CN116978080A

CN116978080A - Information identification method, apparatus and computer readable storage medium

Info

Publication number: CN116978080A
Application number: CN202210406144.1A
Authority: CN
Inventors: 许剑清
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2022-04-18
Filing date: 2022-04-18
Publication date: 2023-10-31

Abstract

The embodiment of the application discloses an information identification method, an information identification device and a computer readable storage medium, which can be applied to various scenes such as cloud technology, artificial intelligence, intelligent traffic, auxiliary driving and the like; obtaining an image sample; extracting image features of the image sample by adopting an identification model, and determining identification information of the image sample based on the extracted image features; acquiring time information corresponding to the image sample, and determining time loss information based on the time information and a time information label; determining the identification loss information according to the identification information and the identification information label, and eliminating the time loss information from the identification loss information to obtain target loss information; and converging the recognition model based on the target loss information to obtain the recognition model after training. Therefore, the recognition model is used for recognizing the image to be recognized according to the trained recognition model, the accuracy of the recognition model for recognizing the face image with different time information is improved, and the information recognition efficiency is further improved.

Description

Information identification method, apparatus and computer readable storage medium

Technical Field

The present application relates to the field of internet technologies, and in particular, to an information identification method, an information identification device, and a computer readable storage medium.

Background

In recent years, with the rapid development of internet technology, artificial intelligence technology is also continuously developing. Among them, the face recognition technology is widely applied to the daily life of people. The popularization of face recognition application scenes requires that a face recognition model has robustness to each scene, for example, a scene in which face images of the same identity in different time information are recognized. In the existing face recognition models, the face recognition models are mostly trained in a multitasking mode, and weights of loss functions among image features of different time information are adjusted in the training process, so that the accuracy of face recognition is improved.

In the research and practice process of the prior art, the method for training the face recognition model in a multitasking mode to improve the accuracy of face recognition in the prior art increases the model volume, so that the training time is relatively long, meanwhile, the accuracy of the face recognition model on face recognition of different time information cannot be stably improved, and further the information recognition efficiency is relatively low.

Disclosure of Invention

The embodiment of the application provides an information identification method, an information identification device and a computer readable storage medium, which can improve the accuracy of identifying face images of different time information by an identification model, thereby improving the efficiency of information identification.

The embodiment of the application provides an information identification method, which comprises the following steps:

acquiring an image sample, wherein the image sample carries an identification information tag and a time information tag;

extracting image features of the image sample by adopting an identification model, and determining identification information of the image sample based on the extracted image features;

acquiring time information corresponding to the image sample, and determining time loss information corresponding to the image sample based on the time information and a time information label, wherein the time information is obtained by time classifying the image features by adopting a classification model;

determining the identification loss information corresponding to the image sample according to the identification information and the identification information label, and eliminating the time loss information from the identification loss information to obtain target loss information;

and converging the recognition model based on the target loss information to obtain a trained recognition model, wherein the trained recognition model is used for recognizing the image to be recognized.

Accordingly, an embodiment of the present application provides an information identifying apparatus, including:

the sample acquisition unit is used for acquiring an image sample, wherein the image sample carries an identification information tag and a time information tag;

The characteristic extraction unit is used for extracting image characteristics of the image sample by adopting an identification model and determining identification information of the image sample based on the extracted image characteristics;

the determining unit is used for obtaining time information corresponding to the image sample, determining time loss information corresponding to the image sample based on the time information and a time information label, wherein the time information is obtained by time classifying the image features by adopting a classification model;

the rejecting unit is used for determining the identification loss information corresponding to the image sample according to the identification information and the identification information label, and rejecting the time loss information from the identification loss information to obtain target loss information;

and the convergence unit is used for converging the recognition model based on the target loss information to obtain a trained recognition model, and the trained recognition model is used for recognizing the image to be recognized.

In an embodiment, the convergence unit includes:

a convergence condition determining subunit, configured to determine a convergence condition of the target loss information based on model information corresponding to the identification model and the classification model;

And the convergence subunit is used for converging the recognition model according to the time loss information and the identification loss information when the target loss information does not meet the convergence condition, so as to obtain a recognition model after training.

In an embodiment, the convergence subunit comprises:

the gradient updating module is used for updating the time gradient of the classification model according to the time loss information to obtain an updated time gradient, and updating the identification gradient of the identification model according to the identification loss information to obtain an updated identification gradient;

and the updating convergence module is used for converging the recognition model based on the updated time gradient and the updated identification gradient to obtain a trained recognition model.

In an embodiment, the update convergence module includes:

the gradient update times acquisition sub-module is used for acquiring the current gradient update times;

the gradient adjustment sub-module is used for adjusting the updated mark gradient based on the gradient updating times and the updated time gradient to obtain an adjusted mark gradient;

and the parameter updating sub-module is used for updating the network parameters of the classification model according to the updated time gradient and updating the network parameters of the recognition model according to the adjusted identification gradient so as to obtain a trained recognition model.

In an embodiment, the gradient adjustment submodule is configured to:

screening out a target time gradient from the updated time gradients, wherein the target time gradient is used for adjusting the updated identification gradient in the updated time gradient;

determining a gradient adjustment parameter of the target time gradient according to the gradient updating times, and weighting the gradient adjustment parameter and the target time gradient to obtain a weighted target time gradient;

and fusing the weighted target time gradient and the updated mark gradient to obtain an adjusted mark gradient.

In an embodiment, the gradient adjustment submodule may be specifically configured to:

calculating the ratio of the gradient update times to preset update parameters to obtain an update frequency ratio;

when the update frequency ratio is of a preset numerical value type, calculating the ratio of the gradient update times to the preset total update times to obtain an update times ratio;

and converting the ratio of the update times into an initial gradient adjustment parameter of the target time gradient, and weighting the initial gradient adjustment parameter according to a preset negative adjustment coefficient to obtain the gradient adjustment parameter.

In an embodiment, the parameter updating sub-module may be specifically configured to:

updating the network parameters of the identification model according to the adjusted identification gradient to obtain an updated identification model, and updating the network parameters of the classification model according to the updated time gradient to obtain an updated classification model;

updating the target loss information by the updated identification model and the updated identification model to obtain updated target loss information;

and converging the updated recognition model based on the updated target loss information to obtain a trained recognition model.

when the updated target loss information meets the convergence condition, the updated identification model is used as the trained identification model;

and when the updated target loss information does not meet the convergence condition, taking the updated target loss information as target loss information, taking an updated identification model as an identification model, taking an updated classification model as a classification model, and returning to execute the step of updating the time gradient of the classification model according to the time loss information to obtain an updated time gradient, updating the identification gradient of the identification model according to the identification loss information to obtain an updated identification gradient until the target loss information meets the convergence condition, and obtaining the trained identification model.

In an embodiment, the information identifying apparatus further includes:

the initial sample acquisition unit is used for acquiring an initial image sample, wherein the initial image sample carries an initial time information label;

the initial feature extraction unit is used for extracting features of the initial image sample by adopting an identification model to obtain initial image features;

the initial classification unit is used for classifying the initial image features by adopting a preset classification model to obtain initial time information corresponding to the initial image features;

and the initial convergence unit is used for converging the preset classification model based on the initial time information and the initial time information label to obtain a classification model.

In an embodiment, the information identifying apparatus further includes:

the image feature extraction subunit to be identified is used for obtaining an image to be identified, and extracting features of the image to be identified by adopting the trained identification model to obtain image features to be identified corresponding to the image to be identified;

the image feature extraction subunit to be compared is used for obtaining an image to be compared, and extracting features of the image to be compared by adopting the recognition model after training to obtain features of the image to be compared corresponding to the image to be compared;

And the recognition subunit is used for recognizing the matching coefficient between the image feature to be recognized and the image feature to be compared based on the recognition model after training, and determining the matching result between the image to be recognized and the image to be compared according to the matching coefficient.

In addition, the embodiment of the application also provides a computer readable storage medium, which stores a plurality of instructions, wherein the instructions are suitable for being loaded by a processor to execute the steps in any information identification method provided by the embodiment of the application.

In addition, the embodiment of the application also provides a computer device, which comprises a processor and a memory, wherein the memory stores an application program, and the processor is used for running the application program in the memory to realize the information identification method provided by the embodiment of the application.

Embodiments of the present application also provide a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device executes the steps in the information identification method provided by the embodiment of the application.

The embodiment of the application obtains the image sample; extracting image features of the image sample by adopting an identification model, and determining identification information of the image sample based on the extracted image features; acquiring time information corresponding to the image sample, and determining time loss information corresponding to the image sample based on the time information and a time information label, wherein the time information is obtained by time classifying the image characteristics by adopting a classification model; determining the identification loss information corresponding to the image sample according to the identification information and the identification information label, and eliminating the time loss information from the identification loss information to obtain target loss information; and converging the recognition model based on the target loss information to obtain the recognition model after training. According to the method, time loss information is calculated by acquiring time information obtained by time classification of image features extracted by the recognition model through the classification model and time information labels corresponding to the image samples, the time loss information is calculated according to identification information and identification information labels determined by the recognition model based on the extracted image features, then the time loss information is removed from the identification loss information to obtain target loss information so as to train the recognition model, the recognition model obtained through training does not contain features of recognition time information in the image features extracted by the image samples, and therefore common features except the time information features can be extracted from the images to be processed so as to recognize the identification information of the images to be processed.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings that are needed in the description of the embodiments will be briefly described below, it being obvious that the drawings in the following description are only some embodiments of the present application, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

Fig. 1 is a schematic diagram of an implementation scenario of an information identification method according to an embodiment of the present application;

fig. 2 is a schematic flow chart of an information identification method according to an embodiment of the present application;

FIG. 3a is a schematic diagram of an identification model training process of an information identification method according to an embodiment of the present application;

FIG. 3b is a schematic diagram of a training flow of a classification model of an information identification method according to an embodiment of the present application;

fig. 4 is a specific flow diagram of an information identification method according to an embodiment of the present application;

fig. 5 is an overall flow diagram of an information identification method according to an embodiment of the present application;

FIG. 6 is another flow chart of an information identification method according to an embodiment of the present application;

fig. 7 is a schematic structural diagram of an information identifying apparatus according to an embodiment of the present application;

Fig. 8 is a schematic structural diagram of a computer device according to an embodiment of the present application.

Detailed Description

The following description of the embodiments of the present application will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present application, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to fall within the scope of the application.

The embodiment of the application provides an information identification method, an information identification device and a computer readable storage medium. The information identifying apparatus may be integrated in a computer device, and the computer device may be a server or a terminal.

The server may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, network acceleration services (Content Delivery Network, CDN), basic cloud computing services such as big data and an artificial intelligent platform. Terminals may include, but are not limited to, cell phones, computers, intelligent voice interaction devices, intelligent appliances, vehicle terminals, aircraft, and the like. The terminal and the server may be directly or indirectly connected through wired or wireless communication, and the present application is not limited herein.

Referring to fig. 1, taking an example that an information recognition device is integrated in a computer device, fig. 1 is a schematic diagram of an implementation scenario of an information recognition method provided by an embodiment of the present application, where the computer device may be a server or a terminal, and the computer device may obtain an image sample; extracting image features of the image sample by adopting an identification model, and determining identification information of the image sample based on the extracted image features; acquiring time information corresponding to the image sample, and determining time loss information corresponding to the image sample based on the time information and the time information label; determining the identification loss information corresponding to the image sample according to the identification information and the identification information label, and eliminating the time loss information from the identification loss information to obtain target loss information; and converging the recognition model based on the target loss information to obtain the recognition model after training.

It should be noted that the embodiments of the present application may be applied to various scenarios, including, but not limited to, cloud technology, artificial intelligence, intelligent transportation, driving assistance, and the like. The schematic view of the implementation environment of the information identifying method shown in fig. 1 is only an example, and the implementation environment of the information identifying method described in the embodiment of the present application is for more clearly describing the technical solution of the embodiment of the present application, and does not constitute a limitation on the technical solution provided by the embodiment of the present application. As can be known by those skilled in the art, with the evolution of information identification and the appearance of new service scenarios, the technical scheme provided by the application is also applicable to similar technical problems.

The scheme provided by the embodiment of the application relates to the technology of computer vision and the like of artificial intelligence, and is specifically described by the following embodiment. The following description of the embodiments is not intended to limit the preferred embodiments.

The present embodiment will be described from the viewpoint of an information identifying apparatus, which may be integrated in a computer device, which may be a server, to which the present application is not limited.

Referring to fig. 2, fig. 2 is a flowchart illustrating an information identifying method according to an embodiment of the application. The information identification method comprises the following steps:

in step 101, an image sample is acquired.

The image sample may be a sample of a human face image, and the image sample may carry an identification information tag and a time information tag, and optionally, the identification information tag and the time information tag may also be information contained in the image sample. The identification information tag may be information, carried by the image sample, of labeling an identity tag corresponding to the image sample, for indicating an identity of an object in the image sample, the time information tag may be information, carried by the image sample, of labeling a tag of time information corresponding to the image sample, for indicating time information of the object in the image sample, and the time information may be information indicating an age or an age stage of the object in the image sample.

The image sample may be acquired in various ways, for example, from a memory connected to the information recognition device, or from another data storage terminal. Or may be obtained from a memory of the entity terminal, or may be obtained from a virtual memory space, such as a data set or a picture library. Alternatively, the image samples may be obtained from one storage location or may be obtained from a plurality of storage locations, for example, the image samples may be stored on a blockchain, and the information identifying apparatus obtains the image samples from the blockchain. The information identifying means may be configured to collectively acquire the image samples in a period of time in response to a certain image sample acquisition instruction, or may be configured to continue acquisition of the image samples according to a certain image sample acquisition logic, or the like.

It will be appreciated that in the specific embodiments of the present application, related data such as user information is involved, and when the above embodiments of the present application are applied to specific products or technologies, user permissions or consents need to be obtained, and the collection, use and processing of related data need to comply with related laws and regulations and standards of related countries and regions.

In step 102, an image feature is extracted from the image sample using the recognition model, and identification information of the image sample is determined based on the extracted image feature.

The recognition model may be a trained model for recognizing a face image, the image features may be features representing face information in an image sample, specifically may be features extracted from the image sample by using the recognition model and used for representing the face information in the image sample, the identification information may be information representing an identity of an object in the image sample, and specifically may be identity information of the object in the image sample obtained by recognizing the image sample by using the recognition model.

Before the image features of the image sample are extracted by using the recognition model, a preset recognition model may be trained to obtain the recognition model, the preset recognition model may be a preset untrained model, and the preset recognition model may be trained by using the face image sample to obtain a trained recognition model, for example, please refer to fig. 3a, fig. 3a is a schematic diagram of a recognition model training flow of the information recognition method provided by the embodiment of the present application, where a training process for the preset recognition model may include the following modules:

(1) Training data preparation module: in the training process, the face image sample is read, and the read data are combined into a batch (batch) to be sent into a depth network unit of a preset recognition model for processing.

Optionally, the obtained face image sample may be preprocessed, and the preprocessing mode may be various, for example, the size of the face image sample may be adjusted to a preset size, so as to better adapt to the processing of the preset recognition model.

(2) Identifying a model unit module: in the module, the spatial features of the face image sample are extracted, and the spatial structure information of the face image is reserved by the output feature map. The module can be a convolutional neural network (Convolutional Neural Networks, CNN for short) structure, and can comprise convolution (convolution) calculation, nonlinear activation function (Relu) calculation, pooling (Pooling) calculation and other operations.

(3) Face recognition objective function calculation module: in this module, the objective function value may be calculated by taking as input the feature output by the full-connection mapping unit of the recognition model unit module and the tag information of the face image sample that generated the feature. The objective function may employ a classification function, such as a softmax (a classification function) function, as well as other types of objective functions.

(4) Face recognition objective function optimizing module: the module can train and optimize the whole network based on a gradient descent mode, for example, gradient descent methods such as random gradient descent, random gradient descent with a quantity item, adam (a gradient descent update algorithm) and adagard (a gradient descent update algorithm) can be adopted. And (3) repeating the steps (1) to (4) in the training process until the training result meets the training termination condition. The training termination condition for terminating the training of the model can generally set that the iteration times meet a set value, or the loss information calculated by the face recognition objective function is smaller than the set value, so that the training of the preset recognition model can be completed.

Along with the continuous development of artificial intelligence technology, the face recognition technology is widely applied to daily life of people. The popularization of face recognition application scenes requires that a face recognition model has robustness to each scene, for example, a scene for identifying face images of the same identity in different time information, for example, a scene of searching people across ages and comparing identity cards with peer-to-peer scenes, and pictures of the same identity in different ages need to be compared. In a general face recognition model, registered base pictures are general face images of the same age group acquired recently, so that feature space distribution among face images of different age groups is inconsistent during training of the face recognition model, the face recognition model cannot be aligned with feature space distribution among face images of different age groups, and further the trained face recognition model cannot better recognize comparison among face images of different ages. In order to solve the problem, in the existing face recognition model, the face recognition model is mostly trained by adopting a multitask mode, and the weight of a loss function between image features of different time information is adjusted in the training, so that the accuracy of face recognition is improved, however, in the multitask training mode, the weight of the loss function between each task needs to be adjusted, the weight of the loss function is an ultra-parameter, and the optimal value of the loss function changes along with the change of data distribution, so that the accuracy of face recognition across ages cannot be ensured to be stably improved when model training is performed. Meanwhile, the multitask training mode increases the volume of the recognition model and increases the time consumption of model training and calculation.

In order to solve the problems that in the prior art, the accuracy of face recognition of different time information by a face recognition model cannot be stably improved, meanwhile, the model volume is increased by a multitask training method, so that training time is relatively long, and further information recognition efficiency is relatively low. Meanwhile, an additional deployment module is not required to be added, the time-consuming increase of training is avoided, the light weight of the identification model is further guaranteed, and therefore information identification efficiency is improved. The following is a specific description.

In step 103, time information corresponding to the image sample is acquired, and time loss information corresponding to the image sample is determined based on the time information and the time information tag.

The time information is information obtained by time classifying the image features by using a classification model, and may be, for example, information such as an age group to which an image sample corresponding to the image features belongs. The classification model may be a trained network for classifying time information corresponding to image features, for example, may be an age estimation network, and the time loss information may be information representing a gap between time information and a time information tag, that is, an error representing time information corresponding to the image sample predicted by the classification model based on the image features.

The method for obtaining the time information corresponding to the image sample may be various, for example, a classification model may be used to predict the time information of the image feature, and a probability distribution of which time information the image feature belongs to may be output, so that the time information of the image sample may be determined according to the probability distribution.

Optionally, before the time information corresponding to the image sample is obtained, an identification model may be used to train a preset classification model, so as to obtain a trained classification model. The method includes the steps of training a preset classification model by using a recognition model, for example, obtaining an initial image sample, extracting features of the initial image sample by using the recognition model to obtain initial image features, classifying the initial image features by using the preset classification model to obtain initial time information corresponding to the initial image features, and converging the preset classification model based on the initial time information and an initial time information label to obtain the classification model.

The preset classification model may be a preset model which is not trained and is used for predicting time information of image features, and the recognition model may be used for training the capability of the preset recognition model for predicting the time information of the image features so as to obtain a trained recognition model. The initial image sample can be a human face image sample and is used for training a preset classification model, the initial image sample carries an initial time information label, and the initial time information label can be a time information label carried by the initial image sample and is used for labeling time information of an object in the initial image sample. The initial image features may be image features extracted from an initial image sample by using an identification model, and the initial time information may be a result obtained by predicting time information corresponding to the initial image features based on a preset classification model.

In an embodiment, please refer to fig. 3b, fig. 3b is a schematic diagram of a training flow chart of a classification model of an information recognition method according to an embodiment of the present application, where the training flow chart of the classification model shown in fig. 3b may be used to train a preset classification model, and specifically may include the following modules:

S1, training data preparation module: in the training process, the initial image sample is read, the read sample data is combined into a batch, and the batch is sent into a depth network unit of the recognition model for processing.

Optionally, the obtained initial image sample may be preprocessed, and the preprocessing mode may be various, for example, the size of the initial image sample may be adjusted to a preset size, so as to better adapt to the processing of the preset classification model.

S2, a classification model unit module: in this module, the prediction of the time information is performed on the initial image features extracted in the initial image sample by the recognition model unit module. The category labels may be labels of different time information, for example, labels of different age groups. The preset classification model can be a classification network, and the network structure of the classification model can be composed of a convolutional neural network and can comprise operations such as convolutional calculation, nonlinear activation function calculation, pooling calculation and the like. The network structure of the network can also be a fully connected network, and can be adjusted according to the input of a preset classification model. Alternatively, the input of the classification model unit module may be the output of the middle layer of the recognition model, or may be the final output feature of the recognition model. Alternatively, the selection of the access point for the classification model may be optimized by using a network structure search (NAS) method to obtain the best access point.

S3, a cross entropy objective function calculation module: the module can take the initial time information output by the classification model and the initial time information label of the initial image sample as inputs to calculate the loss information of the preset classification model, and the loss function corresponding to the loss information can adopt a cross entropy loss function or a regression type loss function.

S4, a cross entropy objective function optimization module: the module can perform training optimization on the whole network based on a gradient descent mode, for example, gradient descent algorithms such as random gradient descent, random gradient descent with a driving quantity item, adam optimization algorithm, adagard algorithm and the like can be adopted. Steps S1 to S4 may be repeated during the training process until the training result satisfies the training termination condition. The training termination condition for terminating the training of the model can generally set that the iteration times meet a set value, or the loss information calculated by the cross entropy objective function is smaller than the set value, so that the training of the preset classification model can be completed, and the trained classification model is obtained based on the model parameters at the moment.

In step 104, according to the identification information and the identification information label, determining the identification loss information corresponding to the image sample, and eliminating the time loss information from the identification loss information to obtain the target loss information.

The identification loss information may be information representing a gap between the identification information and the identification information label, that is, representing an error of the identification information corresponding to the identification model predicted image sample, and optionally, the identification loss information corresponding to the image sample may be obtained by calculating a loss function of the identification model based on the identification information and the identification information label. The target loss information may be a gap between a characterization recognition model and an optimization target, and the optimization target may be an optimization target for performing time information decoupling training on image features extracted by the recognition model by using a classification model.

In order to improve accuracy of face image recognition comparison between different time information by using a recognition model, time information decoupling training can be performed on image features extracted by using a classification model, for example, please refer to fig. 4, fig. 4 is a specific flow diagram of an information recognition method provided by an embodiment of the present application, and the accuracy of face image recognition comparison between different time information can be improved by connecting a time information decoupling layer between a recognition model unit module and a classification model unit module and performing time information decoupling training on the image features extracted by using the classification model based on the time information decoupling layer, so that features extracted by using the recognition model obtained by training in an image sample do not include relevant features of time information, thereby learning feature distribution after age decoupling. In order to enable features extracted from the image sample by the recognition model obtained through training to not contain relevant features of time information, the time loss information can be removed from the identification loss information to obtain target loss information, so that the target loss information can be used as an optimization target to converge a training process of performing time information decoupling on the recognition model by using the classification model, and the recognition model capable of learning common feature distribution among face images with different time information is obtained.

There are various ways to remove the time loss information from the identification loss information, for example, the time loss information may be subtracted from the identification loss information, so that the target loss information may be obtained. Alternatively, as shown in formula (1), the loss function corresponding to the target loss information may be expressed as

Wherein E () represents a loss function corresponding to target loss information, that is, an optimization target for performing time information decoupling training on image features extracted by a recognition model by using a classification model, and θ _f Representing the characteristics of the image, _y representing identification information tag, θ _d Representing a time information tag, n may represent the total number of samples of image samples in one batch, and i represents the ith image sample in one batch.Representing the loss function of the identification model, i.e. the loss function corresponding to the identification loss information, +.>The loss function of the classification model is represented, and the loss function corresponding to the time loss information is also represented, and lambda can represent the super parameter corresponding to the loss function of the classification model. The purpose of this optimization objective is to enable the image features extracted by the recognition model to preserve the features of the face recognition classification, but not to include the relevant features that distinguish the temporal information. Therein is provided with

Wherein argmin () represents the function of making the targetThe variable value at minimum, argmax () represents the value of the objective function +.>When taking the maximum valueIs used for the variable value of (a).

In step 105, the recognition model is converged based on the target loss information, and a trained recognition model is obtained.

The recognition model after training can be a recognition model after training, in particular can be a recognition model obtained by performing time information decoupling training on image features extracted by the recognition model by adopting a classification model, and can be used for recognizing an image to be recognized. The image to be identified may be a face image for which information identification is not performed.

The method for converging the recognition model based on the target loss information may be various, for example, the convergence condition of the target loss information may be determined based on model information corresponding to the recognition model and the classification model, and when the target loss information does not meet the convergence condition, the recognition model may be converged according to the time loss information and the identification loss information, so as to obtain the recognition model after training.

The model information may be attribute information of the recognition model and the classification model, the attribute information may be information representing properties and relationships of the recognition model and the classification model, for example, may include a loss function of the recognition model and the classification model, the convergence condition may be a termination condition of performing time information decoupling training on image features extracted by the recognition model by using the classification model, for example, may be that the number of iterations is set to satisfy a set value, or the training may be completed by setting the loss information calculated by the objective function to be smaller than the set value, so that the recognition model after training is obtained based on the model parameters at this time.

When the target loss information does not meet the convergence condition, there may be multiple ways of converging the recognition model according to the time loss information and the identification loss information, for example, the time gradient of the classification model may be updated according to the time loss information to obtain an updated time gradient, the identification gradient of the recognition model may be updated according to the identification loss information to obtain an updated identification gradient, and the recognition model may be converged based on the updated time gradient and the updated identification gradient to obtain a trained recognition model.

The time gradient may be a gradient corresponding to the classification model, the updated time gradient may be a gradient obtained by updating the time gradient based on the time loss information, the identification gradient may be a gradient corresponding to the identification model, and the updated identification gradient may be a gradient obtained by updating the identification gradient based on the identification loss information.

The time gradient of the classification model may be updated according to the time loss information in various manners, for example, the gradient of the loss function corresponding to the time loss information may be calculated according to the time loss information, so that the calculated gradient may be updated to obtain an updated time gradient.

Wherein, based on the updated time gradient and the updated identification gradient, various ways of converging the identification model are available, for example, the current gradient update times can be obtained, based on the gradient update times and the updated time gradient, the updated identification gradient is adjusted, and updating the network parameters of the classification model according to the updated time gradient, and updating the network parameters of the recognition model according to the adjusted identification gradient to obtain the trained recognition model.

The number of gradient updating times can be the number of times that the identification model and the classification model are subjected to gradient updating in the decoupling training process of the time information, the adjusted identification gradient can be a result obtained by adjusting the updated identification gradient based on the number of gradient updating times and the updated time gradient, and the network parameters of the identification model are updated based on the adjusted identification gradient, so that the identification model can update parameters in a direction opposite to the updated time gradient of the classification network, and extraction of the characteristics of the time information is continuously weakened.

The method for adjusting the updated marking gradient may be various based on the number of times of gradient update and the updated time gradient, for example, a target time gradient may be selected from the updated time gradients, a gradient adjustment parameter of the target time gradient may be determined according to the number of times of gradient update, the gradient adjustment parameter and the target time gradient may be weighted to obtain a weighted target time gradient, and the weighted target time gradient and the updated marking gradient may be fused to obtain an adjusted marking gradient.

The target time gradient may be a time gradient used for adjusting the updated identification gradient in the updated time gradient, the gradient adjustment parameter may be a parameter for adjusting the target time gradient, and the weighted target time gradient may be a result obtained by weighting the gradient adjustment parameter and the target time gradient.

The method of screening the target time gradient from the updated time gradients may be various, for example, the updated time gradient corresponding to the network layer connected to the identification model in the updated time gradient may be determined as the target time gradient, for example, please continue to refer to fig. 4, and the updated time gradient corresponding to the network layer connected to the information decoupling layer in the classification model of the classification model unit module may be determined as the target time gradient.

After the target time gradient is screened out from the updated time gradients, the gradient adjustment parameters of the target time gradient can be determined according to the gradient update times, wherein the gradient adjustment parameters of the target time gradient can be determined according to the gradient update times in various manners, for example, the ratio of the gradient update times to the preset update parameters can be calculated to obtain an update frequency ratio, when the update frequency ratio is of a preset numerical value type, the ratio of the gradient update times to the preset total update times is calculated to obtain an update times ratio, the update times ratio is converted into the initial gradient adjustment parameters of the target time gradient, and the initial gradient adjustment parameters are weighted according to the preset negative adjustment coefficients to obtain the gradient adjustment parameters.

The ratio of the update frequency may be a ratio of a gradient update frequency to a preset update parameter, where the preset update parameter may be a parameter preset according to an actual requirement and is used to control a frequency of gradient update of the identification model based on the adjusted identification gradient, for example, the preset update parameter may be 3, that is, each 3 times of gradient update of the identification model is indicated, so that the updated identification gradient of the identification model may be adjusted based on an updated time gradient of the identification model, so that the gradient update of the identification model may be performed once based on the adjusted identification gradient. The preset value type may be a preset value type, for example, may be an integer type, that is, when the update frequency ratio is an integer type, the frequency of gradient update performed on the identification model based on the adjusted identification gradient is indicated that the current gradient update frequency accords with the preset update parameter control, so that the ratio of the gradient update frequency to the preset total update frequency may be calculated, the preset total update frequency may be the total iteration frequency of performing time information decoupling training set according to the actual training requirement, the update frequency ratio may be the ratio of the gradient update frequency to the preset total update frequency, and the initial gradient adjustment parameter may be the gradient adjustment parameter of the target time gradient obtained by performing conversion based on the update frequency ratio. The preset negative adjustment coefficient can be a preset adjustment coefficient, and the gradient of the classification model can be reversely acted on the recognition model based on the preset negative adjustment coefficient, so that the features extracted from the image sample by the recognition model obtained through training do not contain the relevant features of the time information, the common feature distribution among the face images of different time information is learned, and the accuracy of face image recognition among the different time information is improved.

The method of converting the ratio of the number of updates to the initial gradient adjustment parameter of the target time gradient may be various, for example, the method of converting the ratio of the number of updates to the initial gradient adjustment parameter of the target time gradient as shown in the formula (2) may be specifically expressed as

Wherein γ may represent an initial gradient adjustment parameter, where the initial gradient adjustment parameter may be a function that varies with the number of iteration steps (i.e. the number of gradient updates), k is the total number of iteration steps, i.e. the preset total number of updates, p is the number of gradient updates, p/k represents the ratio of the number of updates, 3 is the preset update parameter, and may be set to other values according to the actual training requirement, exp () represents an exponential function, if represents a conditional function, and% represents the sign of the remainder.

After the ratio of the update times is converted into the initial gradient adjustment parameter of the target time gradient, the initial gradient adjustment parameter can be weighted according to a preset negative adjustment coefficient to obtain the gradient adjustment parameter. There may be various ways to weight the initial gradient adjustment parameter according to a preset negative adjustment coefficient, for example, the preset negative adjustment coefficient may be-1, and the gradient adjustment parameter may be represented as- γ.

After determining the gradient adjustment parameter of the target time gradient according to the gradient update times, the gradient adjustment parameter and the target time gradient may be weighted. The gradient adjustment parameter and the target time gradient may be weighted in various manners, for example, the gradient adjustment parameter may be multiplied by the target time gradient to obtain a weighted target time gradient of- γi, where I represents the target time gradient.

After weighting the gradient adjustment parameter and the target time gradient, the weighted target time gradient and the updated marker gradient may be fused. The method for fusing the weighted target time gradient and the updated identification gradient may be various, for example, the weighted target time gradient and the updated identification gradient may be superimposed to act on the identification model, so that the identification model performs gradient update based on the updated identification gradient, and then the gradient update based on the weighted target time gradient is superimposed.

For example, please continue to refer to fig. 4, a time information decoupling layer is introduced between the recognition model and the classification model, and the time information decoupling layer has the function of ensuring that when the recognition model is forward, the output of the recognition model is consistent with the input of the classification model, that is, the image features extracted by the recognition model can be indiscriminately input into the classification model, and in addition, when the classification model is subjected to gradient update, the classification model can be returned to the gradient of the recognition model through the time information decoupling layer to adjust, so that the gradient returned to the recognition model by the classification model acts on the recognition model in the opposite direction, and further, the features extracted by the recognition model obtained by training in the image sample do not contain the relevant features of time information, so that the common feature distribution between face images of different time information is learned, and the accuracy of face image recognition comparison between different time information is improved. Alternatively, the mathematical expression of the time information decoupling layer may be as shown in formula (3):

Wherein X can represent the image features extracted by the recognition model, namely the input of the time information decoupling layer, R (X) can represent the image features input by the classification model, namely the output of the time information decoupling layer,the gradient transmitted back to the recognition model through the time learning decoupling layer, namely the adjusted mark gradient, wherein I is the updated time gradient, the negative sign (namely-1) can be a preset negative adjustment coefficient, and gamma can be an initial gradient adjustment parameter.

After the updated identification gradient is adjusted based on the gradient update times and the updated time gradient, the network parameters of the classification model can be updated according to the updated time gradient, and the network parameters of the recognition model are updated according to the adjusted identification gradient to obtain the trained recognition model. The method comprises the steps of updating network parameters of the classification model according to the updated time gradient, updating the network parameters of the identification model according to the adjusted identification gradient to obtain a trained identification model, for example, updating the network parameters of the identification model according to the adjusted identification gradient to obtain an updated identification model, updating the network parameters of the classification model according to the updated time gradient to obtain an updated classification model, updating the target loss information by the updated identification model and the updated identification model to obtain updated target loss information, converging the updated identification model based on the updated target loss information, and obtaining the trained identification model.

The updated identification model may be an identification model obtained by updating network parameters of the identification model according to the adjusted identification gradient, and the updated classification model may be a classification model obtained by updating network parameters of the classification model according to the updated time gradient. The updated target loss information may be target loss information calculated when the time information decoupling training is performed based on the updated recognition model and the updated recognition model.

The method for updating the target loss information by using the updated identification model and the updated identification model may have various ways for obtaining updated target loss information, for example, the updated identification model may be used to extract image features of the image sample, and the updated identification information of the image sample is determined based on the extracted updated image features; classifying the image features by adopting an updated classification model to obtain updated time information of the image sample, and determining updated time loss information corresponding to the image sample based on the updated time information and the updated time information label; and determining updated identification loss information corresponding to the image sample according to the updated identification information and the identification information label, and eliminating updated time loss information from the updated identification loss information to obtain updated target loss information.

The updated image features may be image features obtained by extracting image features from an image sample based on the updated recognition model, and the image sample obtained by the updated recognition model may be the same as or different from the image sample obtained by the recognition model, which is not limited herein.

After the updated recognition model and the updated recognition model update the target loss information to obtain updated target loss information, the updated recognition model may be converged based on the updated target loss information, where there may be various ways of converging the updated recognition model based on the updated target loss information, for example, the updated recognition model may be used as the trained recognition model when the updated target loss information satisfies the convergence condition, the updated target loss information may be used as the target loss information when the updated target loss information does not satisfy the convergence condition, the updated recognition model may be used as the recognition model, the updated classification model may be used as the classification model, and the time gradient of the classification model is updated based on the time loss information is returned to obtain the updated time gradient, and the identification gradient of the recognition model is updated based on the identification loss information until the target loss information satisfies the convergence condition, thereby obtaining the trained recognition model.

With reference to fig. 4, in this embodiment of the present application, a difference learning network as shown in fig. 4 is used to autonomously learn the distribution difference of face images between different time information, and the difference learning network introduces a time information decoupling layer between the recognition model and the classification model and assists the recognition model in performing training of time information decoupling, so that the recognition model learns the common feature distribution between face images of different time information after time information decoupling, thereby significantly and effectively improving the accuracy of face recognition of cross-time information. Meanwhile, the information identification method provided by the embodiment of the application starts from the image features extracted by the identification model, and the accuracy of the identification model on the cross-time information face identification can be effectively improved by directly carrying out fine adjustment on the image features extracted by the identification model, rather than carrying out matching on the feature space distribution of the face images of different time information through re-weighting or geometric transformation.

After the recognition model is converged based on the target loss information, the recognition model is obtained, the image to be recognized can be recognized according to the recognition model after training, wherein various ways of recognizing the image to be recognized according to the recognition model after training can be adopted, for example, the image to be recognized can be obtained, the recognition model after training is adopted to conduct feature extraction on the image to be recognized, the image feature to be recognized corresponding to the image to be recognized is obtained, the image to be compared is obtained, the recognition model after training is adopted to conduct feature extraction on the image to be compared, the image feature to be compared corresponding to the image to be compared is obtained, the matching coefficient between the image feature to be recognized and the image feature to be compared is recognized based on the recognition model after training, and the matching result between the image to be recognized and the image to be compared is determined according to the matching coefficient.

The image to be identified may be a face image to be identified, the image feature to be identified may be an image feature corresponding to the image to be identified, the image to be compared may be a face image to be compared, the image feature to be compared may be an image feature corresponding to the image to be compared, the matching coefficient may be a coefficient representing a matching degree between the image feature to be identified and the image feature to be compared, for example, may be a coefficient representing a similarity degree between the image feature to be identified and the image feature to be compared, and the matching result may include information such as matching or unmatching between the image feature to be identified and the image feature to be compared.

For example, when the image feature to be recognized and the image feature to be compared are face images with the same identification information of different time information, the matching result can be the matching between the image feature to be recognized and the image feature to be compared or the information with the same identity, for example, the recognition model after training can be applied to the cross-age face recognition service, for example, in the application scene of cross-age person searching, the common distribution feature except for the age information between the image feature to be recognized and the image feature to be compared can be extracted through the recognition model after training, so that the image feature to be recognized and the image feature to be compared can be obtained, and the matching coefficient between the image feature to be recognized and the image feature to be compared can be recognized, so that the influence of the age factor on the matching coefficient between the face images can be avoided, and the accuracy of the cross-age face recognition can be improved.

Therefore, referring to fig. 5, fig. 5 is an overall flow chart of an information identifying method according to an embodiment of the present application, where the information identifying method includes an identifying model training stage and an identifying model deployment stage. In the training stage of the recognition model, a preset recognition model is trained through a training recognition model unit module to obtain a trained recognition model, and a preset classification model is trained through a training classification model unit module to obtain a trained classification model, so that a time information decoupling layer is introduced to perform time information decoupling on extracted features of the recognition model by using the classification model. The time information decoupling layer is connected with the recognition model, and the time information decoupling can be carried out on the image features extracted by the recognition model in an anti-learning mode, so that the gradient of the classification model acts on the recognition model in a reverse direction, the features extracted by the recognition model do not contain features which can be used for distinguishing the time information, and the common feature distribution among the face images of different time information learned by the recognition model is achieved. In this way, in the information recognition method provided by the embodiment of the application, the proportion configuration between the face images of different time information is not required to be adjusted, only the time information of the face images is required to be consistent, and in the training process, the additionally added classification model occupies negligible proportion of the video memory to the time occupied by training, so that the light weight of the information recognition method is ensured. In the recognition model deployment stage, after training is finished, the classification model does not need to participate in application deployment, so that the information recognition method can be applied to a large model or a small model, and the original recognition model does not need to be changed, and therefore the trained recognition model can be combined into a conventional face recognition system to recognize face images, the accuracy of face comparison of different time information is improved under the condition that the deployment module is not additionally increased, and the information recognition efficiency is improved.

From the above, the embodiment of the application obtains the image sample; extracting image features of the image sample by adopting an identification model, and determining identification information of the image sample based on the extracted image features; acquiring time information corresponding to the image sample, and determining time loss information corresponding to the image sample based on the time information and the time information label; determining the identification loss information corresponding to the image sample according to the identification information and the identification information label, and eliminating the time loss information from the identification loss information to obtain target loss information; and converging the recognition model based on the target loss information to obtain the recognition model after training. The method comprises the steps of classifying image features extracted by a recognition model by adopting a classification model, calculating time loss information according to time information labels corresponding to time information obtained by classification and image samples, calculating the identification loss information according to identification information and identification information labels determined by the recognition model based on the extracted image features, eliminating the time loss information in the identification loss information to obtain target loss information so as to train the recognition model, so that the recognition model does not contain the characteristics of the recognition time information in the image features extracted by the image samples, and recognizing an image to be processed based on the trained recognition model.

According to the method described in the above embodiments, examples are described in further detail below.

In this embodiment, an example will be described in which the information identifying apparatus is specifically integrated in a computer device. The information identification method uses a server as an execution subject, and uses the time information as age information as an example to specifically describe the information identification method. It will be appreciated that, in the specific embodiment of the present application, related data such as user information is involved, and when the above embodiments of the present application are applied to specific products or technologies, user permission or consent is required to be obtained, and the collection, use and processing of related data is required to comply with related laws and regulations and standards of related countries and regions.

For better describing the embodiment of the present application, please refer to fig. 6, fig. 6 is another flow chart of the information identifying method provided in the embodiment of the present application. The specific flow is as follows:

in step 201, the server acquires an initial image sample, and performs feature extraction on the initial image sample by using the identification model to obtain an initial image feature.

With continued reference to fig. 3b, when the server acquires an initial image sample, the server may read the initial image sample, combine the read sample data into a batch, and send the batch into a depth network unit of the recognition model to extract image features, so as to obtain initial image features.

In step 202, the server classifies the initial image features by using a preset classification model to obtain initial time information corresponding to the initial image features, and converges the preset classification model based on the initial time information and the initial time information label to obtain a classification model.

The server can predict time information through the initial image features extracted from the initial image sample by the classification model unit identification model unit module. The class labels can be labels of different age segments, so that initial time information output by the preset classification model and initial time information labels of initial image samples can be used as inputs to calculate loss information of the preset classification model, and a loss function corresponding to the loss information can be a cross entropy loss function or a regression type loss function. And furthermore, the whole network of the preset classification model can be trained and optimized based on the gradient descent method until the training result meets the training termination condition. The training termination condition for terminating the training of the model can generally set that the iteration times meet a set value, or the loss information calculated by the cross entropy objective function is smaller than the set value, so that the training of the preset classification model can be completed, and the trained classification model is obtained based on the model parameters at the moment.

In step 203, the server acquires an image sample, performs image feature extraction on the image sample using the identification model, and determines identification information of the image sample based on the extracted image features.

The image sample carries an identification information tag and a time information tag.

The server may acquire the image sample in various manners, for example, the server may acquire the image sample from a memory connected to the information identifying apparatus, or may acquire the image sample from another data storage terminal. Or may be obtained from a memory of the entity terminal, or may be obtained from a virtual memory space, such as a data set or a picture library. Alternatively, the image samples may be obtained from one storage location or may be obtained from a plurality of storage locations, for example, the image samples may be stored on a blockchain, and the information identifying apparatus obtains the image samples from the blockchain. The information identifying means may be configured to collectively acquire the image samples in a period of time in response to a certain image sample acquisition instruction, or may be configured to continue acquisition of the image samples according to a certain image sample acquisition logic, or the like.

In step 204, the server obtains time information corresponding to the image sample, and determines time loss information corresponding to the image sample based on the time information and the time information tag.

The server may obtain the time information corresponding to the image sample in various manners, for example, the server may use a classification model to predict the time information of the image feature, and output a probability distribution of which time information the image feature belongs to, so that the time information of the image sample may be determined according to the probability distribution.

After the time information corresponding to the image sample is acquired, the time loss information corresponding to the image sample can be determined based on the time information and the time information tag. The time information and the time information tag may be used to determine the time loss information corresponding to the image sample in various manners, for example, the time loss information corresponding to the image sample may be calculated by using a cross entropy loss function based on the time information and the time information tag.

In step 205, the server determines the identification loss information corresponding to the image sample according to the identification information and the identification information label, and eliminates the time loss information from the identification loss information to obtain the target loss information.

The method of eliminating the time loss information from the identification loss information may be various, for example, the time loss information may be subtracted from the identification loss information, so that the target loss information may be obtained. For example, as shown in formula (1).

In step 206, the server determines a convergence condition of the target loss information based on the model information corresponding to the identification model and the classification model, and updates the time gradient of the classification model according to the time loss information to obtain an updated time gradient, and updates the identification gradient of the identification model according to the identification loss information to obtain an updated identification gradient when the target loss information does not meet the convergence condition.

The server may update the time gradient of the classification model according to the time loss information in various manners, for example, the server may calculate the gradient of the loss function corresponding to the time loss information according to the time loss information, so as to update the calculated gradient to the time gradient of the classification model, and obtain the updated time gradient.

In step 207, the server obtains the current number of gradient updates, screens out the target time gradient from the updated time gradient, calculates the ratio of the number of gradient updates to the preset update parameter to obtain an update frequency ratio, and calculates the ratio of the number of gradient updates to the preset total number of updates to obtain an update frequency ratio when the update frequency ratio is of a preset value type.

The target time gradient may be a time gradient used to adjust the updated marker gradient in the updated time gradient.

The server may screen the target time gradient from the updated time gradients, for example, the server may determine the updated time gradient corresponding to the network layer connected to the identification model in the updated time gradient as the target time gradient, for example, please continue to refer to fig. 4, and may determine the updated time gradient corresponding to the network layer connected to the information decoupling layer in the classification model of the classification model unit module as the target time gradient.

In step 208, the server converts the ratio of the update times into an initial gradient adjustment parameter of the target time gradient, and weights the initial gradient adjustment parameter according to a preset negative adjustment coefficient to obtain a gradient adjustment parameter.

The method of converting the ratio of the number of updates to the initial gradient adjustment parameter of the target time gradient may be various, for example, the method shown in the formula (2) may be used to convert the ratio of the number of updates to the initial gradient adjustment parameter of the target time gradient.

After the server converts the ratio of the update times into the initial gradient adjustment parameter of the target time gradient, the initial gradient adjustment parameter can be weighted according to a preset negative adjustment coefficient. There may be various ways to weight the initial gradient adjustment parameter according to a preset negative adjustment coefficient, for example, the preset negative adjustment coefficient may be-1, and the gradient adjustment parameter may be represented as- γ.

In step 209, the server weights the gradient adjustment parameter and the target time gradient to obtain a weighted target time gradient, and fuses the weighted target time gradient and the updated marker gradient to obtain an adjusted marker gradient.

The server may weight the gradient adjustment parameter and the target time gradient after determining the gradient adjustment parameter of the target time gradient according to the gradient update times. The server may weight the gradient adjustment parameter and the target time gradient in various manners, for example, the server may multiply the gradient adjustment parameter and the target time gradient to obtain a weighted target time gradient as- γi, where I represents the target time gradient.

After weighting the gradient adjustment parameter and the target time gradient, the server may fuse the weighted target time gradient and the updated identified gradient. The server may perform fusion on the weighted target time gradient and the updated identifier gradient in various manners, for example, the server may superimpose the weighted target time gradient and the updated identifier gradient on the recognition model, so that the recognition model performs gradient update based on the updated identifier gradient, and superimpose a gradient update based on the weighted target time gradient.

For example, please continue to refer to fig. 4, a time information decoupling layer is introduced between the recognition model and the classification model, and the time information decoupling layer has the function of ensuring that when the recognition model is forward, the output of the recognition model is consistent with the input of the classification model, that is, the image features extracted by the recognition model can be indiscriminately input into the classification model, and in addition, when the classification model is subjected to gradient update, the classification model can be returned to the gradient of the recognition model through the time information decoupling layer to adjust, so that the gradient returned to the recognition model by the classification model acts on the recognition model in the opposite direction, and further, the features extracted by the recognition model obtained by training in the image sample do not contain the relevant features of time information, so that the common feature distribution between face images of different time information is learned, and the accuracy of face image recognition comparison between different time information is improved. Alternatively, the mathematical expression of the time information decoupling layer may be as shown in equation (3).

In step 210, the server updates the network parameters of the recognition model according to the adjusted identification gradient to obtain an updated recognition model, and updates the network parameters of the classification model according to the updated time gradient to obtain an updated classification model.

The updated identification model may be an identification model obtained by updating network parameters of the identification model according to the adjusted identification gradient, and the updated classification model may be a classification model obtained by updating network parameters of the classification model according to the updated time gradient.

In step 211, the server updates the updated recognition model and the updated recognition model to the target loss information to obtain updated target loss information, and converges the updated recognition model based on the updated target loss information to obtain a trained recognition model.

The server updates the target loss information by using the updated identification model and the updated identification model, and various ways of obtaining updated target loss information may be used, for example, the server may use the updated identification model to extract image features of the image sample, and determine updated identification information of the image sample based on the extracted updated image features; classifying the image features by adopting an updated classification model to obtain updated time information of the image sample, and determining updated time loss information corresponding to the image sample based on the updated time information and the updated time information label; and determining updated identification loss information corresponding to the image sample according to the updated identification information and the identification information label, and eliminating updated time loss information from the updated identification loss information to obtain updated target loss information.

After updating the post-update recognition model and the post-update recognition model with respect to the target loss information, the server may converge the post-update recognition model based on the post-update target loss information, where the server may converge the post-update recognition model based on the post-update target loss information, for example, the server may use the post-update recognition model as the post-training recognition model when the post-update target loss information satisfies the convergence condition, use the post-update target loss information as the target loss information when the post-update target loss information does not satisfy the convergence condition, use the post-update recognition model as the recognition model, use the post-update classification model as the classification model, and return to perform the step of updating the time gradient of the classification model according to the time loss information to obtain the post-update time gradient, update the identification gradient of the recognition model according to the identification loss information to obtain the post-update identification gradient until the target loss information satisfies the convergence condition.

After the updated recognition model is converged based on the updated target loss information to obtain a trained recognition model, the server can recognize the image to be recognized according to the trained recognition model. The server may obtain the image to be identified according to the trained identification model, and perform feature extraction on the image to be identified by using the trained identification model to obtain the feature of the image to be identified corresponding to the image to be identified, obtain the image to be compared, perform feature extraction on the image to be compared by using the trained identification model to obtain the feature of the image to be compared corresponding to the image to be compared, identify a matching coefficient between the feature of the image to be identified and the feature of the image to be compared based on the trained identification model, and determine a matching result between the image to be identified and the image to be compared according to the matching coefficient.

For example, when the image feature to be recognized and the image feature to be compared are face images with the same identification information of different time information, the matching result can be the matching between the image feature to be recognized and the image feature to be compared or the information with the same identity, for example, the recognition model after training can be applied to the cross-age face recognition service, for example, in the application scene of cross-age person searching, the server can extract the common distribution feature except for the age information between the image feature to be recognized and the image feature to be compared through the recognition model after training, so that the image feature to be recognized and the image feature to be compared can be obtained, and the matching coefficient between the image feature to be recognized and the image feature to be compared can be recognized, so that the influence of the age factor on the matching coefficient between the face images can be avoided, and the accuracy of the cross-age face recognition can be improved.

As can be seen from the above, in the embodiment of the present application, an initial image sample is obtained through a server, and an identification model is used to perform feature extraction on the initial image sample, so as to obtain initial image features; the server classifies the initial image features by adopting a preset classification model to obtain initial time information corresponding to the initial image features, and converges the preset classification model based on the initial time information and an initial time information label to obtain a classification model; the method comprises the steps that a server obtains an image sample, image feature extraction is carried out on the image sample by adopting an identification model, and identification information of the image sample is determined based on the extracted image feature; the server acquires time information corresponding to the image sample, and determines time loss information corresponding to the image sample based on the time information and the time information label; the server determines the identification loss information corresponding to the image sample according to the identification information and the identification information label, and eliminates the time loss information in the identification loss information to obtain target loss information; the server determines a convergence condition of the target loss information based on model information corresponding to the identification model and the classification model, and when the target loss information does not meet the convergence condition, updates the time gradient of the classification model according to the time loss information to obtain an updated time gradient, and updates the identification gradient of the identification model according to the identification loss information to obtain an updated identification gradient; the server acquires the current gradient update times, screens out a target time gradient from the updated time gradient, calculates the ratio of the gradient update times to a preset update parameter to obtain an update frequency ratio, and calculates the ratio of the gradient update times to the preset total update times when the update frequency ratio is of a preset numerical value type to obtain an update time ratio; the server converts the ratio of the update times into an initial gradient adjustment parameter of the target time gradient, and weights the initial gradient adjustment parameter according to a preset negative adjustment coefficient to obtain a gradient adjustment parameter; the server weights the gradient adjusting parameter and the target time gradient to obtain a weighted target time gradient, and fuses the weighted target time gradient and the updated mark gradient to obtain an adjusted mark gradient; the server updates the network parameters of the identification model according to the adjusted identification gradient to obtain an updated identification model, and updates the network parameters of the classification model according to the updated time gradient to obtain an updated classification model; the server updates the target loss information with the updated recognition model and the updated recognition model to obtain updated target loss information, and converges the updated recognition model based on the updated target loss information to obtain a trained recognition model. According to the method, time loss information is calculated by acquiring time information labels corresponding to time information obtained by time classification of image features extracted by the recognition model and the image samples, the time loss information is calculated according to the identification information determined by the recognition model based on the extracted image features and the identification information labels, then the time loss information is removed from the identification loss information to obtain target loss information so as to train the recognition model, so that the recognition model does not contain the features of the recognition time information in the image features extracted by the image samples, then the recognition model can be subjected to gradient update according to the identification loss information and the time loss information when the target loss information does not meet convergence conditions, iterative training can be performed based on the updated recognition model, the images to be processed can be recognized based on the trained recognition model, the accuracy of the identification information of the face images with different time information is improved, meanwhile, an additional deployment module is not required to be added, the increase of training time consumption is avoided, the light weight of the recognition model is further ensured, and the information recognition efficiency is improved.

In order to better implement the above method, the embodiment of the present application further provides an information identifying apparatus, which may be integrated in a computer device, and the computer device may be a server.

For example, as shown in fig. 7, a schematic structural diagram of an information identifying apparatus provided in an embodiment of the present application may include a sample acquiring unit 301, a feature extracting unit 302, a determining unit 303, a culling unit 304, and a converging unit 305, as follows:

a sample acquiring unit 301, configured to acquire an image sample, where the image sample carries an identification information tag and a time information tag;

a feature extraction unit 302, configured to perform image feature extraction on the image sample using the identification model, and determine identification information of the image sample based on the extracted image feature;

a determining unit 303, configured to obtain time information corresponding to the image sample, and determine time loss information corresponding to the image sample based on the time information and the time information tag;

the rejecting unit 304 is configured to determine, according to the identification information and the identification information tag, identification loss information corresponding to the image sample, and reject the time loss information from the identification loss information, so as to obtain target loss information;

The convergence unit 305 is configured to converge the recognition model based on the target loss information, so as to obtain a trained recognition model, where the trained recognition model is used to recognize the image to be recognized.

In one embodiment, the convergence unit 305 includes:

and the convergence subunit is used for converging the recognition model according to the time loss information and the identification loss information when the target loss information does not meet the convergence condition, so as to obtain the recognition model after training.

In one embodiment, the convergence subunit comprises:

In one embodiment, the update convergence module comprises:

the gradient adjusting sub-module is used for adjusting the updated marking gradient based on the gradient updating times and the updated time gradient to obtain an adjusted marking gradient;

and the parameter updating sub-module is used for updating the network parameters of the classification model according to the updated time gradient and updating the network parameters of the recognition model according to the adjusted identification gradient so as to obtain the recognition model after training.

In one embodiment, the gradient adjustment sub-module is configured to:

and based on the updated target loss information, converging the updated recognition model to obtain a trained recognition model.

when the updated target loss information meets the convergence condition, the updated recognition model is used as the recognition model after training;

And when the updated target loss information does not meet the convergence condition, taking the updated target loss information as target loss information, taking the updated identification model as identification model, taking the updated classification model as classification model, and returning to execute the step of updating the time gradient of the classification model according to the time loss information to obtain updated time gradient, updating the identification gradient of the identification model according to the identification loss information to obtain updated identification gradient until the target loss information meets the convergence condition, so as to obtain the trained identification model.

In an embodiment, the information identifying apparatus further includes:

the initial feature extraction unit is used for extracting features of the initial image sample by adopting the identification model to obtain initial image features;

and the initial convergence unit is used for converging the preset classification model based on the initial time information and the initial time information label to obtain the classification model.

In an embodiment, the information identifying apparatus further includes:

the image feature extraction subunit to be identified is used for obtaining an image to be identified, and carrying out feature extraction on the image to be identified by adopting the trained identification model to obtain the image feature to be identified corresponding to the image to be identified;

and the recognition subunit is used for recognizing the matching coefficient between the image feature to be recognized and the image feature to be compared based on the trained recognition model, and determining the matching result between the image to be recognized and the image to be compared according to the matching coefficient.

In the implementation, each unit may be implemented as an independent entity, or may be implemented as the same entity or several entities in any combination, and the implementation of each unit may be referred to the foregoing method embodiment, which is not described herein again.

From the above, the embodiment of the present application acquires an image sample through the acquisition unit 301; the feature extraction unit 302 performs image feature extraction on the image sample by adopting an identification model, and determines identification information of the image sample based on the extracted image features; the determining unit 303 obtains time information corresponding to the image sample, and determines time loss information corresponding to the image sample based on the time information and the time information tag; the eliminating unit 304 determines the identification loss information corresponding to the image sample according to the identification information and the identification information label, and eliminates the time loss information in the identification loss information to obtain target loss information; the convergence unit 305 converges the recognition model based on the target loss information, resulting in a trained recognition model. The method comprises the steps of classifying image features extracted by a recognition model by adopting a classification model, calculating time loss information according to time information labels corresponding to time information obtained by classification and image samples, calculating the identification loss information according to identification information and identification information labels determined by the recognition model based on the extracted image features, eliminating the time loss information in the identification loss information to obtain target loss information so as to train the recognition model, so that the recognition model does not contain the characteristics of the recognition time information in the image features extracted by the image samples, and recognizing an image to be processed based on the trained recognition model.

The embodiment of the application also provides a computer device, as shown in fig. 8, which shows a schematic structural diagram of the computer device according to the embodiment of the application, wherein the computer device may be a server, in particular:

the computer device may include one or more processors 401 of a processing core, memory 402 of one or more computer readable storage media, a power supply 403, and an input unit 404, among other components. Those skilled in the art will appreciate that the computer device structure shown in FIG. 8 is not limiting of the computer device and may include more or fewer components than shown, or may be combined with certain components, or a different arrangement of components. Wherein:

the processor 401 is a control center of the computer device, connects various parts of the entire computer device using various interfaces and lines, performs various functions of the computer device and processes data by running or executing software programs and/or modules stored in the memory 402, and calling data stored in the memory 402. Optionally, processor 401 may include one or more processing cores; preferably, the processor 401 may integrate an application processor and a modem processor, wherein the application processor mainly processes an operating system, a user interface, an application program, etc., and the modem processor mainly processes wireless communication. It will be appreciated that the modem processor described above may not be integrated into the processor 401.

The memory 402 may be used to store software programs and modules, and the processor 401 executes various functional applications and information identification by executing the software programs and modules stored in the memory 402. The memory 402 may mainly include a storage program area and a storage data area, wherein the storage program area may store an operating system, an application program (such as a sound playing function, an image playing function, etc.) required for at least one function, and the like; the storage data area may store data created according to the use of the computer device, etc. In addition, memory 402 may include high-speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory 402 may also include a memory controller to provide the processor 401 with access to the memory 402.

The computer device further comprises a power supply 403 for supplying power to the various components, preferably the power supply 403 may be logically connected to the processor 401 by a power management system, so that functions of charge, discharge, and power consumption management may be performed by the power management system. The power supply 403 may also include one or more of any of a direct current or alternating current power supply, a recharging system, a power failure detection circuit, a power converter or inverter, a power status indicator, and the like.

The computer device may also include an input unit 404, which input unit 404 may be used to receive input numeric or character information and to generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

Although not shown, the computer device may further include a display unit or the like, which is not described herein. In particular, in this embodiment, the processor 401 in the computer device loads executable files corresponding to the processes of one or more application programs into the memory 402 according to the following instructions, and the processor 401 executes the application programs stored in the memory 402, so as to implement various functions as follows:

acquiring an image sample; extracting image features of the image sample by adopting an identification model, and determining identification information of the image sample based on the extracted image features; acquiring time information corresponding to the image sample, and determining time loss information corresponding to the image sample based on the time information and the time information label; determining the identification loss information corresponding to the image sample according to the identification information and the identification information label, and eliminating the time loss information from the identification loss information to obtain target loss information; and converging the recognition model based on the target loss information to obtain the recognition model after training.

The specific implementation of each operation may be referred to the previous embodiments, and will not be described herein. It should be noted that, the computer device provided in the embodiment of the present application and the method applicable to information identification in the above embodiment belong to the same concept, and detailed implementation processes of the computer device are shown in the above method embodiment, which is not repeated here.

Those of ordinary skill in the art will appreciate that all or a portion of the steps of the various methods of the above embodiments may be performed by instructions, or by instructions controlling associated hardware, which may be stored in a computer-readable storage medium and loaded and executed by a processor.

To this end, embodiments of the present application provide a computer readable storage medium having stored therein a plurality of instructions capable of being loaded by a processor to perform any of the steps of the information identification method provided by the embodiments of the present application. For example, the instructions may perform the steps of:

Wherein the computer-readable storage medium may comprise: read Only Memory (ROM), random access Memory (RAM, random Access Memory), magnetic or optical disk, and the like.

Because the instructions stored in the computer readable storage medium can execute the steps in any information identification method provided by the embodiments of the present application, the beneficial effects that any information identification method provided by the embodiments of the present application can achieve can be achieved, which are detailed in the previous embodiments and are not described herein.

Wherein according to an aspect of the application, a computer program product or a computer program is provided, the computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the methods provided in the various alternative implementations provided in the above embodiments.

The foregoing has described in detail the methods, apparatuses and computer readable storage medium provided by embodiments of the present application, and specific examples have been presented herein to illustrate the principles and implementations of the present application, and the above description of the embodiments is only for aiding in the understanding of the methods and core ideas of the present application; meanwhile, as those skilled in the art will have variations in the specific embodiments and application scope in light of the ideas of the present application, the present description should not be construed as limiting the present application.

Claims

1. An information identification method, comprising:

2. The information identifying method as claimed in claim 1, wherein the converging the identifying model based on the target loss information to obtain a trained identifying model includes:

Determining convergence conditions of the target loss information based on model information corresponding to the identification model and the classification model;

and when the target loss information does not meet the convergence condition, converging the recognition model according to the time loss information and the identification loss information to obtain a trained recognition model.

3. The information identifying method as claimed in claim 2, wherein the converging the identifying model according to the time loss information and the identification loss information to obtain a trained identifying model includes:

updating the time gradient of the classification model according to the time loss information to obtain an updated time gradient, and updating the identification gradient of the identification model according to the identification loss information to obtain an updated identification gradient;

and converging the recognition model based on the updated time gradient and the updated identification gradient to obtain a trained recognition model.

4. The information recognition method of claim 3, wherein the converging the recognition model based on the updated time gradient and the updated identification gradient to obtain a trained recognition model comprises:

Acquiring the current gradient update times;

based on the gradient updating times and the updated time gradient, adjusting the updated identification gradient to obtain an adjusted identification gradient;

and updating the network parameters of the classification model according to the updated time gradient, and updating the network parameters of the recognition model according to the adjusted identification gradient to obtain a trained recognition model.

5. The method of information identification of claim 4, wherein adjusting the updated marker gradient based on the number of gradient updates and the updated time gradient to obtain an adjusted marker gradient comprises:

6. The information identifying method as claimed in claim 5, wherein the determining the gradient adjustment parameter of the target time gradient according to the number of gradient updates includes:

7. The method of claim 4, wherein updating the network parameters of the classification model according to the updated time gradient and updating the network parameters of the recognition model according to the adjusted identification gradient to obtain the trained recognition model comprises:

8. The method of claim 7, wherein the converging the updated recognition model based on the updated target loss information to obtain a trained recognition model comprises:

9. The method for identifying information according to claim 1, wherein before the time information corresponding to the image sample is obtained, further comprising:

acquiring an initial image sample, wherein the initial image sample carries an initial time information tag;

extracting features of the initial image sample by adopting an identification model to obtain initial image features;

classifying the initial image features by adopting a preset classification model to obtain initial time information corresponding to the initial image features;

and converging the preset classification model based on the initial time information and the initial time information label to obtain a classification model.

10. The information identifying method according to any one of claims 1 to 9, characterized in that the identifying model is converged based on the target loss information, and after obtaining a trained identifying model, further comprising:

acquiring an image to be identified, and extracting features of the image to be identified by adopting the trained identification model to obtain features of the image to be identified corresponding to the image to be identified;

obtaining images to be compared, and extracting features of the images to be compared by adopting the recognition model after training to obtain features of the images to be compared, which correspond to the images to be compared;

And identifying the matching coefficient between the image feature to be identified and the image feature to be compared based on the trained identification model, and determining the matching result between the image to be identified and the image to be compared according to the matching coefficient.

11. An information identifying apparatus, comprising:

12. A computer readable storage medium storing a plurality of instructions adapted to be loaded by a processor to perform the steps in the information identification method of any one of claims 1 to 10.

13. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, the processor implementing the information identification method of any one of claims 1 to 10 when the computer program is executed.

14. A computer program product, characterized in that it comprises a computer program/instruction which, when executed by a processor, implements the steps of the information identification method of any of claims 1 to 10.