Information identification method, device and computer readable storage medium

Abstract

Embodiments of the present application disclose an information recognition method and device and a computer readable storage medium, which can be applied to various scenes such as cloud technology, artificial intelligence, intelligent transportation and auxiliary driving. An image sample is obtained, an identification model is used to extract image features of the image sample, identification information of the image sample is determined based on the extracted image features, time information corresponding to the image sample is obtained, time loss information is determined based on the time information and a time information label, identification loss information is determined according to the identification information and an identification information label, the time loss information is removed from the identification loss information to obtain target loss information, the identification model is converged based on the target loss information to obtain a trained identification model. In this way, the trained identification model is used to recognize a to-be-recognized image, the accuracy of the identification model in recognizing face images of different time information is improved, and the efficiency of information recognition is improved.

Description

Technical Field

[0001] This application relates to the field of Internet technology, specifically to an information identification method, apparatus, and computer-readable storage medium. Background Technology

[0002] In recent years, with the rapid development of internet technology, artificial intelligence technology has also been continuously evolving. Among these advancements, facial recognition technology has been widely applied in people's daily lives. The widespread adoption of facial recognition applications demands that facial recognition models be robust across various scenarios, such as recognizing facial images of the same person from different times. Currently, most facial recognition models are trained using a multi-task approach, adjusting the weights of the loss function between image features from different times during training to improve accuracy.

[0003] Research and practice on existing technologies have revealed that the current method of training face recognition models through multi-task approaches to improve the accuracy of face recognition increases the model size and leads to longer training times. Furthermore, it cannot consistently improve the accuracy of face recognition models for face information at different times, resulting in low information recognition efficiency. Summary of the Invention

[0004] This application provides an information recognition method, apparatus, and computer-readable storage medium, which can improve the accuracy of the recognition model in recognizing face images with information from different times, thereby improving the efficiency of information recognition.

[0005] This application provides an information identification method, including: Acquire image samples, wherein the image samples carry identification information tags and time information tags; The image sample is used to extract image features using a recognition model, and the identification information of the image sample is determined based on the extracted image features; The time information corresponding to the image sample is obtained, and based on the time information and the time information label, the time loss information corresponding to the image sample is determined. The time information is obtained by performing time classification on the image features using a classification model. Based on the identification information and identification information label, the identification loss information corresponding to the image sample is determined, and the time loss information is removed from the identification loss information to obtain the target loss information; The recognition model is converged based on the target loss information to obtain a trained recognition model, which is used to recognize the image to be recognized.

[0006] Accordingly, embodiments of this application provide an information identification device, including: A sample acquisition unit is used to acquire image samples, wherein the image samples carry identification information tags and time information tags; The feature extraction unit is used to extract image features from the image sample using a recognition model, and to determine the identification information of the image sample based on the extracted image features; The determining unit is used to acquire the time information corresponding to the image sample, and determine the time loss information corresponding to the image sample based on the time information and the time information label, wherein the time information is the information obtained by performing time classification on the image features using a classification model; The elimination unit is used to determine the identification loss information corresponding to the image sample based on the identification information and the identification information label, and to remove the time loss information from the identification loss information to obtain the target loss information; A convergence unit is used to converge the recognition model based on the target loss information to obtain a trained recognition model, which is used to recognize the image to be recognized.

[0007] In one embodiment, the convergence unit includes: The convergence condition determination subunit is used to determine the convergence condition of the target loss information based on the model information corresponding to the recognition model and the classification model; The convergence subunit is used to converge the recognition model based on the time loss information and the labeling loss information when the target loss information does not meet the convergence condition, so as to obtain the trained recognition model.

[0008] In one embodiment, the convergence subunit includes: The gradient update module is used to update the temporal gradient of the classification model according to the temporal loss information to obtain the updated temporal gradient, and to update the label gradient of the recognition model according to the label loss information to obtain the updated label gradient. An update convergence module is used to converge the recognition model based on the updated time gradient and the updated label gradient to obtain the trained recognition model.

[0009] In one embodiment, the update convergence module includes: The gradient update count acquisition submodule is used to obtain the current gradient update count; The gradient adjustment submodule is used to adjust the updated identifier gradient based on the number of gradient updates and the updated time gradient to obtain the adjusted identifier gradient. The parameter update submodule is used to update the network parameters of the classification model according to the updated time gradient, and to update the network parameters of the recognition model according to the adjusted label gradient, so as to obtain the trained recognition model.

[0010] In one embodiment, the gradient adjustment submodule is configured to: A target time gradient is selected from the updated time gradients, and the target time gradient is the time gradient used to adjust the updated identifier gradient in the updated time gradients; Based on the number of gradient updates, the gradient adjustment parameter of the target temporal gradient is determined, and the gradient adjustment parameter and the target temporal gradient are weighted to obtain the weighted target temporal gradient. The weighted target temporal gradient and the updated label gradient are fused to obtain the adjusted label gradient.

[0011] In one embodiment, the gradient adjustment submodule can specifically be used for: Calculate the ratio of the gradient update count to the preset update parameter to obtain the update frequency ratio; When the update frequency ratio is a preset numerical type, the ratio of the gradient update count to the preset total update count is calculated to obtain the update count ratio; The ratio of update times is converted into the initial gradient adjustment parameter of the target time gradient, and the initial gradient adjustment parameter is weighted according to the preset negative adjustment coefficient to obtain the gradient adjustment parameter.

[0012] In one embodiment, the parameter update submodule can be specifically used for: Based on the adjusted identifier gradient, the network parameters of the recognition model are updated to obtain the updated recognition model. Based on the updated time gradient, the network parameters of the classification model are updated to obtain the updated classification model. The updated target loss information is obtained by updating the target loss information using the updated recognition model and the updated recognition model. Based on the updated target loss information, the updated recognition model is converged to obtain the trained recognition model.

[0013] In one embodiment, the parameter update submodule can be specifically used for: When the updated target loss information satisfies the convergence condition, the updated recognition model is used as the trained recognition model. When the updated target loss information does not meet the convergence condition, the updated target loss information is used as the target loss information, the updated recognition model is used as the recognition model, the updated classification model is used as the classification model, and the process of updating the temporal gradient of the classification model based on the temporal loss information to obtain the updated temporal gradient, and updating the label gradient of the recognition model based on the label loss information to obtain the updated label gradient is repeated until the target loss information meets the convergence condition, thus obtaining the trained recognition model.

[0014] In one embodiment, the information identification device further includes: An initial sample acquisition unit is used to acquire an initial image sample, wherein the initial image sample carries an initial time information label; An initial feature extraction unit is used to extract features from the initial image sample using a recognition model to obtain initial image features; An initial classification unit is used to classify the initial image features using a preset classification model to obtain the initial time information corresponding to the initial image features; An initial convergence unit is used to converge the preset classification model based on the initial time information and the initial time information label to obtain a classification model.

[0015] In one embodiment, the information identification device further includes: The image feature extraction subunit is used to acquire the image to be identified and to extract features from the image to be identified using the trained recognition model to obtain the image features corresponding to the image to be identified. The image feature extraction subunit is used to acquire the image to be compared and to use the trained recognition model to extract features from the image to be compared, thereby obtaining the image features corresponding to the image to be compared. The recognition subunit is used to identify the matching coefficient between the features of the image to be recognized and the features of the image to be compared based on the trained recognition model, and to determine the matching result between the image to be recognized and the image to be compared based on the matching coefficient.

[0016] Furthermore, embodiments of this application also provide a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute steps in any of the information identification methods provided in embodiments of this application.

[0017] Furthermore, this application also provides a computer device, including a processor and a memory, wherein the memory stores an application program, and the processor is used to run the application program in the memory to implement the information identification method provided in this application.

[0018] This application also provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in the information identification method provided in this application.

[0019] This application embodiment acquires image samples; extracts image features from the image samples using a recognition model, and determines the identification information of the image samples based on the extracted image features; acquires the time information corresponding to the image samples, and determines the time loss information corresponding to the image samples based on the time information and time information labels. This time information is obtained by classifying the image features by time using a classification model; determines the identification loss information corresponding to the image samples based on the identification information and identification information labels, and removes the time loss information from the identification loss information to obtain the target loss information; and converges the recognition model based on the target loss information to obtain the trained recognition model. Therefore, by obtaining the time information obtained by time classification of image features extracted by the recognition model using a classification model and the time information labels corresponding to the image samples, time loss information is calculated. Then, based on the identification information determined by the recognition model based on the extracted image features and the identification information labels, identification loss information is calculated. Finally, time loss information is removed from the identification loss information to obtain target loss information for training the recognition model. This ensures that the image features extracted by the trained recognition model do not contain features related to time information in the image samples. As a result, common features other than time information features can be extracted from the image to be processed to identify the identification information of the image. Thus, by using the trained recognition model for image recognition, the accuracy of the face recognition model in identifying the identification information of face images with different time information is improved, thereby improving the efficiency of information recognition. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram illustrating an implementation scenario of an information identification method provided in this application. Figure 2 This is a flowchart illustrating an information identification method provided in an embodiment of this application; Figure 3aThis is a schematic diagram of the recognition model training process of an information recognition method provided in an embodiment of this application; Figure 3b This is a schematic diagram of the classification model training process of an information recognition method provided in an embodiment of this application; Figure 4 This is a schematic diagram illustrating the specific process of an information identification method provided in an embodiment of this application; Figure 5 This is a schematic diagram of the overall process of an information identification method provided in an embodiment of this application; Figure 6 This is another flowchart illustrating an information identification method provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of the information recognition device provided in the embodiments of this application; Figure 8 This is a schematic diagram of the structure of the computer device provided in the embodiments of this application. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] This application provides an information identification method, apparatus, and computer-readable storage medium. The information identification apparatus can be integrated into a computer device, which may be a server or a terminal, etc.

[0024] The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery network (CDN) acceleration services, and big data and artificial intelligence platforms. The terminal can include, but is not limited to, mobile phones, computers, smart voice interaction devices, smart home appliances, vehicle terminals, and aircraft. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited herein.

[0025] Please see Figure 1 Taking the integration of information identification devices into computer equipment as an example, Figure 1This is a schematic diagram of an implementation scenario for the information recognition method provided in this application. The computer device can be a server or a terminal. The computer device can acquire image samples; extract image features from the image samples using a recognition model, and determine the identification information of the image samples based on the extracted image features; acquire the time information corresponding to the image samples, and determine the time loss information corresponding to the image samples based on the time information and time information labels; determine the identification loss information corresponding to the image samples based on the identification information and identification information labels, and remove the time loss information from the identification loss information to obtain the target loss information; and converge the recognition model based on the target loss information to obtain the trained recognition model.

[0026] It should be noted that the embodiments of the present invention can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, and assisted driving. Figure 1 The illustrated scenario of the information recognition method is merely an example. The implementation environment of the information recognition method described in this application is intended to more clearly illustrate the technical solution of this application and does not constitute a limitation on the technical solution provided in this application. Those skilled in the art will understand that, with the evolution of information recognition and the emergence of new business scenarios, the technical solution provided in this application is equally applicable to similar technical problems.

[0027] The solutions provided in this application relate to technologies such as computer vision in artificial intelligence, and are specifically illustrated through the following embodiments. It should be noted that the order of description of the following embodiments is not intended to limit the preferred order of the embodiments.

[0028] This embodiment will be described from the perspective of an information identification device, which can be integrated into a computer device, which can be a server, and this application does not limit it.

[0029] Please see Figure 2 , Figure 2 This is a flowchart illustrating the information recognition method provided in an embodiment of this application. The information recognition method includes: In step 101, image samples are acquired.

[0030] The image sample can be a facial image sample, and it can carry identification information tags and time information tags. Optionally, the identification information tags and time information tags can also be information contained within the image sample. The identification information tag can be information carried by the image sample to mark the identity of the object in the image sample, used to indicate the identity of the object in the image sample. The time information tag can be information carried by the image sample to mark the time information of the object in the image sample, used to indicate the time information of the object in the image sample. The time information can be information indicating the age or age group of the object in the image sample.

[0031] There are several ways to acquire image samples. For example, they can be acquired from a memory connected to the information recognition device, or from other data storage terminals. They can also be acquired from the memory of a physical terminal, or from virtual storage spaces such as datasets or image libraries. Optionally, image samples can be acquired from one storage location or multiple storage locations. For example, image samples can be stored on a blockchain, and the information recognition device can acquire these image samples from the blockchain. The information recognition device can acquire image samples centrally over a period of time in response to an image sample acquisition command, or it can continuously acquire image samples according to a certain image sample acquisition logic.

[0032] It is understood that in the specific embodiments of this application, data such as user information are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0033] In step 102, an identification model is used to extract image features from the image samples, and the identification information of the image samples is determined based on the extracted image features.

[0034] The recognition model can be a trained model for recognizing face images. The image features can be features that characterize face information in the image sample. Specifically, they can be features extracted from the image sample by the recognition model to characterize face information in the image sample. The identification information can be information that characterizes the identity of the object in the image sample. Specifically, it can be the identity information of the object in the image sample obtained by recognizing the image sample by the recognition model.

[0035] Before using a recognition model to extract image features from image samples, a preset recognition model can be trained to obtain a pre-defined, untrained model. This pre-defined model can be trained using face image samples to obtain a trained recognition model. For example, please refer to... Figure 3a , Figure 3a This is a schematic diagram of the recognition model training process of an information recognition method provided in this application embodiment. The training process of the preset recognition model may include the following modules: (1) Training data preparation module: During the training process, face image samples are read and the read data is combined into a batch and sent to the deep network unit of the preset recognition model for processing.

[0036] Optionally, the acquired face image samples can be preprocessed. There are various preprocessing methods, such as adjusting the size of the face image samples to a preset size to better adapt to the processing of the preset recognition model.

[0037] (2) Recognition Model Unit Module: In this module, spatial features are extracted from face image samples, and the output feature map retains the spatial structure information of the face image. This module can be a Convolutional Neural Network (CNN) structure, which can include operations such as convolution calculation, nonlinear activation function (ReLU) calculation, and pooling calculation.

[0038] (3) Face recognition objective function calculation module: In this module, the features output by the fully connected mapping unit of the recognition model unit module and the label information of the face image sample that generates the features can be used as input to calculate the objective function value. The objective function can be a classification function, such as the softmax function, or other types of objective functions.

[0039] (4) Face recognition objective function optimization module: This module can train and optimize the entire network based on gradient descent. For example, it can use stochastic gradient descent, stochastic gradient descent with momentum term, Adam (a gradient descent update algorithm), and Adamard (a gradient descent update algorithm). During training, steps (1) to (4) can be repeated until the training result meets the training termination condition. The training termination condition for terminating model training can generally be set to meet the set value of the number of iterations, or the loss information calculated by the face recognition objective function is less than the set value to complete the training of the preset recognition model.

[0040] With the continuous development of artificial intelligence technology, facial recognition technology has been widely applied in people's daily lives. The widespread adoption of facial recognition applications requires models to be robust across various scenarios, such as recognizing facial images of the same person from different times, like cross-age identity searches or ID card comparisons. These scenarios require comparing images of the same person from different age groups. In typical facial recognition models, the registered database usually consists of recently collected facial images of the same age group. This leads to inconsistencies in the feature space distribution between facial images of different age groups during model training. The model cannot align the feature space distribution between these age groups, resulting in a poor ability to accurately recognize facial images across different age groups. To address this issue, most existing face recognition models employ a multi-task training approach, adjusting the weights of the loss function between image features from different time periods during training to improve accuracy. However, multi-task training requires adjusting the weights of the loss function for each task, and these weights are hyperparameters whose optimal values ​​change with data distribution. This makes it difficult to guarantee consistent improvements in cross-age face recognition accuracy during model training. Furthermore, multi-task training increases the size of the recognition model and the time required for training and computation.

[0041] To address the limitations of existing technologies in consistently improving the accuracy of face recognition models across different time periods, and the increased model size and training time resulting from multi-task training methods, which ultimately lead to lower recognition efficiency, this application provides an information recognition method. This method employs a different approach than traditional metrics to measure the distributional differences between different time periods. By adjusting the image features extracted by the recognition model using a classification model, the temporal information of face image samples from different time periods is decoupled. This ensures that the features extracted by the recognition model based on this method do not include time-related features, allowing it to learn the common distributional features of face images from different time periods. This effectively improves the accuracy of face recognition comparison across different time periods. Furthermore, it eliminates the need for additional deployment modules, avoiding increased training time and ensuring the lightweight nature of the recognition model, thereby improving recognition efficiency. The specific details are described below.

[0042] In step 103, the time information corresponding to the image sample is obtained, and the time loss information corresponding to the image sample is determined based on the time information and the time information label.

[0043] The time information refers to information obtained by classifying the image features using a classification model. For example, it could be information such as the age group to which the image sample corresponding to the image feature belongs. The classification model can be a pre-trained network used to classify the time information corresponding to the image features, such as an age estimation network. The time loss information can be information representing the difference between the time information and the time information label, that is, representing the error of the classification model in predicting the time information corresponding to the image sample based on the image features.

[0044] There are several ways to obtain the time information corresponding to the image sample. For example, a classification model can be used to predict the time information of the image features and output the probability distribution of which time information the image features belong to. Thus, the time information of the image sample can be determined based on the probability distribution.

[0045] Optionally, before obtaining the time information corresponding to the image sample, a recognition model can be used to train a preset classification model to obtain a trained classification model. There are several ways to train the preset classification model using a recognition model. For example, an initial image sample can be obtained, features can be extracted from the initial image sample using a recognition model to obtain initial image features, the preset classification model can be used to classify the initial image features to obtain the initial time information corresponding to the initial image features, and the preset classification model can be converged based on the initial time information and the initial time information label to obtain the classification model.

[0046] The preset classification model can be a pre-defined, untrained model used to predict the temporal information of image features. A recognition model can be used to train this preset model's ability to predict the temporal information of image features to obtain a trained recognition model. The initial image sample can be a face image sample used to train the preset classification model. This initial image sample carries an initial temporal information label, which is used to annotate the temporal information of objects in the initial image sample. The initial image features can be image features extracted from the initial image sample using the recognition model, and the initial temporal information can be the result of predicting the temporal information corresponding to the initial image features based on the preset classification model.

[0047] In one embodiment, please refer to Figure 3b , Figure 3b This is a schematic diagram of the classification model training process of an information recognition method provided in this application embodiment. The classification model training flowchart shown in Figure 3 can be used to train the preset classification model. Specifically, it can include the following modules: S1, Training Data Preparation Module: During the training process, the initial image samples are read, the read sample data are combined into a batch, and then sent to the deep network unit of the recognition model for processing.

[0048] Optionally, the obtained initial image samples can be preprocessed. There are various ways to preprocess them, such as adjusting the size of the initial image samples to a preset size to better adapt to the processing of the preset classification model.

[0049] S2. Classification Model Unit Module: This module predicts the temporal information of the initial image features extracted from the initial image samples by the recognition model unit module. The category labels can be labels representing different time periods, such as different age groups. The preset classification model can be a classification network, whose structure can be composed of convolutional neural networks, including operations such as convolution, non-linear activation functions, and pooling. The network structure can also be a fully connected network, adjustable according to the input of the preset classification model. Optionally, the input to the classification model unit module can be the output of intermediate layers of the recognition model or the final output features of the recognition model. Optionally, the selection of the access points for the classification model can be optimized using Neural Architecture Search (NAS) to obtain the optimal access points.

[0050] S3, Cross-entropy Objective Function Calculation Module: This module can take the initial time information output by the classification model and the initial time information labels of the initial image samples as input to calculate the loss information of the preset classification model. The loss function corresponding to this loss information can be the cross-entropy loss function or a regression-type loss function.

[0051] S4. Cross-entropy Objective Function Optimization Module: This module can train and optimize the entire network based on gradient descent. For example, it can use stochastic gradient descent, stochastic gradient descent with a moving average, Adam optimization algorithm, Adamard algorithm, and other gradient descent algorithms. During training, steps S1 to S4 can be repeated until the training results meet the training termination condition. The training termination condition can generally be set by setting the number of iterations to a predetermined value, or by setting the loss information calculated by the cross-entropy objective function to less than a predetermined value. This completes the training of the preset classification model, and the trained classification model is obtained based on the model parameters at this point.

[0052] In step 104, based on the identification information and the identification information label, the identification loss information corresponding to the image sample is determined, and the time loss information is removed from the identification loss information to obtain the target loss information.

[0053] The label loss information can represent the difference between the label information and the label information tag, that is, the error of the recognition model in predicting the label information corresponding to the image sample. Optionally, the label loss information corresponding to the image sample can be obtained by calculating the loss function of the recognition model based on the label information and the label information tag. The target loss information can represent the difference between the recognition model and the optimization target. The optimization target can be the optimization target of decoupling the image features extracted by the recognition model from the classification model during time information training.

[0054] To improve the accuracy of the recognition model in identifying and comparing face images from different time periods, a classification model can be used to decouple the image features extracted by the recognition model from temporal information during training. For example, please refer to... Figure 4 , Figure 4 This is a schematic diagram illustrating a specific process of an information recognition method provided in this application. A temporal information decoupling layer is connected between the recognition model unit module and the classification model unit module. Based on this temporal information decoupling layer, the classification model is used to train the image features extracted by the recognition model to decouple the temporal information. This ensures that the features extracted by the trained recognition model from image samples do not contain temporal information-related features, thereby learning the feature distribution after age decoupling and improving the accuracy of face image recognition and comparison between different time information. Specifically, to ensure that the features extracted by the trained recognition model from image samples do not contain temporal information-related features, temporal loss information can be removed from the labeling loss information to obtain target loss information. This target loss information can then be used as an optimization objective to converge the training process of decoupling the temporal information of the recognition model using the classification model, resulting in a recognition model that can learn the common feature distribution between face images with different time information.

[0055] There are several ways to remove time loss information from the identification loss information. For example, the time loss information can be subtracted from the identification loss information to obtain the target loss information. Optionally, as shown in formula (1), the loss function corresponding to the target loss information can be expressed as follows:

[0056] Where E() represents the loss function corresponding to the target loss information, which is the optimization objective of using the classification model to decouple the image features extracted by the recognition model from the temporal information during training. Representing image features, Labels indicating identification information The label represents the time information. n can represent the total number of image samples in a batch, and i represents the i-th image sample in a batch. This represents the loss function of the recognition model, that is, the loss function corresponding to the identification loss information. This represents the loss function of the classification model, that is, the loss function corresponding to the time loss information. This can represent the hyperparameters corresponding to the loss function of the classification model. The goal of this optimization is to ensure that the image features extracted by the recognition model retain the features of face recognition classification, but do not include features related to distinguishing temporal information. Among these are...

[0057]

[0058] Where argmin() represents making the objective function The variable value that is minimized when the objective function is minimized; argmax() represents the value that makes the objective function... The value of the variable when the maximum value is reached.

[0059] In step 105, the recognition model is converged based on the target loss information to obtain the trained recognition model.

[0060] The trained recognition model can be a pre-trained recognition model, specifically a recognition model obtained by decoupling the image features extracted by the recognition model from the classification model using temporal information. This model can be used to recognize the image to be recognized. The image to be recognized can be a face image that has not undergone information recognition.

[0061] There are several ways to converge the recognition model based on the target loss information. For example, the convergence condition of the target loss information can be determined based on the model information corresponding to the recognition model and the classification model. When the target loss information does not meet the convergence condition, the recognition model can be converged based on the time loss information and the label loss information to obtain the trained recognition model.

[0062] The model information can be attribute information of the recognition model and the classification model. This attribute information can be information that characterizes the properties and relationships of the recognition model and the classification model. For example, it can include the loss functions of the recognition model and the classification model. The convergence condition can be the termination condition for decoupling the image features extracted by the recognition model from the classification model using time information. For example, it can be that the number of iterations meets a set value, or the loss information calculated by the objective function is less than a set value, and the training can be completed. Thus, the trained recognition model is obtained based on the model parameters at this time.

[0063] When the target loss information does not meet the convergence condition, there are multiple ways to converge the recognition model based on the time loss information and the label loss information. For example, the time gradient of the classification model can be updated based on the time loss information to obtain the updated time gradient, and the label gradient of the recognition model can be updated based on the label loss information to obtain the updated label gradient. Based on the updated time gradient and the updated label gradient, the recognition model can be converged to obtain the trained recognition model.

[0064] Here, the temporal gradient can be the gradient corresponding to the classification model, the updated temporal gradient can be the gradient obtained after updating the temporal gradient based on the temporal loss information, the label gradient can be the gradient corresponding to the identification model, and the updated label gradient can be the gradient obtained after updating the label gradient based on the label loss information.

[0065] There are several ways to update the temporal gradient of the classification model based on the temporal loss information. For example, the gradient of the loss function corresponding to the temporal loss information can be calculated based on the temporal loss information, and then the calculated gradient can be used to update the temporal gradient of the classification model to obtain the updated temporal gradient.

[0066] There are several ways to converge the recognition model based on the updated temporal gradient and the updated label gradient. For example, the current number of gradient updates can be obtained, and the updated label gradient can be adjusted based on the number of gradient updates and the updated temporal gradient to obtain the adjusted label gradient. The network parameters of the classification model can be updated according to the updated temporal gradient, and the network parameters of the recognition model can be updated according to the adjusted label gradient to obtain the trained recognition model.

[0067] The gradient update count can be the cumulative number of gradient updates performed by the recognition model and the classification model during the decoupled training of temporal information. The adjusted label gradient can be the result obtained by adjusting the updated label gradient based on the gradient update count and the updated temporal gradient. Updating the network parameters of the recognition model based on the adjusted label gradient can make the recognition model update its parameters in the opposite direction to the updated temporal gradient of the classification network, so as to continuously weaken the extraction of temporal information features.

[0068] There are several ways to adjust the updated identifier gradient based on the number of gradient updates and the updated time gradient. For example, a target time gradient can be selected from the updated time gradient, the gradient adjustment parameter of the target time gradient can be determined according to the number of gradient updates, and the gradient adjustment parameter and the target time gradient can be weighted to obtain the weighted target time gradient. The weighted target time gradient and the updated identifier gradient can be fused to obtain the adjusted identifier gradient.

[0069] Wherein, the target temporal gradient can be the temporal gradient used to adjust the updated identifier gradient in the updated temporal gradient, the gradient adjustment parameter can be the parameter used to adjust the target temporal gradient, and the weighted target temporal gradient can be the result obtained by weighting the gradient adjustment parameter and the target temporal gradient.

[0070] There are several ways to select the target temporal gradient from the updated temporal gradient. For example, the updated temporal gradient corresponding to the network layer connected to the recognition model can be determined as the target temporal gradient. (Please refer to [reference needed] for more details.) Figure 4 The updated temporal gradient of the network layer connected to the information decoupling layer in the classification model of the classification model unit module can be determined as the target temporal gradient.

[0071] After selecting the target temporal gradient from the updated temporal gradients, the gradient adjustment parameter of the target temporal gradient can be determined based on the number of times the gradient is updated. There are several ways to determine the gradient adjustment parameter of the target temporal gradient based on the number of times the gradient is updated. For example, the ratio of the number of times the gradient is updated to the preset update parameter can be calculated to obtain the update frequency ratio. When the update frequency ratio is a preset value type, the ratio of the number of times the gradient is updated to the preset total number of updates can be calculated to obtain the update frequency ratio. The update frequency ratio can be converted into the initial gradient adjustment parameter of the target temporal gradient, and the initial gradient adjustment parameter can be weighted according to the preset negative adjustment coefficient to obtain the gradient adjustment parameter.

[0072] The update frequency ratio can be the ratio of the number of gradient updates to the preset update parameter. The preset update parameter can be a parameter pre-set according to actual needs, used to control the frequency of gradient updates to the recognition model based on the adjusted label gradient. For example, the preset update parameter can be 3, which means that for every 3 updates to the gradient of the classification model, the updated label gradient of the recognition model can be adjusted based on the updated temporal gradient of the classification model, so that the gradient of the recognition model can be updated once based on the adjusted label gradient. It should be noted that when the current gradient update count does not meet the condition for updating the gradient of the recognition model based on the classification model, the network parameters of the classification model can be updated according to the updated temporal gradient, and the network parameters of the recognition model can be updated according to the updated label gradient. The preset numerical type can be a pre-defined numerical type, such as an integer. When the update frequency ratio is an integer, it indicates that the current gradient update count conforms to the frequency of gradient updates to the recognition model based on the adjusted identifier gradient, controlled by the preset update parameters. The ratio of this gradient update count to the preset total update count can be calculated. This preset total update count can be the total number of iterations for time-information decoupling training, set according to actual training needs. The initial gradient adjustment parameter can be the gradient adjustment parameter for the target temporal gradient obtained by converting the update frequency ratio. The preset negative adjustment coefficient can be a pre-defined adjustment coefficient. Based on this preset negative adjustment coefficient, the gradient of the classification model can be applied inversely to the recognition model, ensuring that the features extracted by the trained recognition model from image samples do not contain temporal information-related features. This allows the model to learn the common feature distribution between face images with different time information, improving the accuracy of face image recognition across different time information.

[0073] There are several ways to convert the update ratio into the initial gradient adjustment parameter of the target temporal gradient. For example, the update ratio can be converted into the initial gradient adjustment parameter of the target temporal gradient as shown in formula (2), which can be specifically expressed as follows:

[0074] in, Let represent the initial gradient adjustment parameter, which can be a function that changes with the number of iterations (i.e., the number of gradient updates). Here, k is the total number of iterations (i.e., the preset total number of updates), and p is the number of gradient updates. This represents the ratio of update times. 3 is the preset update parameter, but it can also be set to other values ​​according to actual training needs. exp() represents the exponential function, if represents the conditional function, and % represents the remainder symbol.

[0075] After converting the update ratio into initial gradient adjustment parameters for the target temporal gradient, these initial gradient adjustment parameters can be weighted according to a preset negative adjustment coefficient to obtain the final gradient adjustment parameters. There are several ways to weight these initial gradient adjustment parameters using the preset negative adjustment coefficient; for example, if the preset negative adjustment coefficient is -1, then the gradient adjustment parameter can be expressed as - .

[0076] After determining the gradient adjustment parameter for the target temporal gradient based on the number of gradient updates, the gradient adjustment parameter and the target temporal gradient can be weighted. There are several ways to weight the gradient adjustment parameter and the target temporal gradient; for example, the gradient adjustment parameter can be multiplied by the target temporal gradient to obtain a weighted target temporal gradient of - I, where I represents the target time gradient.

[0077] After weighting the gradient adjustment parameter and the target temporal gradient, the weighted target temporal gradient and the updated label gradient can be fused. There are several ways to fuse the weighted target temporal gradient and the updated label gradient. For example, the weighted target temporal gradient and the updated label gradient can be superimposed on the recognition model, so that the recognition model updates its gradient based on the updated label gradient, and then the gradient update based on the weighted target temporal gradient is added on top.

[0078] For example, please continue to refer to Figure 4 By introducing a temporal information decoupling layer between the recognition model and the classification model, the function of this temporal information decoupling layer is to ensure that the output of the recognition model is consistent with the input of the classification model when the recognition model is performing forward propagation. That is, it ensures that the image features extracted by the recognition model can be input into the classification model without discrimination. In addition, when the classification model is performing gradient updates, the gradient of the classification model back to the recognition model can be adjusted through the temporal information decoupling layer so that the gradient back to the recognition model from the classification model acts inversely on the recognition model. This makes the features extracted by the trained recognition model in the image samples not include temporal information-related features, thereby learning the common feature distribution between face images with different time information and improving the accuracy of face image recognition and comparison between different time information. Optionally, the mathematical expression of this temporal information decoupling layer can be as shown in formula (3):

[0079] Here, X can represent the image features extracted by the recognition model, which is the input to the temporal information decoupling layer. It can represent the image features input to the classification model, that is, the output of the temporal information decoupling layer. The gradient can be fed back to the recognition model through the time-learning decoupling layer, which is also the adjusted identification gradient. I is the updated time gradient, the negative sign (i.e. -1) can be the preset negative adjustment coefficient, and γ can be the initial gradient adjustment parameter.

[0080] After adjusting the updated label gradient based on the number of gradient updates and the updated temporal gradient, the network parameters of the classification model can be updated according to the updated temporal gradient, and the network parameters of the recognition model can be updated according to the adjusted label gradient to obtain the trained recognition model. There are several ways to update the network parameters of the classification model and the recognition model according to the updated label gradient to obtain the trained recognition model. For example, the network parameters of the recognition model can be updated according to the adjusted label gradient to obtain the updated recognition model, and the network parameters of the classification model can be updated according to the updated temporal gradient to obtain the updated classification model. The updated recognition model and the updated recognition model can then be used to update the target loss information to obtain the updated target loss information. Based on the updated target loss information, the updated recognition model can be converged to obtain the trained recognition model.

[0081] Specifically, the updated recognition model can be the recognition model obtained by updating the network parameters of the recognition model based on the adjusted label gradient, and the updated classification model can be the classification model obtained by updating the network parameters of the classification model based on the updated temporal gradient. The updated target loss information can be the target loss information calculated when the updated recognition model and the updated recognition model are decoupled from each other by temporal information during training.

[0082] There are several ways to update the target loss information using the updated recognition model and the updated identification model. For example, the updated recognition model can be used to extract image features from the image sample and determine the updated label information of the image sample based on the extracted updated image features; the updated classification model can be used to classify the image features to obtain the updated time information of the image sample, and the updated time loss information corresponding to the image sample can be determined based on the updated time information and the updated time information label; the updated label loss information corresponding to the image sample can be determined according to the updated label information and the label information label, and the updated time loss information can be removed from the updated label loss information to obtain the updated target loss information.

[0083] The updated image features can be image features obtained by extracting image features from image samples based on the updated recognition model. The image samples obtained by the updated recognition model can be the same as or different from the image samples obtained by the recognition model, and no limitation is made here.

[0084] After updating the target loss information using the updated recognition model and the updated target loss information, the updated recognition model can be converged based on this updated target loss information. There are several ways to converge the updated recognition model based on the updated target loss information. For example, when the updated target loss information meets the convergence condition, the updated recognition model can be used as the trained recognition model. When the updated target loss information does not meet the convergence condition, the updated target loss information can be used as the target loss information, the updated recognition model can be used as the recognition model, and the updated classification model can be used as the classification model. Then, the steps of updating the temporal gradient of the classification model according to the temporal loss information to obtain the updated temporal gradient, and updating the label gradient of the recognition model according to the label loss information to obtain the updated label gradient are repeated until the target loss information meets the convergence condition, thus obtaining the trained recognition model.

[0085] Therefore, please refer to this. Figure 4 The embodiments of this application employ a method such as Figure 4 The differential learning network shown autonomously learns the differences in the distribution of face images across different time periods. This network introduces a temporal information decoupling layer between the recognition and classification models, assisting the recognition model in training for temporal information decoupling. This enables the recognition model to learn the common feature distributions between face images from different time periods after temporal information decoupling, thereby significantly and effectively improving the accuracy of cross-time information face recognition. Furthermore, the information recognition method provided in this application starts from the image features extracted by the recognition model and directly fine-tunes these features, rather than using reweighting or geometric transformations, to match the feature space distribution of face images from different time periods. This effectively improves the accuracy of the recognition model in cross-time information face recognition.

[0086] After the recognition model converges based on the target loss information to obtain the trained recognition model, the image to be recognized can be recognized according to the trained recognition model. There are several ways to recognize the image to be recognized according to the trained recognition model. For example, the image to be recognized can be acquired, and the trained recognition model can be used to extract features from the image to be recognized to obtain the features corresponding to the image to be recognized. The image to be compared can be acquired, and the trained recognition model can be used to extract features from the image to be compared to obtain the features corresponding to the image to be compared. Based on the trained recognition model, the matching coefficient between the features of the image to be recognized and the features of the image to be compared can be identified, and the matching result between the image to be recognized and the image to be compared can be determined according to the matching coefficient.

[0087] The image to be identified can be a face image to be identified, the image feature to be identified can be the image feature corresponding to the image to be identified, the image to be compared can be a face image to be compared, the image feature to be compared can be the image feature corresponding to the image to be compared, the matching coefficient can be a coefficient that characterizes the degree of matching between the image feature to be identified and the image feature to be compared, for example, it can be a coefficient that characterizes the degree of similarity between the image feature to be identified and the image feature to be compared, and the matching result can include information such as whether the image feature to be identified matches or does not match the image feature to be compared.

[0088] For example, when the features of the image to be identified and the features of the image to be compared are the same identification information with different time information, the matching result can be a match between the features of the image to be identified and the features of the image to be compared, or information of the same identity. For example, this trained recognition model can be applied to cross-age face recognition business. For instance, in the application scenario of finding people across ages, the common distribution features between the image to be identified and the image to be compared, except for age information, can be extracted by the trained recognition model. Thus, the features of the image to be identified and the features of the image to be compared can be obtained, and the matching coefficient between the features of the image to be identified and the features of the image to be compared can be identified. This can avoid the influence of age factors on the matching coefficient between face images and improve the accuracy of cross-age face recognition.

[0089] Therefore, please refer to this. Figure 5 , Figure 5This is a schematic diagram of the overall process of an information recognition method provided in this application embodiment. The information recognition method provided in this application embodiment includes a recognition model training stage and a recognition model deployment stage. In the recognition model training stage, a preset recognition model is trained using a training recognition model unit module to obtain a trained recognition model, and a preset classification model is trained using a training classification model unit module to obtain a trained classification model. A temporal information decoupling layer is then introduced to decouple the features extracted by the recognition model from the temporal information using the classification model. Specifically, a small classification model for temporal information estimation is added to the existing recognition model training framework. This classification model distinguishes between different temporal information in the features extracted by the recognition model. During training, the classification model classifies the image features extracted by the recognition model based on temporal information, and normal iterative optimization occurs within the classification model. The introduction of the temporal information decoupling layer, connected to the recognition model, allows for adversarial learning to decouple the image features extracted by the recognition model from the temporal information. This reverses the gradient of the classification model, causing the features extracted by the recognition model to not contain features that can distinguish temporal information, thereby enabling the recognition model to learn the common feature distribution among face images with different temporal information. Therefore, in the information recognition method provided in this application embodiment, there is no need to adjust the proportional configuration between face images with different time information; it is only necessary to keep the various time information consistent. During the training process, the additional classification model occupies negligible amounts of GPU memory and training time, ensuring the lightweight nature of the information recognition method. In the recognition model deployment stage, the classification model does not need to participate in application deployment after training is completed, ensuring that the information recognition method can be used in both large and small models without modifying the original recognition model. Thus, the trained recognition model can be integrated into a conventional face recognition system to recognize face images, improving the accuracy of face comparison with different time information without adding additional deployment modules, thereby improving the efficiency of information recognition.

[0090] As can be seen from the above, the embodiments of this application obtain image samples; use a recognition model to extract image features from the image samples, and determine the identification information of the image samples based on the extracted image features; obtain the time information corresponding to the image samples, and determine the time loss information corresponding to the image samples based on the time information and time information labels; determine the identification loss information corresponding to the image samples according to the identification information and identification information labels, and remove the time loss information from the identification loss information to obtain the target loss information; and converge the recognition model based on the target loss information to obtain the trained recognition model. Therefore, by using a classification model to classify the image features extracted by the recognition model, time loss information is calculated based on the time information obtained from the classification and the time information labels corresponding to the image samples. Then, identification loss information is calculated based on the identification information and labels determined by the recognition model based on the extracted image features. Finally, time loss information is removed from the identification loss information to obtain the target loss information for training the recognition model. This ensures that the image features extracted by the recognition model from the image samples do not contain features with time information. The trained recognition model can then be used to recognize the images to be processed. This improves the accuracy of the recognition model in identifying the identification information of face images with different time information. Simultaneously, it eliminates the need for additional deployment modules, avoiding increased training time and further ensuring the lightweight nature of the recognition model, thereby improving the efficiency of information recognition.

[0091] Based on the method described in the above embodiments, the following examples will provide further detailed explanations.

[0092] In this embodiment, the information recognition device will be specifically integrated into a computer device as an example for explanation. The information recognition method uses a server as the executing entity and takes the time information as age information as an example for specific description. It is understood that in the specific embodiments of this application, user information and other related data are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0093] For a better description of the embodiments of this application, please refer to Figure 6 , Figure 6 Another flowchart illustrating the information recognition method provided in this application embodiment. The specific process is as follows: In step 201, the server obtains an initial image sample and uses a recognition model to extract features from the initial image sample to obtain the initial image features.

[0094] Please continue to refer to the following: Figure 3bWhen the server obtains the initial image sample, it can read the initial image sample, combine the read sample data into a batch, and send it into the deep network unit of the recognition model to extract image features, thereby obtaining the initial image features.

[0095] In step 202, the server uses a preset classification model to classify the initial image features to obtain the initial time information corresponding to the initial image features. Based on the initial time information and the initial time information label, the preset classification model is converged to obtain the classification model.

[0096] The server can predict time information by using the initial image features extracted from the initial image samples by the classification model unit identification module. The category labels can be labels for different age groups. The initial time information output by the preset classification model and the initial time information labels of the initial image samples can be used as input to calculate the loss information of the preset classification model. The loss function corresponding to this loss information can be a cross-entropy loss function or a regression-type loss function. Then, based on the gradient descent method, the entire network of the preset classification model can be trained and optimized until the training result meets the training termination condition. The training termination condition can generally be set to a set number of iterations or a loss calculated by the cross-entropy objective function being less than a set value. The trained classification model is then obtained based on the model parameters at this point.

[0097] In step 203, the server acquires an image sample, uses a recognition model to extract image features from the image sample, and determines the identification information of the image sample based on the extracted image features.

[0098] The image sample carries identification information tags and time information tags.

[0099] The server can acquire image samples in various ways. For example, it can acquire them from a memory connected to the information recognition device, or from other data storage terminals. It can also acquire them from the memory of a physical terminal, or from virtual storage spaces such as datasets or image libraries. Optionally, image samples can be acquired from one storage location or multiple storage locations. For example, image samples can be stored on a blockchain, and the information recognition device can acquire them from the blockchain. The information recognition device can acquire image samples centrally over a period of time in response to an image sample acquisition command, or it can continuously acquire image samples according to a certain image sample acquisition logic.

[0100] In step 204, the server obtains the time information corresponding to the image sample, and determines the time loss information corresponding to the image sample based on the time information and the time information label.

[0101] There are several ways for the server to obtain the time information corresponding to the image sample. For example, the server can use a classification model to predict the time information of the image features and output the probability distribution of which time information the image features belong to. Thus, the time information of the image sample can be determined based on the probability distribution.

[0102] After obtaining the time information corresponding to the image sample, the time loss information corresponding to the image sample can be determined based on the time information and the time information label. There are several ways to determine the time loss information corresponding to the image sample based on the time information and the time information label. For example, the time loss information corresponding to the image sample can be calculated using the cross-entropy loss function based on the time information and the time information label.

[0103] In step 205, the server determines the identification loss information corresponding to the image sample based on the identification information and the identification information label, and removes the time loss information from the identification loss information to obtain the target loss information.

[0104] There are several ways to remove time loss information from the identification loss information. For example, the identification loss information can be subtracted from the time loss information to obtain the target loss information. For example, as shown in formula (1).

[0105] In step 206, the server determines the convergence condition of the target loss information based on the model information corresponding to the recognition model and the classification model. When the target loss information does not meet the convergence condition, the server updates the temporal gradient of the classification model according to the temporal loss information to obtain the updated temporal gradient. The server also updates the label gradient of the recognition model according to the label loss information to obtain the updated label gradient.

[0106] There are several ways for the server to update the temporal gradient of the classification model based on the temporal loss information. For example, the server can calculate the gradient of the loss function corresponding to the temporal loss information based on the temporal loss information, and then update the temporal gradient of the classification model with the calculated gradient to obtain the updated temporal gradient.

[0107] In step 207, the server obtains the current gradient update count, filters the target time gradient from the updated time gradient, calculates the ratio of the gradient update count to the preset update parameter to obtain the update frequency ratio. When the update frequency ratio is a preset numerical type, the server calculates the ratio of the gradient update count to the preset total number of updates to obtain the update count ratio.

[0108] The target temporal gradient can be the temporal gradient used to adjust the updated identifier gradient in the updated temporal gradient.

[0109] There are several ways for the server to select the target temporal gradient from the updated temporal gradient. For example, the server can determine the updated temporal gradient corresponding to the network layer connected to the recognition model as the target temporal gradient. (Please refer to [reference needed] for more details.) Figure 4 The updated temporal gradient of the network layer connected to the information decoupling layer in the classification model of the classification model unit module can be determined as the target temporal gradient.

[0110] In step 208, the server converts the update ratio into an initial gradient adjustment parameter for the target time gradient, and weights the initial gradient adjustment parameter according to a preset negative adjustment coefficient to obtain the gradient adjustment parameter.

[0111] There are several ways to convert the update ratio into the initial gradient adjustment parameter of the target time gradient. For example, the update ratio can be converted into the initial gradient adjustment parameter of the target time gradient as shown in formula (2).

[0112] After the server converts the update count ratio into initial gradient adjustment parameters for the target time gradient, it can weight these initial gradient adjustment parameters according to a preset negative adjustment coefficient. There are several ways to weight these initial gradient adjustment parameters using the preset negative adjustment coefficient; for example, if the preset negative adjustment coefficient is -1, then the gradient adjustment parameter can be expressed as - .

[0113] In step 209, the server weights the gradient adjustment parameter and the target time gradient to obtain the weighted target time gradient, and then fuses the weighted target time gradient with the updated identifier gradient to obtain the adjusted identifier gradient.

[0114] After determining the gradient adjustment parameter for the target temporal gradient based on the number of gradient updates, the server can then weight the gradient adjustment parameter and the target temporal gradient. There are several ways the server can weight the gradient adjustment parameter and the target temporal gradient; for example, the server can multiply the gradient adjustment parameter by the target temporal gradient to obtain a weighted target temporal gradient of - I, where I represents the target time gradient.

[0115] After weighting the gradient adjustment parameter and the target temporal gradient, the server can then fuse the weighted target temporal gradient with the updated identifier gradient. There are several ways the server can fuse the weighted target temporal gradient and the updated identifier gradient. For example, the server can superimpose the weighted target temporal gradient and the updated identifier gradient onto the recognition model, so that the recognition model updates its gradient based on the updated identifier gradient, and then further updates its gradient based on the weighted target temporal gradient.

[0116] For example, please continue to refer to Figure 4 By introducing a temporal information decoupling layer between the recognition model and the classification model, the function of this temporal information decoupling layer is to ensure that the output of the recognition model is consistent with the input of the classification model when the recognition model is performing forward propagation. That is, it ensures that the image features extracted by the recognition model can be input into the classification model without discrimination. In addition, when the classification model performs gradient updates, the gradient of the classification model back to the recognition model through the temporal information decoupling layer can be adjusted so that the gradient back to the recognition model from the classification model acts inversely on the recognition model. In this way, the features extracted by the trained recognition model in the image samples do not contain temporal information-related features, thereby learning the common feature distribution between face images with different time information and improving the accuracy of face image recognition and comparison between different time information. Optionally, the mathematical expression of this temporal information decoupling layer can be as shown in formula (3).

[0117] In step 210, the server updates the network parameters of the recognition model according to the adjusted identifier gradient to obtain the updated recognition model, and updates the network parameters of the classification model according to the updated time gradient to obtain the updated classification model.

[0118] The updated recognition model can be the recognition model obtained by updating the network parameters of the recognition model according to the adjusted label gradient, and the updated classification model can be the classification model obtained by updating the network parameters of the classification model according to the updated time gradient.

[0119] In step 211, the server updates the target loss information using the updated recognition model and the updated recognition model to obtain the updated target loss information. Based on the updated target loss information, the updated recognition model is converged to obtain the trained recognition model.

[0120] The server updates the target loss information using the updated recognition model and the updated identification model. There are several ways to obtain the updated target loss information. For example, the server can use the updated recognition model to extract image features from the image sample and determine the updated label information of the image sample based on the extracted updated image features; use the updated classification model to classify the image features to obtain the updated time information of the image sample, and determine the updated time loss information corresponding to the image sample based on the updated time information and the updated time information label; determine the updated label loss information corresponding to the image sample based on the updated label information and the label information label, and remove the updated time loss information from the updated label loss information to obtain the updated target loss information.

[0121] The updated image features can be image features obtained by extracting image features from image samples based on the updated recognition model. The image samples obtained by the updated recognition model can be the same as or different from the image samples obtained by the recognition model, and no limitation is made here.

[0122] After updating the target loss information using the updated recognition model and the updated target loss information, the server can converge the updated recognition model based on the updated target loss information. There are several ways the server can converge the updated recognition model based on the updated target loss information. For example, if the updated target loss information meets the convergence condition, the server uses the updated recognition model as the trained recognition model. If the updated target loss information does not meet the convergence condition, the server uses the updated target loss information as the target loss information, the updated recognition model as the recognition model, and the updated classification model as the classification model. It then returns to the steps of updating the temporal gradient of the classification model based on the temporal loss information to obtain the updated temporal gradient, and updating the label gradient of the recognition model based on the label loss information to obtain the updated label gradient, until the target loss information meets the convergence condition, thus obtaining the trained recognition model.

[0123] Therefore, please refer to this. Figure 4 The embodiments of this application employ a method such as Figure 4The differential learning network shown autonomously learns the differences in the distribution of face images across different time periods. This network introduces a temporal information decoupling layer between the recognition and classification models, assisting the recognition model in training for temporal information decoupling. This enables the recognition model to learn the common feature distributions between face images from different time periods after temporal information decoupling, thereby significantly and effectively improving the accuracy of cross-time information face recognition. Furthermore, the information recognition method provided in this application starts from the image features extracted by the recognition model and directly fine-tunes these features, rather than using reweighting or geometric transformations, to match the feature space distribution of face images from different time periods. This effectively improves the accuracy of the recognition model in cross-time information face recognition.

[0124] After the updated target loss information is used to converge the updated recognition model and obtain the trained recognition model, the server can recognize the image to be recognized according to the trained recognition model. There are several ways the server can recognize the image to be recognized using the trained recognition model. For example, the server can acquire the image to be recognized, use the trained recognition model to extract features from the image to be recognized, obtain the features corresponding to the image to be recognized, acquire the image to be compared, use the trained recognition model to extract features from the image to be compared, obtain the features corresponding to the image to be compared, identify the matching coefficient between the features of the image to be recognized and the features of the image to be compared based on the trained recognition model, and determine the matching result between the image to be recognized and the image to be compared based on the matching coefficient.

[0125] For example, when the features of the image to be identified and the features of the image to be compared are the same identification information with different time information, the matching result can be a match between the features of the image to be identified and the features of the image to be compared, or information of the same identity. For example, this trained recognition model can be applied to cross-age face recognition business. For instance, in the application scenario of finding people across ages, the server can extract the common distribution features between the image to be identified and the image to be compared, except for age information, through the trained recognition model. In this way, the features of the image to be identified and the features of the image to be compared can be obtained, and the matching coefficient between the features of the image to be identified and the features of the image to be compared can be identified. This can avoid the influence of age factors on the matching coefficient between face images and improve the accuracy of cross-age face recognition.

[0126] As described above, in this embodiment, the server acquires an initial image sample, uses a recognition model to extract features from the initial image sample, and obtains initial image features. The server uses a preset classification model to classify the initial image features, obtaining the initial time information corresponding to the initial image features. Based on the initial time information and the initial time information label, the preset classification model is converged to obtain a classification model. The server acquires an image sample, uses a recognition model to extract image features from the image sample, and determines the identification information of the image sample based on the extracted image features. The server acquires the time information corresponding to the image sample, and determines the 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 based on the identification information and the identification information label, and removes the time loss information from the identification loss information to obtain the target loss information. The server determines the convergence condition of the target loss information based on the model information corresponding to the recognition model and the classification model. When the target loss information does not meet the convergence condition, the time gradient of the classification model is updated based on the time loss information to obtain the updated time gradient. The identification gradient of the recognition model is then updated based on the identification loss information. The process involves updating the gradient to obtain the updated identifier gradient. The server retrieves the current gradient update count, filters the target temporal gradient from the updated temporal gradients, calculates the ratio of this gradient update count to a preset update parameter, obtaining the update frequency ratio. When the update frequency ratio is a preset numerical type, the server calculates the ratio of this gradient update count to a preset total update count, obtaining the update count ratio. The server converts the update count ratio into initial gradient adjustment parameters for the target temporal gradient and weights these initial gradient adjustment parameters according to a preset negative adjustment coefficient, obtaining the gradient adjustment parameters. The server then applies these gradient adjustment parameters and the target temporal gradient... The target temporal gradient is obtained by weighting the target gradient. This weighted target temporal gradient is then fused with the updated label gradient to obtain the adjusted label gradient. The server updates the network parameters of the recognition model based on the adjusted label gradient to obtain the updated recognition model. The server also updates the network parameters of the classification model based on the updated temporal gradient to obtain the updated classification model. The server then updates the target loss information using the updated recognition model and the updated recognition model to obtain the updated target loss information. Based on the updated target loss information, the updated recognition model is converged to obtain the trained recognition model.Therefore, by acquiring the time information obtained by time classification of image features extracted by the recognition model using a classification model and calculating the time information labels corresponding to the image samples, time loss information is calculated. Then, based on the identification information determined by the recognition model based on the extracted image features and the identification information labels, identification loss information is calculated. Finally, time loss information is removed from the identification loss information to obtain target loss information for training the recognition model. This ensures that the image features extracted by the recognition model from the image samples do not contain features with time information. When the target loss information does not meet the convergence condition, gradient updates can be performed on the recognition model based on the identification loss information and time loss information. Iterative training can then be performed based on the updated recognition model. This allows for the recognition of images under test based on the trained recognition model, improving the accuracy of the recognition model in identifying the identification information of face images with different time information. Simultaneously, no additional deployment modules are required, avoiding increased training time and further ensuring the lightweight nature of the recognition model, thereby improving the efficiency of information recognition.

[0127] To better implement the above methods, embodiments of the present invention also provide an information identification device, which can be integrated into a computer device, which can be a server.

[0128] For example, such as Figure 7 The diagram shown is a structural schematic of the information recognition device provided in this application embodiment. The information recognition device may include a sample acquisition unit 301, a feature extraction unit 302, a determination unit 303, a rejection unit 304, and a convergence unit 305, as follows: The sample acquisition unit 301 is used to acquire image samples, which carry identification information tags and time information tags; The feature extraction unit 302 is used to extract image features from the image sample using a recognition model, and to determine the identification information of the image sample based on the extracted image features; The determining unit 303 is used to obtain the time information corresponding to the image sample, and determine the time loss information corresponding to the image sample based on the time information and the time information label; The elimination unit 304 is used to determine the identification loss information corresponding to the image sample based on the identification information and the identification information label, and to remove the time loss information from the identification loss information to obtain the target loss information; The convergence unit 305 is used to converge the recognition model based on the target loss information to obtain a trained recognition model, which is used to recognize the image to be recognized.

[0129] In one embodiment, the convergence unit 305 includes: The convergence condition determination subunit is used to determine the convergence condition of the target loss information based on the model information corresponding to the recognition model and the classification model. The convergence subunit is used to converge the recognition model based on the time loss information and the label loss information when the target loss information does not meet the convergence condition, so as to obtain the trained recognition model.

[0130] In one embodiment, the convergence subunit includes: The gradient update module is used to update the temporal gradient of the classification model based on the temporal loss information to obtain the updated temporal gradient, and to update the label gradient of the recognition model based on the label loss information to obtain the updated label gradient. The update convergence module is used to converge the recognition model based on the updated temporal gradient and the updated label gradient, thus obtaining the trained recognition model.

[0131] In one embodiment, the update convergence module includes: The gradient update count acquisition submodule is used to obtain the current gradient update count; The gradient adjustment submodule is used to adjust the updated identifier gradient based on the number of gradient updates and the updated time gradient to obtain the adjusted identifier gradient. The parameter update submodule is used to update the network parameters of the classification model according to the updated temporal gradient, and to update the network parameters of the recognition model according to the adjusted label gradient, so as to obtain the trained recognition model.

[0132] In one embodiment, the gradient adjustment submodule is used for: The target time gradient is selected from the updated time gradient. This target time gradient is the time gradient used to adjust the updated identifier gradient in the updated time gradient. Based on the number of gradient updates, determine the gradient adjustment parameter of the target temporal gradient, and then weight the gradient adjustment parameter and the target temporal gradient to obtain the weighted target temporal gradient. The weighted target temporal gradient and the updated label gradient are fused to obtain the adjusted label gradient.

[0133] In one embodiment, the gradient adjustment submodule can specifically be used for: Calculate the ratio of the number of gradient updates to the preset update parameters to obtain the update frequency ratio; When the update frequency ratio is a preset numerical type, the ratio of the gradient update count to the preset total update count is calculated to obtain the update count ratio. The ratio of update times is converted into the initial gradient adjustment parameter of the target time gradient, and the initial gradient adjustment parameter is weighted according to the preset negative adjustment coefficient to obtain the gradient adjustment parameter.

[0134] In one embodiment, this parameter update submodule can be specifically used for: Based on the adjusted identifier gradient, the network parameters of the recognition model are updated to obtain the updated recognition model. Based on the updated time gradient, the network parameters of the classification model are updated to obtain the updated classification model. The updated identification model and the updated identification model are used to update the target loss information to obtain the updated target loss information; Based on the updated target loss information, the updated recognition model is converged to obtain the trained recognition model.

[0135] In one embodiment, this parameter update submodule can be specifically used for: When the updated target loss information satisfies the convergence condition, the updated recognition model is used as the trained recognition model. When the updated target loss information does not meet the convergence condition, the updated target loss information is used as the target loss information, the updated recognition model is used as the recognition model, the updated classification model is used as the classification model, and the process of updating the temporal gradient of the classification model based on the temporal loss information to obtain the updated temporal gradient, and updating the label gradient of the recognition model based on the label loss information to obtain the updated label gradient is repeated until the target loss information meets the convergence condition, thus obtaining the trained recognition model.

[0136] In one embodiment, the information identification device further includes: An initial sample acquisition unit is used to acquire an initial image sample, which carries an initial time information label. The initial feature extraction unit is used to extract features from the initial image sample using a recognition model to obtain the initial image features; The initial classification unit is used to classify the initial image features using a preset classification model to obtain the initial time information corresponding to the initial image features; The initial convergence unit is used to converge the preset classification model based on the initial time information and the initial time information label to obtain the classification model.

[0137] In one embodiment, the information identification device further includes: The image feature extraction subunit is used to acquire the image to be identified and to use the trained recognition model to extract features from the image to be identified, thereby obtaining the image features corresponding to the image to be identified. The image feature extraction subunit is used to acquire the image to be compared and to use the trained recognition model to extract features from the image to be compared, thereby obtaining the image features corresponding to the image to be compared. The recognition subunit is used to identify the matching coefficient between the features of the image to be recognized and the features of the image to be compared based on the trained recognition model, and to determine the matching result between the image to be recognized and the image to be compared based on the matching coefficient.

[0138] In practice, each of the above units can be implemented as an independent entity or can be arbitrarily combined to be implemented as the same or several entities. For the specific implementation of each of the above units, please refer to the previous method embodiments, which will not be repeated here.

[0139] As can be seen from the above, in this embodiment of the application, the acquisition unit 301 acquires image samples; the feature extraction unit 302 uses a recognition model to extract image features from the image samples and determines the identification information of the image samples based on the extracted image features; the determination unit 303 acquires the time information corresponding to the image sample and determines the time loss information corresponding to the image sample based on the time information and the time information label; the elimination unit 304 determines the identification loss information corresponding to the image sample based on 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 convergence unit 305 converges the recognition model based on the target loss information to obtain the trained recognition model. Therefore, by using a classification model to classify the image features extracted by the recognition model, time loss information is calculated based on the time information obtained from the classification and the time information labels corresponding to the image samples. Then, identification loss information is calculated based on the identification information and labels determined by the recognition model based on the extracted image features. Finally, time loss information is removed from the identification loss information to obtain the target loss information for training the recognition model. This ensures that the image features extracted by the recognition model from the image samples do not contain features with time information. The trained recognition model can then be used to recognize the images to be processed. This improves the accuracy of the recognition model in identifying the identification information of face images with different time information. Simultaneously, it eliminates the need for additional deployment modules, avoiding increased training time and further ensuring the lightweight nature of the recognition model, thereby improving the efficiency of information recognition.

[0140] This application also provides a computer device, such as... Figure 8 As shown, it illustrates a structural diagram of a computer device involved in an embodiment of this application. This computer device may be a server, specifically: The computer device may include components such as a processor 401 with one or more processing cores, a memory 402 with one or more computer-readable storage media, a power supply 403, and an input unit 404. Those skilled in the art will understand that... Figure 8 The computer device structure shown does not constitute a limitation on the computer device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein: Processor 401 is the control center of the computer device, connecting various parts of the computer device through various interfaces and lines. It performs various functions and processes data by running or executing software programs and / or modules stored in memory 402, and by calling data stored in memory 402. Optionally, processor 401 may include one or more processing cores; preferably, processor 401 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into processor 401.

[0141] The memory 402 can be used to store software programs and modules. The processor 401 executes various functional applications and information recognition by running the software programs and modules stored in the memory 402. The memory 402 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 402 may include high-speed random access memory, and may also include non-volatile memory, such as at least one 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.

[0142] The computer device also includes a power supply 403 that supplies power to the various components. Preferably, the power supply 403 can be logically connected to the processor 401 through a power management system, thereby enabling functions such as charging, discharging, and power consumption management through the power management system. The power supply 403 may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components.

[0143] The computer device may also include an input unit 404, which can be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0144] Although not shown, the computer device may also include a display unit, etc., which will not be described in detail here. Specifically, in this embodiment, the processor 401 in the computer device loads the executable files corresponding to the processes of one or more applications into the memory 402 according to the following instructions, and the processor 401 runs the applications stored in the memory 402 to realize various functions, as follows: Acquire image samples; use a recognition model to extract image features from the image samples, and determine the identification information of the image samples based on the extracted image features; acquire the time information corresponding to the image samples, and determine the time loss information corresponding to the image samples based on the time information and time information labels; determine the identification loss information corresponding to the image samples based on the identification information and identification information labels, and remove the time loss information from the identification loss information to obtain the target loss information; converge the recognition model based on the target loss information to obtain the trained recognition model.

[0145] The specific implementation of each of the above operations can be found in the preceding embodiments, and will not be repeated here. It should be noted that the computer device provided in this application embodiment and the information identification method in the above embodiments belong to the same concept, and its specific implementation process can be found in the above method embodiments, and will not be repeated here.

[0146] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0147] Therefore, embodiments of this application provide a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to execute steps in any of the information recognition methods provided in embodiments of this application. For example, the instructions can execute the following steps: Acquire image samples; use a recognition model to extract image features from the image samples, and determine the identification information of the image samples based on the extracted image features; acquire the time information corresponding to the image samples, and determine the time loss information corresponding to the image samples based on the time information and time information labels; determine the identification loss information corresponding to the image samples based on the identification information and identification information labels, and remove the time loss information from the identification loss information to obtain the target loss information; converge the recognition model based on the target loss information to obtain the trained recognition model.

[0148] The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0149] Since the instructions stored in the computer-readable storage medium can execute the steps of any of the information identification methods provided in the embodiments of this application, the beneficial effects that any of the information identification methods provided in the embodiments of this application can achieve can be realized. For details, please refer to the previous embodiments, which will not be repeated here.

[0150] According to one aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations of the above embodiments.

[0151] The above provides a detailed description of an information identification method, apparatus, and computer-readable storage medium provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. An information identification method, characterized in that, include: Acquire image samples, wherein the image samples carry identification information tags and time information tags; The image sample is used to extract image features using a recognition model, and the identification information of the image sample is determined based on the extracted image features; The time information corresponding to the image sample is obtained, and based on the time information and the time information label, the time loss information corresponding to the image sample is determined. The time information is obtained by performing time classification on the image features using a classification model. Based on the identification information and identification information label, the identification loss information corresponding to the image sample is determined, and the time loss information is removed from the identification loss information to obtain the target loss information; The recognition model is converged based on the target loss information to obtain a trained recognition model. Specifically, the temporal gradient of the classification model is updated according to the temporal loss information to obtain an updated temporal gradient; the label gradient of the recognition model is updated according to the label loss information to obtain an updated label gradient; the current gradient update count is obtained; a target temporal gradient is selected from the updated temporal gradients, and the target temporal gradient is the temporal gradient used to adjust the updated label gradient; the gradient adjustment parameter of the target temporal gradient is determined based on the gradient update count, and the gradient adjustment parameter and the target... The temporal gradient is weighted to obtain a weighted target temporal gradient; the weighted target temporal gradient and the updated label gradient are fused to obtain an adjusted label gradient; the network parameters of the classification model are updated according to the updated temporal gradient, and the network parameters of the recognition model are updated according to the adjusted label gradient to obtain a trained recognition model. This trained recognition model extracts image features from image samples that do not contain temporal information. The trained recognition model is used to recognize the image to be recognized, extracting image features other than temporal information from the image to be recognized to determine the label information of the image to be recognized.

2. The information identification method as described in claim 1, characterized in that, The step of converging the recognition model based on the target loss information to obtain the trained recognition model includes: Based on the model information corresponding to the recognition model and the classification model, the convergence condition of the target loss information is determined; When the target loss information does not meet the convergence condition, the recognition model is converged based on the time loss information and the label loss information to obtain the trained recognition model.

3. The information recognition method as described in claim 1, characterized in that, The step of determining the gradient adjustment parameter of the target temporal gradient based on the number of gradient updates includes: Calculate the ratio of the gradient update count to the preset update parameter to obtain the update frequency ratio; When the update frequency ratio is a preset numerical type, the ratio of the gradient update count to the preset total update count is calculated to obtain the update count ratio; The ratio of update times is converted into the initial gradient adjustment parameter of the target time gradient, and the initial gradient adjustment parameter is weighted according to the preset negative adjustment coefficient to obtain the gradient adjustment parameter.

4. The information recognition method as described in claim 1, characterized in that, The step of updating the network parameters of the classification model according to the updated temporal gradient, and updating the network parameters of the recognition model according to the adjusted label gradient, to obtain the trained recognition model, includes: Based on the adjusted identifier gradient, the network parameters of the recognition model are updated to obtain the updated recognition model. Based on the updated time gradient, the network parameters of the classification model are updated to obtain the updated classification model. The updated target loss information is obtained by updating the target loss information using the updated recognition model and the updated recognition model. Based on the updated target loss information, the updated recognition model is converged to obtain the trained recognition model.

5. The information recognition method as described in claim 4, characterized in that, The step of converging the updated recognition model based on the updated target loss information to obtain the trained recognition model includes: When the updated target loss information satisfies the convergence condition, the updated recognition model is used as the trained recognition model. When the updated target loss information does not meet the convergence condition, the updated target loss information is used as the target loss information, the updated recognition model is used as the recognition model, the updated classification model is used as the classification model, and the process of updating the temporal gradient of the classification model based on the temporal loss information to obtain the updated temporal gradient, and updating the label gradient of the recognition model based on the label loss information to obtain the updated label gradient is repeated until the target loss information meets the convergence condition, thus obtaining the trained recognition model.

6. The information recognition method as described in claim 1, characterized in that, Before obtaining the time information corresponding to the image sample, the method further includes: Acquire an initial image sample, the initial image sample carrying an initial time information label; The initial image features are obtained by using a recognition model to extract features from the initial image samples; The initial image features are classified using a preset classification model to obtain the initial time information corresponding to the initial image features; Based on the initial time information and the initial time information label, the preset classification model is converged to obtain the classification model.

7. The information identification method according to any one of claims 1 to 6, characterized in that, After the recognition model is converged based on the target loss information to obtain the trained recognition model, the process further includes: The image to be identified is obtained, and the trained recognition model is used to extract features from the image to be identified to obtain the image features corresponding to the image to be identified. The image to be compared is obtained, and the trained recognition model is used to extract features from the image to be compared to obtain the features of the image to be compared. Based on the trained recognition model, the matching coefficient between the features of the image to be recognized and the features of the image to be compared is identified, and the matching result between the image to be recognized and the image to be compared is determined according to the matching coefficient.

8. An information identification device, characterized in that, include: A sample acquisition unit is used to acquire image samples, wherein the image samples carry identification information tags and time information tags; The feature extraction unit is used to extract image features from the image sample using a recognition model, and to determine the identification information of the image sample based on the extracted image features; The determining unit is used to acquire the time information corresponding to the image sample, and determine the time loss information corresponding to the image sample based on the time information and the time information label, wherein the time information is the information obtained by performing time classification on the image features using a classification model; The elimination unit is used to determine the identification loss information corresponding to the image sample based on the identification information and the identification information label, and to remove the time loss information from the identification loss information to obtain the target loss information; A convergence unit is used to converge the recognition model based on the target loss information to obtain a post-trained recognition model, such that the image features extracted by the post-trained recognition model in the image samples do not contain features with time information. The post-trained recognition model is used to recognize the image to be recognized, and to extract image features other than features with time information from the image to be recognized to determine the identification information of the image to be recognized. The convergence unit includes: The gradient update module is used to update the temporal gradient of the classification model according to the temporal loss information to obtain the updated temporal gradient, and to update the label gradient of the recognition model according to the label loss information to obtain the updated label gradient; The convergence module is updated to obtain the current gradient update count; a target temporal gradient is selected from the updated temporal gradients, which is the temporal gradient used to adjust the updated label gradient; a gradient adjustment parameter for the target temporal gradient is determined based on the gradient update count, and the gradient adjustment parameter and the target temporal gradient are weighted to obtain a weighted target temporal gradient; the weighted target temporal gradient and the updated label gradient are fused to obtain an adjusted label gradient; the network parameters of the classification model are updated based on the updated temporal gradient, and the network parameters of the recognition model are updated based on the adjusted label gradient to obtain a trained recognition model.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to perform the steps of the information identification method according to any one of claims 1 to 7.

10. A computer device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the information recognition method according to any one of claims 1 to 7.

11. A computer program product, characterized in that, The computer program product includes a computer program / instruction that, when executed by a processor, implements the steps of the information identification method according to any one of claims 1 to 7.

Images