A cross-age face recognition method and system based on representation mapping
By constructing a cross-age face recognition network model based on feature extraction and representation mapping and deploying it on an FPGA platform, the problems of high model complexity and low recognition accuracy in existing methods are solved, and efficient cross-age face recognition is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SUN YAT SEN UNIV
- Filing Date
- 2024-04-19
- Publication Date
- 2026-08-04
AI Technical Summary
Existing cross-age face recognition methods suffer from high model complexity and low recognition accuracy. In particular, discriminative methods are unable to effectively suppress age interference, and existing platforms have high deployment costs.
A cross-age face recognition network model is constructed using a feature extraction network module and an attention module. The model complexity is reduced through feature extraction, representation mapping, and domain separation processing. The optimized model is then deployed on an FPGA platform, and network performance is further optimized using domain separation loss function and identity aggregation loss function.
The model complexity was reduced, the accuracy of cross-age face recognition was improved, and higher operating efficiency and recognition performance were achieved on the FPGA platform.
Smart Images
Figure CN118430040B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of cross-age face recognition technology, and in particular to a cross-age face recognition method and system based on representation mapping. Background Technology
[0002] Existing cross-age face recognition methods can be mainly divided into generative and discriminative methods. The general idea of generative methods is to use image generation models to transform face images of different ages into face images of the same age, and then perform general face recognition. Generative methods can effectively reduce the interference of age factors; however, their model structures are relatively complex, and they can introduce other interfering information during the face image generation process, thus affecting the stability of the recognition results. Unlike generative methods, discriminative methods aim to directly extract age-independent identity information from face images, thereby achieving cross-age face recognition. Existing discriminative methods typically design network structures based on linear assumptions to separate age and identity information from facial features. However, because the aging process of the face is influenced by many factors such as innate genes and acquired growth environment, there is a complex relationship between facial identity information and age information, making it difficult for existing discriminative methods to effectively suppress age interference.
[0003] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the Invention
[0004] The main objective of this application is to propose a cross-age face recognition method and system based on representation mapping, which can simplify the complexity of cross-age face recognition network models and improve the accuracy of cross-age face recognition.
[0005] To achieve the above objectives, one aspect of this application proposes a cross-age face recognition method based on representation mapping, the method comprising:
[0006] A cross-age face recognition network model is constructed based on a feature extraction network module and an attention module;
[0007] Based on the cross-age face recognition network model, feature extraction and representation mapping are performed sequentially on the face image to obtain the cross-age face recognition result;
[0008] Based on the cross-age face recognition results, the cross-age face recognition network model is optimized to obtain an optimized cross-age face recognition network model.
[0009] The optimized cross-age face recognition network model was deployed to an FPGA platform to construct a cross-age face recognition network model based on the FPGA platform.
[0010] In some embodiments, the cross-age face recognition network model includes a feature extraction network module, an attention module, an identity domain representation mapping network module, and an age domain representation mapping network module. The output of the feature extraction network module is connected to the input of the attention module, the first output of the attention module is connected to the input of the identity domain representation mapping network module, and the second output of the attention module is connected to the input of the age domain representation mapping network module, wherein:
[0011] The feature extraction network module consists of a deep residual network ResNet34 with fully connected layers removed;
[0012] The attention module includes an average pooling layer, a max pooling layer, a convolutional layer, and an activation layer;
[0013] Both the identity domain representation mapping network module and the age domain representation mapping network module are composed of fully connected layers.
[0014] In some embodiments, the step of sequentially performing feature extraction and representation mapping processing on the face image based on the cross-age face recognition network model to obtain the cross-age face recognition result includes:
[0015] Acquire facial images;
[0016] The facial image is input into the cross-age facial recognition network model;
[0017] Based on the feature extraction network module of the cross-age face recognition network model, feature extraction processing is performed on the face image to obtain face hybrid features;
[0018] Based on the attention module of the cross-age face recognition network model, the face hybrid features are preprocessed to obtain the identity features and age features of the face image.
[0019] Based on the identity domain representation mapping network module of the cross-age face recognition network model, the identity features of the face image are processed by domain representation mapping to obtain the identity domain representation of the face image.
[0020] Based on the age domain representation mapping network module of the cross-age face recognition network model, the age features of the face image are processed by domain representation mapping to obtain the age domain representation of the face image.
[0021] The identity domain representation and age domain representation of the face image are separated by a domain separation loss function, and the cross-age face recognition result is output.
[0022] In some embodiments, the attention module based on the cross-age face recognition network model performs feature preprocessing on the face blending features to obtain the identity features and age features of the face image, including:
[0023] The facial hybrid features are input into the attention module of the cross-age facial recognition network model;
[0024] Based on the average pooling layer of the attention module, the face blending features are subjected to average pooling processing to obtain the face blending feature average pooling result;
[0025] Based on the max pooling layer of the attention module, the face blending features are subjected to max pooling to obtain the face blending feature max pooling result;
[0026] The average pooling result of the face blending features is concatenated with the max pooling result of the face blending features to obtain the concatenated face blending feature pooling result;
[0027] Based on the convolutional and activation layers of the attention module, the pooling result of the concatenated face hybrid features is subjected to convolutional activation processing to obtain the attention mask features;
[0028] The attention mask features and the face blending features are weighted and calculated to obtain the identity features of the face image;
[0029] The age feature of the face image is obtained by inverting the attention mask feature and then performing a weighted calculation with the face blending feature.
[0030] In some embodiments, the expression for the domain separation loss function is as follows:
[0031]
[0032] In the above formula, L DD Let C represent the domain separation loss function, and C represent the total number of training sample classes. This represents the square of the number of samples in class c. This represents the identity domain representation of the i-th sample. This represents the age domain of the i-th sample.
[0033] In some embodiments, optimizing the cross-age face recognition network model based on the cross-age face recognition results to obtain an optimized cross-age face recognition network model includes:
[0034] The identity features of the face image are constrained by the identity loss function to obtain the identity feature constraint result;
[0035] The age features of the face image are constrained by the age loss function to obtain the age feature constraint result;
[0036] The identity domain representation of the face image is constrained by the identity aggregation loss function to obtain the identity domain representation constraint result.
[0037] The age domain representation of the face image is constrained by the age aggregation loss function to obtain the age domain representation constraint result.
[0038] Based on the identity feature constraint results and the age feature constraint results, the feature extraction network module and the attention module of the cross-age face recognition network model are updated by backpropagation of errors to obtain the optimized feature extraction network module and the optimized attention module.
[0039] Based on the identity domain representation constraint results and the age domain representation constraint results, the identity domain representation mapping network module and the age domain representation mapping network module of the cross-age face recognition network model are updated by error backpropagation to obtain the optimized identity domain representation mapping network module and the optimized age domain representation mapping network module.
[0040] By integrating the optimized feature extraction network module, the optimized attention module, the optimized identity domain representation mapping network module, and the optimized age domain representation mapping network module, the optimized cross-age face recognition network model is obtained.
[0041] In some embodiments, the expressions for the identity loss function and the age loss function are as follows:
[0042]
[0043]
[0044] In the above formula, L I L represents the identity loss function. A Let represent the age loss function, N represent the number of training samples, s represent the amplification factor, m represent the redundant parameter, and cos(θ) represent the loss function. j ,i) represents the cosine of the angle between the weight of the i-th feature and the weight of the j-th class, T represents the total number of age categories, and y t Indicates the actual age label, p t This represents the probability value of the age feature after Softmax calculation. θ represents the angle between the features of the i-th sample and the weights of its true label class. j This represents the angle between the features of the i-th sample and the weights of the j-th class. Let represent the cosine of the angle between the features of the i-th sample and the weights of its true label class.
[0045] In some embodiments, the expressions for the identity aggregation loss function and the age aggregation loss function are specifically as follows:
[0046]
[0047]
[0048] In the above formula, L IP L represents the identity aggregation loss function. AP Let f represent the age aggregation loss function, where n represents the total number of features, and f is the sum of the values of n and n. I f represents the identity domain representation mapping. A This represents the mapping of the age domain. This represents the i-th identity feature. Let m represent the i-th age characteristic. I m represents the average vector of all identity features. A A vector representing the average values of all age characteristics.
[0049] In some embodiments, deploying the optimized cross-age face recognition network model to an FPGA platform to construct an FPGA-based cross-age face recognition network model includes:
[0050] The optimized cross-age face recognition network model is quantized using the Vitis AI quantizer and the AdaQuant algorithm to generate the xmodel model file.
[0051] The xmodel model file is cross-compiled and deployed to the FPGA platform to build a cross-age face recognition network model based on the FPGA platform.
[0052] To achieve the above objectives, another aspect of this application proposes a cross-age face recognition system based on representation mapping, the system comprising:
[0053] The first module is used to build a cross-age face recognition network model based on the feature extraction network module and the attention module;
[0054] The second module is used to sequentially perform feature extraction and representation mapping processing on the face image based on the cross-age face recognition network model to obtain the cross-age face recognition result.
[0055] The third module is used to optimize the cross-age face recognition network model based on the cross-age face recognition results to obtain an optimized cross-age face recognition network model.
[0056] The fourth module is used to deploy the optimized cross-age face recognition network model to the FPGA platform to build a cross-age face recognition network model based on the FPGA platform.
[0057] The embodiments of this application include at least the following beneficial effects: This application provides a cross-age face recognition method and system based on representation mapping. This scheme introduces a feature extraction network module and an attention module to construct a cross-age face recognition network model, which reduces the model complexity and is more conducive to the deployment of the model on the FPGA platform. Furthermore, based on the cross-age face recognition network model, feature extraction and representation mapping are performed on the face image in sequence. By adopting the method of separating the identity domain and the age domain, the difference between the identity features and age features extracted from the face image is expanded, effectively suppressing the age information in the identity features, improving the accuracy of cross-age face recognition, and optimizing the cross-age face recognition network model. Finally, the optimized cross-age face recognition network model is deployed to the FPGA platform, which can achieve better operating efficiency while ensuring the performance of cross-age face recognition. Attached Figure Description
[0058] Figure 1 This is a flowchart of a cross-age face recognition method based on representation mapping provided in an embodiment of this application;
[0059] Figure 2 This is a schematic diagram of the cross-age face recognition network structure provided in the embodiments of this application;
[0060] Figure 3 This is a schematic diagram of the structure of a cross-age face recognition system based on representation mapping provided in an embodiment of this application. Detailed Implementation
[0061] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of systems and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.
[0062] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various concepts, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to a determination” as used herein may be interpreted as “when…” or “when…” or “in response to a determination.”
[0063] As used in this application, the terms "at least one", "multiple", "each", "any", etc., "at least one" includes one, two or more, "multiple" includes two or more, "each" refers to each of the corresponding multiples, and "any" refers to any one of the multiples.
[0064] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.
[0065] Before providing a detailed description of the embodiments of this application, the terms involved in the embodiments of this application will be explained first. The terms involved in the embodiments of this application are subject to the following interpretations.
[0066] Cross-age face recognition is a major challenge in the field of face recognition. It aims to use computer vision technology to identify individuals from two facial images of different ages, determining whether they belong to the same person. Because facial aging causes significant changes in facial features, it introduces substantial intra-class variability into facial images, degrading the performance of general face recognition algorithms. Therefore, the primary goal of cross-age face recognition methods is to mitigate the impact of age on facial feature extraction and recognition.
[0067] Some shortcomings exist in related technologies. Current cross-age face recognition methods can be mainly divided into generative and discriminative methods. The general idea of generative methods is to use image generation models to transform face images of different ages into face images of the same age, and then perform general face recognition. Generative methods can effectively reduce the interference of age factors; however, the model structure of generative methods is relatively complex, and other interference information may be introduced during the face image generation process, thus affecting the stability of the recognition results. Unlike generative methods, discriminative methods aim to directly extract age-independent identity information from face images, thereby achieving cross-age face recognition. Existing discriminative methods typically design network structures based on linear assumptions to separate age information and identity information in facial features. However, because the aging process of the face is affected by many factors such as innate genes and acquired growth environment, there is a complex relationship between facial identity information and age information, making it difficult for existing discriminative methods to effectively suppress age interference.
[0068] Regarding the operating platform, deploying cross-age face recognition algorithms on a CPU platform results in low CPU computational efficiency, making real-time output of recognition results difficult. Deploying them on a GPU platform increases deployment and operating costs due to GPU's high power consumption and large size. FPGAs, with their high computational efficiency, low average power consumption, strong compatibility, and small size, are suitable for large-scale deployment of cross-age face recognition algorithms. However, considering model conversion and module support issues, the model structure deployed on an FPGA platform should not be too complex.
[0069] In view of this, this application provides a cross-age face recognition method based on representation mapping, which improves the accuracy of cross-age face recognition results by reducing the complexity of the network structure and providing a feasible FPGA platform deployment scheme.
[0070] Reference Figure 1 , Figure 1 A flowchart illustrating a cross-age face recognition method based on representation mapping provided in this embodiment of the invention is shown below. Figure 1 The method includes the following steps:
[0071] S100. Construct a cross-age face recognition network model based on the feature extraction network module and the attention module;
[0072] In some specific embodiments, such as Figure 2As shown, the cross-age face recognition network model includes a feature extraction network module, an attention module, an identity domain representation mapping network module, and an age domain representation mapping network module. The output of the feature extraction network module is connected to the input of the attention module. The first output of the attention module is connected to the input of the identity domain representation mapping network module, and the second output of the attention module is connected to the input of the age domain representation mapping network module. The feature extraction network module is composed of a deep residual network ResNet34 with fully connected layers removed. The attention module includes average pooling layers, max pooling layers, convolutional layers, and activation layers. Both the identity domain representation mapping network module and the age domain representation mapping network module are composed of fully connected layers.
[0073] It should be noted that the feature extraction network consists of a deep residual network ResNet34 with fully connected layers removed, and was pre-trained on a face recognition dataset. The face image is input into the feature extraction network, and the output of the last convolutional layer is taken as a hybrid feature x containing facial identity information and facial age information. M .
[0074] S200. Based on the cross-age face recognition network model, feature extraction and representation mapping are performed on the face image in sequence to obtain the cross-age face recognition result;
[0075] It should be noted that in some embodiments, step S200 may include: S210, acquiring a face image; S220, inputting the face image into a cross-age face recognition network model; S230, performing feature extraction processing on the face image based on the feature extraction network module of the cross-age face recognition network model to obtain mixed face features; S240, performing feature preprocessing on the mixed face features based on the attention module of the cross-age face recognition network model to obtain the identity features and age features of the face image; S250, performing domain representation mapping processing on the identity features of the face image based on the identity domain representation mapping network module of the cross-age face recognition network model to obtain the identity domain representation of the face image; S260, performing domain representation mapping processing on the age features of the face image based on the age domain representation mapping network module of the cross-age face recognition network model to obtain the age domain representation of the face image; S270, performing separation processing on the identity domain representation and age domain representation of the face image through a domain separation loss function, and outputting the cross-age face recognition result.
[0076] In this embodiment, a face image is input into a feature extraction network to obtain a hybrid feature x that includes facial identity information and facial age information. M , mix features x M The input is fed into the attention module, which learns to generate an attention mask M. I Focus will be placed on mask M. I For mixed features xM After weighting, the identity feature x is obtained I Focus will be on mask M I Inverting the x-axis yields the mixed feature x. M Weighted age feature x A Using the identity loss function L I For identity feature x I Constraints are applied using the age loss function L. A Age characteristic x A By imposing constraints and updating the parameters of the feature extraction network and the focus module through error backpropagation, the performance of the network in extracting identity and age features is improved. This is achieved by assigning identity feature x... I and age characteristics x A Perform domain representation mapping to obtain identity domain representations D respectively. I and age domain characterization D A Using the identity aggregation loss function L IP and age aggregation loss function L AP Constraints are applied to the identity domain representation mapping network and the age domain representation mapping network respectively. The domain separation loss function L is then used. DD Identity domain representation D I and age domain characterization D A Separating the features, the parameters of the feature extraction network and the focus module are updated through error backpropagation to optimize the network's performance in extracting identity and age features. Finally, the trained network model is used to extract features from the input face image, removing age information from the identity feature x. I Used for facial feature comparison and recognition, enabling cross-age facial recognition.
[0077] Additionally, it should be noted that the domain separation loss function L... DD The aim is to expand the obtained identity domain representation D I and age domain characterization D A The domain separation loss function L maximizes the difference between identity and age features by identifying the differences between domains. DD The expression is:
[0078]
[0079] In the above formula, L DD Let C represent the domain separation loss function, and C represent the total number of training sample classes. This represents the square of the number of samples in class c. This represents the identity domain representation of the i-th sample. This represents the age domain of the i-th sample.
[0080] During network training, the network parameters for identity domain representation mapping and age domain representation mapping are fixed, while the parameters of the feature extraction module and the focus module are updated to maximize the difference between the extracted identity features and age features. Finally, the identity feature x extracted by the network is used. I Used for facial feature comparison and recognition, thereby enabling cross-age facial recognition.
[0081] It should be noted that in some embodiments, step S240 may include: S241, inputting the face blending features into the attention module of the cross-age face recognition network model; S242, performing average pooling processing on the face blending features based on the average pooling layer of the attention module to obtain the average pooling result of the face blending features; S243, performing max pooling processing on the face blending features based on the max pooling layer of the attention module to obtain the max pooling result of the face blending features; S244, concatenating the average pooling result of the face blending features with the max pooling result of the face blending features to obtain the concatenated face blending feature pooling result; S245, performing convolution activation processing on the concatenated face blending feature pooling result based on the convolutional layer and activation layer of the attention module to obtain the attention mask features; S246, performing weighted calculation processing on the attention mask features and the face blending features to obtain the identity features of the face image; S247, performing weighted calculation processing on the attention mask features after inverting them and the face blending features to obtain the age features of the face image.
[0082] In some specific embodiments, the attention module consists of an average pooling layer, a max pooling layer, a convolutional layer, and an activation layer. The attention module first processes the mixed feature x. M Perform average pooling and max pooling along the channel dimension, concatenate the two pooling results, pass them through a 7×7 convolution, and then apply sigmoid activation to obtain the attention mask M. I Mixed features x M With attention mask M I Perform corresponding bit multiplication to obtain identity feature x I Subtract the attention mask M from a matrix of all 1s of the same dimension. I After and mixed features x M Perform corresponding part-wise multiplication to obtain the age feature x. A .
[0083] S300. Based on the cross-age face recognition results, the cross-age face recognition network model is optimized to obtain the optimized cross-age face recognition network model.
[0084] It should be noted that in some embodiments, step S300 may include: S310, constraining the identity features of the face image using an identity loss function to obtain an identity feature constraint result; S311, constraining the age features of the face image using an age loss function to obtain an age feature constraint result; S312, constraining the identity domain representation of the face image using an identity aggregation loss function to obtain an identity domain representation constraint result; S313, constraining the age domain representation of the face image using an age aggregation loss function to obtain an age domain representation constraint result; S314, based on the identity feature constraint result and the age feature constraint result, optimizing the features of the cross-age face recognition network model. S315. Based on the identity domain representation constraint results and age domain representation constraint results, the identity domain representation mapping network module and age domain representation mapping network module of the cross-age face recognition network model are updated by error backpropagation to obtain the optimized identity domain representation mapping network module and age domain representation mapping network module; S316. The optimized feature extraction network module, optimized attention module, optimized identity domain representation mapping network module, and optimized age domain representation mapping network module are integrated to obtain the optimized cross-age face recognition network model.
[0085] In some specific embodiments, the identity loss function L I The CosFace loss is used, and its expression is as follows:
[0086]
[0087] Age loss function L A The cross-entropy loss is used, and its expression is as follows:
[0088]
[0089] In the above formula, L I L represents the identity loss function. A Let represent the age loss function, N represent the number of training samples, s represent the amplification factor, m represent the redundant parameter, and cos(θ) represent the loss function. j ,i) represents the cosine of the angle between the weight of the i-th feature and the weight of the j-th class, T represents the total number of age categories, and y t Indicates the actual age label, p t This represents the probability value of the age feature after Softmax calculation. θ represents the angle between the features of the i-th sample and the weights of its true label class. j This represents the angle between the features of the i-th sample and the weights of the j-th class. Let represent the cosine of the angle between the features of the i-th sample and the weights of its true label class.
[0090] identity loss function L I and age loss function L A To train the loss function, update the parameters of the feature extraction network and the attention module.
[0091] Furthermore, the domain representation mapping network consists of a fully connected layer. The identity aggregation loss function L... IP The purpose is to supervise the identity domain representation mapping f I To learn the identity domain representation, its expression is:
[0092]
[0093] Age aggregation loss function L AP The purpose is to supervise the age domain representation mapping f A To learn the age domain representation, its expression is:
[0094]
[0095] In the above formula, L IP L represents the identity aggregation loss function. AP Let f represent the age aggregation loss function, where n represents the total number of features, and f is the sum of the values of n and n. I f represents the identity domain representation mapping. A This represents the mapping of the age domain. This represents the i-th identity feature. Let m represent the i-th age characteristic. I m represents the average vector of all identity features. A A vector representing the average values of all age characteristics.
[0096] During network training, the parameters of the feature extraction network and the attention module are fixed, and the identity aggregation loss function L is used respectively. IP and age aggregation loss function L AP As the training loss function, supervise the training of the identity domain representation mapping f I and age domain representation mapping f A The network parameters. After training, the identity features x... I and age characteristics x A Perform identity domain representation mapping and age domain representation mapping respectively to obtain the identity domain representation D. I and age domain characterization D A .
[0097] S400. Deploy the optimized cross-age face recognition network model to the FPGA platform to build a cross-age face recognition network model based on the FPGA platform.
[0098] It should be noted that in some embodiments, step S400 may include: S410, quantizing the optimized cross-age face recognition network model using the Vitis AI quantizer and the AdaQuant algorithm to generate an xmodel model file; S420, cross-compiling the xmodel model file and deploying it to the FPGA platform to build a cross-age face recognition network model based on the FPGA platform.
[0099] In some specific embodiments, the Vitis AI quantizer is used to quantize the model, employing the AdaQuant algorithm to generate an xmodel model file. The main function of the preprocessing script is face detection and alignment, while the main function of the postprocessing script is feature comparison and recognition result output. The model file, along with the preprocessing and postprocessing scripts, is cross-compiled to generate an executable file, which is then deployed to FPGA platforms such as ZCU102 to achieve cross-age face recognition based on the FPGA platform.
[0100] Please see Figure 3 This application also provides a cross-age face recognition system based on representation mapping, which can implement the above-mentioned cross-age face recognition method based on representation mapping. The system includes:
[0101] The first module is used to build a cross-age face recognition network model based on the feature extraction network module and the attention module;
[0102] The second module is used to perform feature extraction and representation mapping on face images sequentially based on a cross-age face recognition network model to obtain cross-age face recognition results.
[0103] The third module is used to optimize the cross-age face recognition network model based on the cross-age face recognition results, so as to obtain the optimized cross-age face recognition network model.
[0104] The fourth module is used to deploy the optimized cross-age face recognition network model to the FPGA platform, and build a cross-age face recognition network model based on the FPGA platform.
[0105] It is understood that the content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.
[0106] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.
Claims
1. A cross-age face recognition method based on representation mapping, characterized in that, The method includes: A cross-age face recognition network model is constructed based on a feature extraction network module and an attention module; Based on the cross-age face recognition network model, feature extraction and representation mapping are performed sequentially on the face image to obtain the cross-age face recognition result; Based on the cross-age face recognition results, the cross-age face recognition network model is optimized to obtain an optimized cross-age face recognition network model. The optimized cross-age face recognition network model is deployed on an FPGA platform to construct a cross-age face recognition network model based on the FPGA platform. The cross-age face recognition network model includes a feature extraction network module, an attention module, an identity domain representation mapping network module, and an age domain representation mapping network module. The output of the feature extraction network module is connected to the input of the attention module. The first output of the attention module is connected to the input of the identity domain representation mapping network module, and the second output of the attention module is connected to the input of the age domain representation mapping network module. The feature extraction network module consists of a deep residual network ResNet34 with fully connected layers removed; The attention module includes an average pooling layer, a max pooling layer, a convolutional layer, and an activation layer; Both the identity domain representation mapping network module and the age domain representation mapping network module are composed of fully connected layers. The process of extracting features and mapping representations from face images sequentially based on the cross-age face recognition network model to obtain cross-age face recognition results includes: Acquire facial images; The facial image is input into the cross-age facial recognition network model; Based on the feature extraction network module of the cross-age face recognition network model, feature extraction processing is performed on the face image to obtain face hybrid features; Based on the attention module of the cross-age face recognition network model, the face hybrid features are preprocessed to obtain the identity features and age features of the face image. Based on the identity domain representation mapping network module of the cross-age face recognition network model, the identity features of the face image are processed by domain representation mapping to obtain the identity domain representation of the face image. Based on the age domain representation mapping network module of the cross-age face recognition network model, the age features of the face image are processed by domain representation mapping to obtain the age domain representation of the face image. The identity domain representation and age domain representation of the face image are separated by a domain separation loss function, and the cross-age face recognition result is output.
2. The method according to claim 1, characterized in that, The attention module based on the cross-age face recognition network model performs feature preprocessing on the face blending features to obtain the identity features and age features of the face image, including: The facial hybrid features are input into the attention module of the cross-age facial recognition network model; Based on the average pooling layer of the attention module, the face blending features are subjected to average pooling processing to obtain the face blending feature average pooling result; Based on the max pooling layer of the attention module, the face blending features are subjected to max pooling to obtain the face blending feature max pooling result; The average pooling result of the face blending features is concatenated with the max pooling result of the face blending features to obtain the concatenated face blending feature pooling result; Based on the convolutional and activation layers of the attention module, the pooling result of the concatenated face hybrid features is subjected to convolutional activation processing to obtain the attention mask features; The attention mask features and the face blending features are weighted and calculated to obtain the identity features of the face image; The age feature of the face image is obtained by inverting the attention mask feature and then performing a weighted calculation with the face blending feature.
3. The method according to claim 2, characterized in that, The specific expression for the domain separation loss function is as follows: In the above formula, Domain separation loss function, This represents the total number of training sample categories. Indicates the first The square of the number of samples in the class, Indicates the first Identity domain representation of each sample Indicates the first Age domain representation of each sample.
4. The method according to claim 1, characterized in that, The optimization of the cross-age face recognition network model based on the cross-age face recognition results to obtain an optimized cross-age face recognition network model includes: The identity features of the face image are constrained by the identity loss function to obtain the identity feature constraint result; The age features of the face image are constrained by the age loss function to obtain the age feature constraint result; The identity domain representation of the face image is constrained by the identity aggregation loss function to obtain the identity domain representation constraint result. The age domain representation of the face image is constrained by the age aggregation loss function to obtain the age domain representation constraint result. Based on the identity feature constraint results and the age feature constraint results, the feature extraction network module and the attention module of the cross-age face recognition network model are updated by backpropagation of errors to obtain the optimized feature extraction network module and the optimized attention module. Based on the identity domain representation constraint results and the age domain representation constraint results, the identity domain representation mapping network module and the age domain representation mapping network module of the cross-age face recognition network model are updated by error backpropagation to obtain the optimized identity domain representation mapping network module and the optimized age domain representation mapping network module. By integrating the optimized feature extraction network module, the optimized attention module, the optimized identity domain representation mapping network module, and the optimized age domain representation mapping network module, the optimized cross-age face recognition network model is obtained.
5. The method according to claim 4, characterized in that, The expressions for the identity loss function and the age loss function are as follows: In the above formula, Represents the identity loss function. Represents the age loss function, Indicates the number of training samples. Indicates the magnification factor. Indicates redundant parameters. Indicates the first The first feature and the second The cosine of the angle between the weights of each class. This represents the total number of age categories. This indicates the actual age label. This represents the probability value of the age feature after Softmax calculation. Indicates the first The angle between the features of a sample and the weights of its true label class. Indicates the first The characteristics of the sample and the first The angle between the weights of each class. Indicates the first The cosine of the angle between the features of a sample and the weights of its true label class.
6. The method according to claim 4, characterized in that, The expressions for the identity aggregation loss function and the age aggregation loss function are as follows: In the above formula, This represents the identity aggregation loss function. This represents the age aggregation loss function. Represents the total number of features. Represents the identity domain representation mapping. This represents the mapping of the age domain. Indicates the first Individual identity characteristics Indicates the first Age characteristics, A vector representing the average of all identity features. A vector representing the average values of all age characteristics.
7. The method according to claim 1, characterized in that, The step of deploying the optimized cross-age face recognition network model to an FPGA platform to construct a cross-age face recognition network model based on the FPGA platform includes: The optimized cross-age face recognition network model is quantized using the Vitis AI quantizer and the AdaQuant algorithm to generate the xmodel model file. The xmodel model file is cross-compiled and deployed to the FPGA platform to build a cross-age face recognition network model based on the FPGA platform.
8. A system for implementing the cross-age face recognition method based on representation mapping as described in any one of claims 1-7, characterized in that, The system includes: The first module is used to build a cross-age face recognition network model based on the feature extraction network module and the attention module; The second module is used to sequentially perform feature extraction and representation mapping processing on the face image based on the cross-age face recognition network model to obtain the cross-age face recognition result. The third module is used to optimize the cross-age face recognition network model based on the cross-age face recognition results to obtain an optimized cross-age face recognition network model. The fourth module is used to deploy the optimized cross-age face recognition network model to the FPGA platform to build a cross-age face recognition network model based on the FPGA platform.