor. The processor The following steps are realized when the computer-readable instructions are executed:

Obtain the target image to be classified;

Based on the optimal parameters in the image classification model, the target image is subjected to feature extraction to obtain image features, and the image features are subjected to classification prediction processing to obtain image classification results, wherein the optimal parameters are in the When the second norm of the loss function is less than the first preset value, it is obtained based on a preset noise value, which is used to make the model parameters determined by the trained image classification model avoid the saddle point during iterative optimization;

The image classification result is output.

A third aspect of the embodiments of the present application provides a terminal device, including:

An obtaining unit, used to obtain the target image to be classified;

The execution unit is configured to perform feature extraction on the target image based on the optimal parameters in the image classification model to obtain image features, and perform classification prediction processing on the image features to obtain image classification results, where the optimal parameters are in all When the second norm of the loss function of the image classification model is less than the first preset value, it is obtained based on a preset noise value, and the preset noise value is used to make iterative optimization of the model parameters determined by the trained image classification model Avoid the saddle point

The output unit is used to output the image classification result.

A fourth aspect of the embodiments of the present application provides a computer nonvolatile readable storage medium, the computer nonvolatile readable storage medium stores computer readable instructions, and the computer readable instructions are executed by a processor The following steps are implemented:

Obtain the target image to be classified;

The image classification result is output.

Beneficial effect

In the embodiment of the present application, the terminal device acquires the target image to be classified; based on the optimal parameters in the image classification model, the target image is subjected to feature extraction to obtain image features, and the image features are subjected to classification prediction processing to obtain image classification results, The optimal parameter is obtained based on a preset noise value when the second norm of the loss function of the image classification model is less than a first preset value, and the preset noise value is used to classify the trained image The model parameters determined by the model avoid the saddle point during iterative optimization, so that the terminal device can extract the feature of the target image based on the optimal parameters in the image classification model to obtain the image feature, and can more accurately extract the image feature corresponding to the target image ; When the terminal device classifies and predicts the image features based on the optimal parameters in the image classification model to obtain the image classification result, the predicted image classification result will also be more accurate.

BRIEF DESCRIPTION

1 is a flowchart of an image classification method provided by the first embodiment of the present application;

2 is a flowchart of an image classification method provided by the second embodiment of the present application;

3 is a schematic diagram of a terminal device according to a third embodiment of the present application;

4 is a schematic diagram of a terminal device according to a fourth embodiment of the present application.

Embodiments of the invention

In the following description, for the purpose of illustration rather than limitation, specific details such as specific system structure and technology are proposed to thoroughly understand the embodiments of the present application. However, those skilled in the art should understand 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 so as not to obscure the description of the present application with unnecessary details.

Referring to FIG. 1, FIG. 1 is a flowchart of an image classification method in the first embodiment of the present application. The execution subject of the image classification method in this embodiment is a terminal device. The image classification method as shown in the figure may include the following steps:

S101. Acquire a target image to be classified.

In S101, when a user needs to perform classification processing on a target image to be classified through the terminal device, the user may input the target image to be classified into the terminal device, and the terminal device acquires the target image to be classified. Among them, the terminal device classifies the target image based on the pre-stored image classification model pre-stored in the terminal device. The image classification model may specifically be a classification model that implements a race classification function. All classification results that the image classification model can predict include at least two Of course, it is not limited to this.

S102: Perform feature extraction on the target image based on the optimal parameters in the image classification model to obtain image features, and perform classification prediction processing on the image features to obtain image classification results, where the optimal parameters are classified in the image When the second norm of the loss function of the model is less than the first preset value, it is obtained based on a preset noise value, and the preset noise value is used to avoid the model parameters determined by the trained image classification model during the iterative optimization. Saddle point.

In S102, after the image classification model has been trained, the terminal device performs feature extraction on the target image based on the optimal parameters in the image classification model to obtain image features, and performs classification prediction processing on the image features to obtain image classification results. Image classification The classification of model prediction is generally only one. Among them, since the optimal parameters in the image classification model are obtained based on a preset noise value when the second norm of the loss function of the image classification model is less than the first preset value, the preset noise value is used to classify the trained image The model parameters determined by the model avoid the saddle point during iterative optimization, so that the optimal parameters are those determined when the image classification model converges to the global extremum during training, and the terminal device is based on the optimal parameters in the image classification model When feature extraction is performed on the target image to obtain image features, the image features corresponding to the target image can be more accurately extracted; the terminal device performs classification prediction processing on the image features based on the optimal parameters in the image classification model to obtain the image classification result. The predicted image classification results will also be more accurate. Specifically, the image classification model may include a convolutional layer and a fully connected layer. The model parameters may specifically be parameters in the convolutional layer and the fully connected layer. The terminal device performs based on the parameter target image corresponding to the convolutional layer in the image classification model. Convolution calculation to extract the image features corresponding to the target image; the terminal device calculates based on the parametric image features corresponding to the fully connected layer in the image classification model, and predicts the image classification results corresponding to the image features.

In S103, the image classification result is output.

In S103, the terminal device outputs the image classification result predicted by the image classification model, so that the user can obtain the corresponding image classification result.

It can be seen from the above that the terminal device acquires the target image to be classified; based on the optimal parameters in the image classification model, the target image is subjected to feature extraction to obtain image features, and the image features are subjected to classification prediction processing to obtain image classification results, The optimal parameter is obtained based on a preset noise value when the second norm of the loss function of the image classification model is less than a first preset value, and the preset noise value is used to classify the trained image The model parameters determined by the model avoid the saddle point during iterative optimization, so that the terminal device can extract the feature of the target image based on the optimal parameters in the image classification model to obtain the image feature, and more accurately extract the image feature corresponding to the target image ; When the terminal device classifies and predicts the image features based on the optimal parameters in the image classification model to obtain the image classification result, the predicted image classification result will also be more accurate.

Referring to FIG. 2, FIG. 2 is an implementation flowchart of the image classification method provided by the second embodiment of the present application. The difference between this embodiment and the first embodiment is that in this embodiment, after S201 and before S202, S2011-S2014 are further included. S201-S204 are the same as S101-S104 in the first embodiment. For details, please refer to the relevant descriptions of S101-S104 in the first embodiment, which will not be repeated here. S2011 ～ S2014 are as follows:

S2011. Determine the first gradient corresponding to the first loss function value according to the first loss function value corresponding to the image classification model trained in the current iteration, and determine the two corresponding to the first gradient according to the first gradient. Norm.

The image classification model needs to be trained to perform the image classification function, and the process of training the image classification model is the process of iterative optimization of the model parameters of the image classification model, so that the model parameters of the image classification model can be optimized . When iteratively optimizing the model parameters in the image classification model, the terminal device determines the first gradient corresponding to the first loss function value according to the first loss function value corresponding to the image classification model under the current iteration optimization times of the image classification model, and The second norm corresponding to the first gradient is determined according to the first gradient. Among them, the first loss function value is the loss function value calculated by the loss function in the current iteration optimization times, and the gradient is used to represent the parameter vector corresponding to the loss function that changes the fastest and has the largest change rate during the current iteration optimization, the first gradient In order to obtain the gradient value corresponding to the first loss function value, the terminal device will also determine and obtain the second norm corresponding to the first gradient according to the first gradient.

S2012: Determine whether the second norm is less than the first preset value.

Since the loss function has a saddle point, and the saddle point is the local minimum value of the loss function, in the prior art, the terminal device cannot distinguish whether the loss function is a local minimum value or a global minimum value, resulting in the image classification model unable to converge to The situation of the global extreme point. In this embodiment, when at the saddle point, the gradient vector corresponding to the corresponding loss function is zero, and the second norm of the corresponding gradient vector is also zero. Therefore, the terminal device determines whether the second norm corresponding to the first gradient is less than the first A preset value to determine whether the loss function reaches the saddle point, where the first preset value is a preset value.

S2013, if the second norm is less than the first preset value, add the preset noise value to the first model parameter determined by the image classification model trained in the current iteration, the preset noise value is used Therefore, the model parameters determined by the trained image classification model avoid corresponding saddle points when iterative optimization is performed.

When the second norm corresponding to the first gradient is less than the first preset value, it means that the loss function reaches the saddle point; when the second norm corresponding to the first gradient is greater than or the first preset value, it means that the loss function has not reached the saddle point Office. When the second norm corresponding to the first gradient is less than the first preset value, the preset noise value is added to the first model parameter determined by the image classification model trained in the current iteration. The preset noise value is used to make the The model parameters determined by the trained image classification model bring disturbance effects when iterative optimization is performed, so that the model parameters determined by the trained image classification model can avoid the saddle point during iterative optimization, and the preset noise value is for the image When the model parameters in the classification model are iteratively optimized, they are obtained by random sampling in the sample library of model parameters. Adding noise values to the model parameters determined by the image classification model can avoid stopping at the saddle point when iteratively optimizing the image classification model, so as to avoid that the terminal device will directly use the corresponding model parameters when converging to the local minimum as image classification The optimal parameters of the model.

S2014, if the difference between the second loss function value corresponding to the image classification model trained in the target iteration after the current iteration and the first loss function value corresponding to the image classification model trained in the current iteration is less than the second preset Value, where the second preset value is generally a constant close to zero, it is determined that the image classification model has converged to the global extremum point during training, and the second model parameter determined in the target iteration is output as the trained image The optimal parameters of the classification model will also complete the training of the image classification model, and the corresponding image classification model can be used to perform the corresponding image classification function.

If in a target iteration after the current iteration, the terminal device determines whether the difference between the value of the second loss function corresponding to the image classification model trained and the value of the first loss function corresponding to the image classification model trained in the current iteration Less than the second preset value, if the difference between the second loss function value corresponding to the trained image classification model and the first loss function value corresponding to the image classification model trained in the current iteration is less than the second preset value For the terminal device, the terminal device determines that the image classification model has converged to the global extremum point during training, and outputs the second model parameter determined in the target iteration as the optimal parameter of the trained image classification model. The terminal device takes the corresponding model parameter when the image classification model converges to the global minimum as the optimal parameter, so that the terminal device can extract the feature of the target image based on the optimal parameter in the image classification model to obtain the image feature more accurately To the image feature corresponding to the target image; when the terminal device classifies and predicts the image feature based on the optimal parameters in the image classification model to obtain the image classification result, the predicted image classification result will be more accurate.

Optionally, in this embodiment, in order to determine the first preset value more accurately, so that the terminal device can accurately determine whether the loss function reaches the saddle point, the calculation method of the first preset value is specifically: Terminal equipment according to preset calculation formula

as well as

The first preset value is calculated. Where g is a preset first preset value, d is the number of corresponding model parameters in the trained image classification model, c, δ, and ∈ are preset constants, and l is a Lipschitz continuous constant, Δf is the gradient function corresponding to the loss function of the trained image classification model.

Optionally, before adding the preset noise value to the first model parameter determined by the image classification model trained in the current iteration, if the second norm is less than the first preset value, including:

It is determined whether the number of iterations without adding a preset noise value to the model parameters determined by the image classification model trained before the current iteration reaches a third preset value.

If the second norm is less than the first preset value, adding the preset noise value to the first model parameter determined by the image classification model trained in the current iteration includes:

If the number of iterations without adding a preset noise value to the model parameters determined by the image classification model trained before the current iteration reaches a third preset value, and the second norm is less than the first preset value, add the preset noise Value into the first model parameter determined by the image classification model trained in the current iteration.

When the second norm corresponding to the first gradient is less than the first preset value, before adding the preset noise value to the first model parameter determined by the image classification model trained in the current iteration, the terminal device also determines Before the current iteration, whether the number of iterations without the preset noise value added to the model parameters determined by the trained image classification model reaches the third preset value, where the third preset value is a positive integer, if the During the iterative optimization process with three preset values, and the corresponding second norm is less than the first preset value, the preset noise value is added to the first model parameter determined by the image classification model trained in the current iteration, so that The terminal equipment can accurately determine whether the loss function reaches the saddle point.

Preferably, the method for calculating the third preset value includes:

According to the preset calculation formula

as well as

The third preset value is calculated, where k is the third preset value, d is the number of corresponding model parameters in the trained image classification model, c, ρ, δ, and ε are preset constants, and l is the profit Pushitz continuous constant, Δf is the gradient function corresponding to the loss function of the trained image classification model.

Terminal equipment according to preset calculation formula

as well as

The third preset value is calculated, where k is the third preset value, d is the number of corresponding model parameters in the trained image classification model, c, ρ, δ, and ∈ are preset constants, and l is Lipschitz continuous constant, Δf is the gradient function corresponding to the loss function of the trained image classification model. It should be noted that, when the third preset value k is not a positive integer, the terminal device will select a positive integer with the smallest difference from the third preset value k to round the third preset value k.

Refer to FIG. 3, which is a schematic diagram of a terminal device according to a third embodiment of the present application. Each unit included in the terminal device is used to execute each step in the embodiment corresponding to FIG. 1 or FIG. 2. For details, please refer to the related descriptions in the embodiments corresponding to FIG. 1 or FIG. 2. For ease of explanation, only parts related to this embodiment are shown. Referring to FIG. 3, the terminal equipment includes:

The obtaining unit 101 is used to obtain a target image to be classified.

The execution unit 102 is configured to perform feature extraction on the target image based on the optimal parameters in the image classification model to obtain image features, and perform classification prediction processing on the image features to obtain an image classification result, where the optimal parameters are When the second norm of the loss function of the image classification model is less than the first preset value, it is obtained based on a preset noise value, and the preset noise value is used to iterate the model parameters determined by the trained image classification model Avoid the saddle point when optimizing.

The output unit 103 is configured to output the image classification result.

Optionally, the terminal device further includes:

A determining unit, configured to determine a first gradient corresponding to the first loss function value according to the first loss function value corresponding to the image classification model trained in the current iteration, and determine the first gradient according to the first gradient Corresponding second norm.

The judging unit is used to judge whether the second norm is less than the first preset value.

An adding unit, configured to add a preset noise value to the first model parameter determined by the image classification model trained in the current iteration if the second norm is less than the first preset value, the preset noise value Used to make the model parameters determined by the trained image classification model avoid the saddle point during iterative optimization.

The determining unit is used if the difference between the second loss function value corresponding to the image classification model trained in the target iteration after the current iteration and the first loss function value corresponding to the image classification model trained in the current iteration is less than the first Two preset values, it is determined that the image classification model has converged to the global extremum point during training, and the second model parameters determined in the target iteration are output as the optimal parameters of the trained image classification model.

Optionally, the determining unit is also used to:

According to the preset calculation formula

as well as

The first preset value is calculated, where g is the first preset value, d is the number of corresponding model parameters in the trained image classification model, c, δ, and ∈ are preset constants, and l is Lipsch Tz continuous constant, Δf is the gradient function corresponding to the loss function of the trained image classification model.

Optionally, the terminal device further includes:

The judging unit is further configured to judge whether the number of iterations without adding a preset noise value to the model parameters determined by the image classification model trained before the current iteration reaches a third preset value;

The adding unit is specifically configured to: if the model parameter determined by the image classification model trained before the current iteration does not add a preset noise value, the number of iterations reaches a third preset value, and the second norm is less than the first preset Value, the preset noise value is added to the first model parameter determined by the image classification model trained in the current iteration.

Optionally, the determining unit is also used to:

According to the preset calculation formula

as well as

It can be seen from the above that the terminal device acquires the target image to be classified; based on the optimal parameters in the image classification model, the target image is subjected to feature extraction to obtain image features, and the image features are subjected to classification prediction processing to obtain image classification results, The optimal parameter is obtained based on a preset noise value when the second norm of the loss function of the image classification model is less than a first preset value, and the preset noise value is used to classify the trained image The model parameters determined by the model avoid the saddle point during iterative optimization, so that the terminal device can extract the feature of the target image based on the optimal parameters in the image classification model to obtain the image feature, and can more accurately extract the image feature corresponding to the target image ; When the terminal device classifies and predicts the image features based on the optimal parameters in the image classification model to obtain the image classification result, the predicted image classification result will also be more accurate.

Refer to FIG. 4, which is a schematic diagram of a terminal device according to a fourth embodiment of the present application. As shown in FIG. 4, the terminal device 4 of this embodiment includes: a processor 40, a memory 41, and computer-readable instructions 42 stored in the memory 41 and executable on the processor 40, such as the terminal device ’s control program. When the processor 40 executes the computer-readable instruction 42, the steps in the above embodiments of the image classification method of each terminal device 4 are implemented, for example, S101 to S103 shown in FIG. 1. Alternatively, when the processor 40 executes the computer-readable instructions 42, the functions of the units in the foregoing device embodiments are realized, for example, the functions of the units 101 to 103 shown in FIG. 3.

Exemplarily, the computer-readable instructions 42 may be divided into one or more units, and the one or more units are stored in the memory 41 and executed by the processor 40 to complete the application . The one or more units may be an instruction segment of a series of computer-readable instructions capable of performing a specific function. The instruction segment is used to describe the execution process of the computer-readable instruction 42 in the terminal device 4. For example, the computer-readable instructions 42 may be divided into an acquisition unit, an execution unit, and an output unit, and the specific functions of each unit are as described above.

The terminal device may include, but is not limited to, the processor 40 and the memory 41. Those skilled in the art may understand that FIG. 4 is only an example of the terminal device 4 and does not constitute a limitation on the terminal device 4, and may include more or fewer components than the illustration, or a combination of certain components, or different components. For example, the terminal device may further include an input and output device, a network access device, a bus, and the like.

The so-called processor 40 may be a central processing unit (Central Processing Unit, CPU), or other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), Ready-made programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may be any conventional processor or the like.

The memory 41 may be an internal storage unit of the terminal device 4, such as a hard disk or a memory of the terminal device 4. The memory 41 may also be an external storage terminal device of the terminal device 4, such as a plug-in hard disk equipped on the terminal device 4, a smart memory card (Smart, Media, Card, SMC), and secure digital (SD) ) Card, flash card (Flash Card), etc. Further, the memory 41 may include both an internal storage unit of the terminal device 4 and an external storage terminal device. The memory 41 is used to store the computer-readable instructions and other programs and data required by the terminal device. The memory 41 can also be used to temporarily store data that has been or will be output.

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, those of ordinary skill in the art should understand that they can still implement the foregoing The technical solutions described in the examples are modified, or some of the technical features are equivalently replaced; and these modifications or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present application. Within the scope of protection of this application.

Claims

An image classification method, characterized in that it includes:

Obtain the target image to be classified;

Based on the optimal parameters in the image classification model, the target image is subjected to feature extraction to obtain image features, and the image features are subjected to classification prediction processing to obtain image classification results, wherein the optimal parameters are in the image classification model When the second norm of the loss function is less than the first preset value, it is obtained based on a preset noise value, which is used to make the model parameters determined by the trained image classification model avoid the saddle point during iterative optimization;

The image classification result is output.
The image classification method according to claim 1, wherein after acquiring the target image to be classified, the target image is subjected to feature extraction based on optimal parameters in the image classification model to obtain image features Previously, the image classification method also included:

Determine the first gradient corresponding to the first loss function value according to the first loss function value corresponding to the image classification model trained in the current iteration, and determine the second norm corresponding to the first gradient according to the first gradient ;

Determine whether the second norm is less than the first preset value;

If the second norm is less than the first preset value, the preset noise value is added to the first model parameter determined by the image classification model trained in the current iteration, and the preset noise value is used to make The model parameters determined by the trained image classification model avoid the saddle point during iterative optimization;

If the difference between the second loss function value corresponding to the image classification model trained in the target iteration after the current iteration and the first loss function value corresponding to the image classification model trained in the current iteration is less than the second preset value, Then, it is determined that the image classification model has converged to the global extremum point during training, and the second model parameter determined in the target iteration is output as the optimal parameter of the trained image classification model.
The image classification method according to claim 2, wherein the method for calculating the first preset value includes:

According to the preset calculation formula
as well as
The first preset value is calculated, where g is the first preset value, d is the number of corresponding model parameters in the trained image classification model, c, δ, and ∈ are preset constants, and l is Lipsch Tz continuous constant, Δf is the gradient function corresponding to the loss function of the trained image classification model.
The image classification method according to claim 2, wherein if the second norm is less than the first preset value, the preset noise value is added to the first determination determined by the image classification model trained in the current iteration Before the model parameters, including:

Determine whether the number of iterations without adding a preset noise value to the model parameters determined by the image classification model trained before the current iteration reaches the third preset value;

If the second norm is less than the first preset value, adding the preset noise value to the first model parameter determined by the image classification model trained in the current iteration includes:

If the number of iterations without adding a preset noise value to the model parameters determined by the image classification model trained before the current iteration reaches a third preset value, and the second norm is less than the first preset value, add the preset noise Value into the first model parameter determined by the image classification model trained in the current iteration.
The image classification method according to claim 4, wherein the method for calculating the third preset value includes:

According to the preset calculation formula
as well as
The third preset value is calculated, where k is the third preset value, d is the number of corresponding model parameters in the trained image classification model, c, ρ, δ, and ε are preset constants, and l is the profit Pushitz continuous constant, Δf is the gradient function corresponding to the loss function of the trained image classification model.
A terminal device is characterized by comprising:

An obtaining unit, used to obtain the target image to be classified;

The execution unit is configured to perform feature extraction on the target image based on the optimal parameters in the image classification model to obtain image features, and perform classification prediction processing on the image features to obtain image classification results, where the optimal parameters are in all When the second norm of the loss function of the image classification model is less than the first preset value, it is obtained based on a preset noise value, and the preset noise value is used to make iterative optimization of the model parameters determined by the trained image classification model Avoid the saddle point

The output unit is used to output the image classification result.
The terminal device according to claim 6, further comprising:

A determining unit, configured to determine a first gradient corresponding to the first loss function value according to the first loss function value corresponding to the image classification model trained in the current iteration, and determine the first gradient according to the first gradient Corresponding second norm;

The judging unit is used to judge whether the second norm is less than the first preset value;

An adding unit, configured to add a preset noise value to the first model parameter determined by the image classification model trained in the current iteration if the second norm is less than the first preset value, the preset noise value Used to make the model parameters determined by the trained image classification model avoid the saddle point during iterative optimization;

The determining unit is used if the difference between the second loss function value corresponding to the image classification model trained in the target iteration after the current iteration and the first loss function value corresponding to the image classification model trained in the current iteration is less than the first Two preset values, it is determined that the image classification model has converged to the global extremum point during training, and the second model parameters determined in the target iteration are output as the optimal parameters of the trained image classification model.
The terminal device according to claim 7, wherein the determination unit is further configured to: according to a preset calculation formula
as well as
The first preset value is calculated, where g is the first preset value, d is the number of corresponding model parameters in the trained image classification model, c, δ, and ∈ are preset constants, and l is Lipsch Tz continuous constant, Δf is the gradient function corresponding to the loss function of the trained image classification model.
The terminal device according to claim 7, wherein:

The judging unit is also used to judge whether the number of iterations without adding a preset noise value to the model parameters determined by the image classification model trained before the current iteration reaches a third preset value;

The adding unit is specifically configured to: if the model parameter determined by the image classification model trained before the current iteration does not add a preset noise value, the number of iterations reaches a third preset value, and the second norm is less than the first preset Value, the preset noise value is added to the first model parameter determined by the image classification model trained in the current iteration.
The terminal device according to claim 9, wherein the determining unit is further configured to:

According to the preset calculation formula
as well as
The third preset value is calculated, where k is the third preset value, d is the number of corresponding model parameters in the trained image classification model, c, ρ, δ, and ∈ are preset constants, and l is Pushitz continuous constant, Δf is the gradient function corresponding to the loss function of the trained image classification model.
A terminal device, characterized in that the terminal device includes a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, and the processor executes the computer-readable instructions The following steps are implemented when instructing:

Obtain the target image to be classified;

Based on the optimal parameters in the image classification model, the target image is subjected to feature extraction to obtain image features, and the image features are subjected to classification prediction processing to obtain image classification results, wherein the optimal parameters are in the image classification model When the second norm of the loss function is less than the first preset value, it is obtained based on a preset noise value, which is used to make the model parameters determined by the trained image classification model avoid the saddle point during iterative optimization;

The image classification result is output.
The terminal device according to claim 11, wherein after acquiring the target image to be classified, before performing feature extraction on the target image based on the optimal parameters in the image classification model to obtain image features , The processor also implements the following steps when executing the computer-readable instructions:

Determine the first gradient corresponding to the first loss function value according to the first loss function value corresponding to the image classification model trained in the current iteration, and determine the second norm corresponding to the first gradient according to the first gradient ;

Determine whether the second norm is less than the first preset value;

If the second norm is less than the first preset value, the preset noise value is added to the first model parameter determined by the image classification model trained in the current iteration, and the preset noise value is used to make The model parameters determined by the trained image classification model avoid the saddle point during iterative optimization;

If the difference between the second loss function value corresponding to the image classification model trained in the target iteration after the current iteration and the first loss function value corresponding to the image classification model trained in the current iteration is less than the second preset value, Then, it is determined that the image classification model has converged to the global extremum point during training, and the second model parameter determined in the target iteration is output as the optimal parameter of the trained image classification model.
The terminal device according to claim 11, wherein before determining whether the second norm is less than a first preset value, the processor further implements the following steps when executing the computer-readable instruction:

According to the preset calculation formula
as well as
The first preset value is calculated, where g is the first preset value, d is the number of corresponding model parameters in the trained image classification model, c, δ, and ∈ are preset constants, and l is Lipsch Tz continuous constant, Δf is the gradient function corresponding to the loss function of the trained image classification model.
The terminal device according to claim 12, wherein if the second norm is less than a first preset value, a preset noise value is added to the first determined by the image classification model trained in the current iteration Before the model parameters, the processor also implements the following steps when executing the computer-readable instructions:

Determine whether the number of iterations without adding a preset noise value to the model parameters determined by the image classification model trained before the current iteration reaches the third preset value;

If the second norm is less than the first preset value, adding the preset noise value to the first model parameter determined by the image classification model trained in the current iteration includes:

If the number of iterations without adding a preset noise value to the model parameters determined by the image classification model trained before the current iteration reaches a third preset value, and the second norm is less than the first preset value, add the preset noise Value into the first model parameter determined by the image classification model trained in the current iteration.
The terminal device according to claim 14, wherein if the model parameters determined by the image classification model trained before the current iteration do not add a preset noise value, the number of iterations reaches a third preset value, and the If the second norm is less than the first preset value, before the preset noise value is added to the first model parameter determined by the image classification model trained in the current iteration, the processor also implements the computer-readable instruction The following steps:

According to the preset calculation formula
as well as
The third preset value is calculated, where k is the third preset value, d is the number of corresponding model parameters in the trained image classification model, c, ρ, δ, and ∈ are preset constants, and l is Pushitz continuous constant, Δf is the gradient function corresponding to the loss function of the trained image classification model.
A computer nonvolatile readable storage medium, the computer nonvolatile readable storage medium storing computer readable instructions, characterized in that, when the computer readable instructions are executed by at least one processor, the following steps are realized :

Obtain the target image to be classified;

Based on the optimal parameters in the image classification model, the target image is subjected to feature extraction to obtain image features, and the image features are subjected to classification prediction processing to obtain image classification results, wherein the optimal parameters are in the image classification model When the second norm of the loss function is less than the first preset value, it is obtained based on a preset noise value, which is used to make the model parameters determined by the trained image classification model avoid the saddle point during iterative optimization;

The image classification result is output.
The computer non-volatile storage medium according to claim 15, wherein after acquiring the target image to be classified, the target image is based on the optimal parameters in the image classification model Before performing feature extraction to obtain image features, the image classification method further includes:

Determine the first gradient corresponding to the first loss function value according to the first loss function value corresponding to the image classification model trained in the current iteration, and determine the second norm corresponding to the first gradient according to the first gradient ;

Determine whether the second norm is less than the first preset value;

If the second norm is less than the first preset value, the preset noise value is added to the first model parameter determined by the image classification model trained in the current iteration, and the preset noise value is used to make The model parameters determined by the trained image classification model avoid the saddle point during iterative optimization;

If the difference between the second loss function value corresponding to the image classification model trained in the target iteration after the current iteration and the first loss function value corresponding to the image classification model trained in the current iteration is less than the second preset value, Then, it is determined that the image classification model has converged to the global extremum point during training, and the second model parameter determined in the target iteration is output as the optimal parameter of the trained image classification model.
The computer non-volatile storage medium according to claim 17, wherein the calculation method of the first preset value includes:

According to the preset calculation formula
as well as
The first preset value is calculated, where g is the first preset value, d is the number of corresponding model parameters in the trained image classification model, c, δ, and ∈ are preset constants, and l is Lipsch Tz continuous constant, Δf is the gradient function corresponding to the loss function of the trained image classification model.
The computer non-volatile readable storage medium according to claim 17, wherein, if the second norm is less than the first preset value, a preset noise value is added to the trained in the current iteration Before the first model parameter determined by the image classification model, it includes:

Determine whether the number of iterations without adding a preset noise value to the model parameters determined by the image classification model trained before the current iteration reaches the third preset value;

If the second norm is less than the first preset value, adding the preset noise value to the first model parameter determined by the image classification model trained in the current iteration includes:

If the number of iterations without adding a preset noise value to the model parameters determined by the image classification model trained before the current iteration reaches a third preset value, and the second norm is less than the first preset value, add the preset noise Value into the first model parameter determined by the image classification model trained in the current iteration.
The computer non-volatile readable storage medium according to claim 19, wherein the method for calculating the third preset value includes:

According to the preset calculation formula
as well as
The third preset value is calculated, where k is the third preset value, d is the number of corresponding model parameters in the trained image classification model, c, ρ, δ, and ε are preset constants, and l is the profit Pushitz continuous constant, Δf is the gradient function corresponding to the loss function of the trained image classification model.