Identity-preserved based cross-style face prototype reconstruction technology

Abstract

The application discloses a cross-style face prototype reconstruction technology based on identity preservation, comprising the following steps: S1, establishing a cross-style face prototype reconstruction model based on GAN; S2, establishing a loss function of the model to guide the training process of the model, so that the generator can preserve the original identity features of the picture while converting the styles of face pictures in different domains; S3, inputting face picture training sets with identity labels from domains A and B into the model for model training, constructing a test set, and checking the face prototype reconstruction effect of the trained model on the test face pictures. The application overcomes the shortcoming that the existing cross-style face generation method cannot simultaneously process style migration and face change elimination, can realize cross-spectrum and cross-artistic style face prototype reconstruction, and avoids the difficulty in face change modeling.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a cross-style face prototyping technique based on identity preservation. Background Technology

[0002] In criminal investigations, photos of suspects may be taken from infrared cameras, resulting in a near-infrared style, or be sketches provided by professional artists based on verbal descriptions from eyewitnesses. Because these near-infrared or sketched facial images differ significantly in style from visible light ID photos in the registered database, they greatly increase the difficulty of identifying the suspect's face. Furthermore, suspects may exhibit facial expressions, large head postures, or partial facial obscuration during the photographing or witnessing process; these complex facial changes also severely interfere with facial recognition.

[0003] In recent years, deep convolutional neural networks have made great strides in many image processing tasks. Among them, generative adversarial networks (GANs) have been widely used for image generation due to their powerful generative capabilities. GANs were proposed in 2014 by Ian Goodfellow and other scholars at the University of Montreal, and were praised by Turing Award winner Yann LeCun as "the coolest idea in machine learning in the last two decades." The main structure of a GAN consists of a generator G and a discriminator D. The generator and discriminator learn alternately through a game-like interaction, prompting the generator to produce incredibly realistic images. Specifically, the generator randomly samples from the feature space as input, and its output should closely mimic real samples from the training set. The discriminator's input is either a real sample or a virtual sample output by the generator, and its purpose is to distinguish between real and virtual samples. The generator and discriminator compete against each other, constantly adjusting the network parameters, until eventually the discriminator can no longer determine whether the generator's output is a real or generated sample, thus reaching a "Nash equilibrium" state.

[0004] Cycle-GAN is an improved model of GAN, adding a backward mapping process to the forward mapping of GAN, thus constructing a recurrent network structure. Cycle-GAN believes that the image-to-image translation process should maintain "cyclic consistency," that is, constraining an image to remain consistent with the original image in content after both forward and backward mappings. Most existing cross-style face generation methods are based on the Cycle-GAN structure, focusing on how to perform style transfer on face images while preserving facial details. Therefore, when the input face image contains facial variations, these methods cannot reconstruct cross-domain, cross-style prototype face images (i.e., frontal, undisturbed, standard face photos under uniform lighting).

[0005] Therefore, in order to facilitate criminal investigation experts to accurately match and identify facial images with different styles (or modalities) and facial variations, this invention provides an identity-preserving cross-style facial prototype reconstruction technology that can simultaneously transfer the target facial style and eliminate facial variations, restoring its visible light style prototype image. Summary of the Invention

[0006] The main objective of this invention is to overcome the shortcomings of existing cross-style face generation methods that cannot simultaneously handle style transfer and facial change elimination, and to propose a cross-style face prototype reconstruction technology based on identity preservation.

[0007] A cross-style face prototyping technique based on identity preservation includes the following steps:

[0008] S1: Establish a cross-style face prototype reconstruction model based on GAN. The model includes a generator and two multi-task discriminators. The generator consists of two symmetric encoder-decoder subnetworks to generate face prototype images of two styles, domain A and domain B. The two multi-task discriminators are applied to domain A and domain B respectively to predict the identity label of the image in the domain and determine whether the image is a real prototype.

[0009] S2: Establish the model's loss function to guide the model's training process, ultimately enabling the generator to retain the original identity features of the image while converting the style of face images from different domains.

[0010] S3: Input the training set of face images with identity labels from domain A and domain B into the model for model training, and build a test set to view the face prototype reconstruction effect of the trained model on the test face images.

[0011] Preferably, the specific process of step S1 is as follows:

[0012] S11: Constructing the generator G: G contains two symmetric encoder-decoder subnetworks, namely G A and G B Each sub-network contains an encoder and a decoder; first, G A via encoder Enc A Encode a 256-dimensional identity feature for an input image x from domain A. At the same time G B via encoder Enc B Encode a 256-dimensional identity feature for the input image y from domain B. Subsequently, P x With a 50-dimensional random noise vector Concatenate and input to decoder Dec A Generate the prototype image of x in domain B. P y With another 50-dimensional random noise vector Concatenate and input to decoder Dec B Generate the prototype image of y in domain A. at last, It will be fed back to its symmetric encoder-decoder subnetwork G B Generate image To approximate the true prototype of x in domain A (i.e., x rp ),at the same time It will also be fed back to its symmetric encoder-decoder subnetwork G A Generate image To approximate the true prototype of y in domain B (i.e., y rp );

[0013] S12: Construct discriminant D and D = [D id D gan ] is a multi-task discriminator, where D id It is an identity-related sub-discriminator used to predict the identity of a face image in domain B; D id Output an N d A dimensional vector, where the position of the maximum value corresponds to the identity label; D gan It is an adversarial discriminator used to distinguish between real prototype images in domain B and those in domain G. A The generated virtual prototype image; similar to D, It is also a multi-task discriminator consisting of two sub-discriminators. and Output an N d A dimensional vector is used to predict the identity of an input face image in domain A. Used to distinguish between real and fake prototype images in domain A.

[0014] Preferably, the specific process of step S2 is as follows:

[0015] S21: Construct the prototype adversarial loss function for the model. Design a prototype adversarial loss function, namely L. adv Used to guide the training of G and D gan and L adv The expression is as follows:

[0016]

[0017] In the formula E[·] represents the expected value of the distribution function, and the loss function L adv The first two items indicate G A Attempting to generate a realistic prototype image of x in domain B. And D gan Then try to distinguish And the real domain B prototype image y rp Loss function L adv The last two terms indicate that G B Attempting to generate a realistic prototype image of y in domain A. and Then try to distinguish And the real domain A prototype image x rp ;

[0018] S22: Construct the identity preservation loss function for the model. Design an identity preservation loss function using the identity label information from the training dataset, namely L... id Used to guide the training of G and D id and L id The expression is as follows:

[0019]

[0020] In the formula and D id and The i-th element, the loss function L id The first two items are used to train D. id and This forces them to correctly predict the identity labels of face images in their respective domains; specifically, given an input image y (or x) from domain B (or domain A), Aimed at accurately predicting the identity tags of images (or Loss function L id The last two terms are used to train G to generate identity-preserving face prototype images, so that the domain identity discriminator classifies them as having the same identity label as the input image.

[0021] S23: Construct a prototype learning recurrent loss function, i.e., L cyc Unlike the cycle consistency loss function in Cycle-GAN, which forces the input image to remain consistent with the original image after two forward and backward mappings, L... cyc The aim is to use a prototype learning recurrent network to revert an input image to its prototype in the source domain after two forward and backward mappings. and L cyc The expression is as follows:

[0022]

[0023] In the formula ||·||1 represents the L1 norm;

[0024] S24: Construct the final loss function of the model. By combining the three loss functions mentioned above, the final loss function L of the model and its optimization are shown below:

[0025]

[0026] In the formula, λ and μ are equilibrium hyperparameters, representing L, respectively. id and L cyc The weight.

[0027] Preferably, the specific process of step S3 is as follows:

[0028] S31: Given a training set of labeled face images from domains A and B, where each training image x in domain A follows a P-order inequality. dataA Distribution and use For each training image y in domain B, the annotation is performed, and the image follows the P-order. dataB Distribution and use Add annotations. (or ) represents the identity label of x (or y). (or ) Reflects whether x (or y) contains facial changes; selected from the training set and Face samples were used to construct real-world prototype corpora for domains A and B, respectively. Each image x in the prototype corpus for domain A... rp Obey P realA Distribution, each image y in the prototype corpus of domain B rp Obey P realB Next, the training set, the prototype corpus of domain A, and the prototype corpus of domain B are input into the model and the model is trained according to the loss function L.

[0029] S32: Build a test set and check the trained model's performance on test face images x from domain A. t With test face image y from domain B t The effect of cross-style face prototype reconstruction.

[0030] Preferably, in step S31, if x includes arbitrary facial variations, then otherwise

[0031] The beneficial effects of this invention are as follows:

[0032] (1) This invention overcomes the shortcomings of existing cross-style face generation methods that cannot simultaneously handle style transfer and facial change elimination, and can realize cross-spectral (e.g. from near-infrared to visible light) and cross-art style (e.g. from sketch to photograph) face prototype reconstruction.

[0033] (2) Unlike previous studies on face prototype reconstruction that overemphasized the categories of facial changes and how to accurately model them mathematically, this invention targets an end-to-end generative network for general facial change removal. By combining the identity information of the input face sample with the style information of the target domain, it directly generates a target domain standard face prototype image that preserves the identity, thus avoiding the problem of difficulty in modeling facial changes. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 A schematic diagram of the model generator G;

[0036] Figure 2 This is a schematic diagram of the model discriminator D;

[0037] Figure 3 For model discriminator Schematic diagram;

[0038] Figure 4 A schematic diagram of a recurrent network for prototype learning;

[0039] Figure 5 For discriminator D and The network structure diagram is shown, where C{i}-f{j}s{k}p{h} represents the i-th convolutional layer, with kernel size jxj, stride k, and padding h;

[0040] Figure 6 For the encoder Enc in generator G A and Enc B Network structure diagram;

[0041] Figure 7 For the decoder Dec in generator G A and Dec B The network structure diagram shows that Dec{i}-f{j}s{k}p{h} represents the i-th deconvolutional layer with kernel size jxj, stride k, and padding h.

[0042] Figure 8Demonstrates the prototype reconstruction effect between the near-infrared and visible light domains;

[0043] Figure 9 This demonstrates the prototype reconstruction effect between the visible light domain and the near-infrared domain.

[0044] Figure 10 This demonstrates the prototype reconstruction effect between the sketch domain and the photo domain.

[0045] Figure 11 This demonstrates the prototype reconstruction effect between the photo domain and the sketch domain. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0047] A cross-style face prototyping technique based on identity preservation includes the following steps:

[0048] S1: Establish a cross-style face prototype reconstruction model based on GAN. The model includes a generator and two multi-task discriminators. The generator consists of two symmetric encoder-decoder subnetworks to generate face prototype images of two styles, domain A and domain B. The two multi-task discriminators are applied to domain A and domain B respectively to predict the identity label of the image in the domain and determine whether the image is a real prototype.

[0049] S11: Constructing the generator G: G contains two symmetric encoder-decoder subnetworks, namely G A and G B Each sub-network contains an encoder and a decoder; first, G A via encoder Enc A Encode a 256-dimensional identity feature for an input image x from domain A. At the same time G B via encoder Enc B Encode a 256-dimensional identity feature for the input image y from domain B. Subsequently, P x With a 50-dimensional random noise vector Concatenate and input to decoder Dec A Generate the prototype image of x in domain B. P y With another 50-dimensional random noise vector Concatenate and input to decoder DecB Generate the prototype image of y in domain A. at last, It will be fed back to its symmetric encoder-decoder subnetwork G B Generate image To approximate the true prototype of x in domain A (i.e., x rp ),at the same time It will also be fed back to its symmetric encoder-decoder subnetwork G A Generate image To approximate the true prototype of y in domain B (i.e., y rp );

[0050] S12: Construct discriminant D and D = [D id D gan ] is a multi-task discriminator, where D id It is an identity-related sub-discriminator used to predict the identity of a face image in domain B; D id Output an N d A dimensional vector, where the position of the maximum value corresponds to the identity label; D gan It is an adversarial discriminator used to distinguish between real prototype images in domain B and those in domain G. A The generated virtual prototype image; similar to D, It is also a multi-task discriminator consisting of two sub-discriminators. and Output an N d A dimensional vector is used to predict the identity of an input face image in domain A. Used to distinguish between real and fake prototype images in domain A.

[0051] S2: Establish the model's loss function to guide the model's training process, ultimately enabling the generator to retain the original identity features of the image while converting the style of face images from different domains.

[0052] S21: Constructing the prototype adversarial loss function for the model. To enable the generator G to generate realistic prototype images, a prototype adversarial loss function, namely L, is designed. adv Used to guide the training of G and D gan and L adv The expression is as follows:

[0053]

[0054] In the formula E[·] represents the expected value of the distribution function, and the loss function L adv The first two items indicate G A Attempting to generate a realistic prototype image of x in domain B. And D gan Then try to distinguish And the real domain B prototype image y rp Loss function L adv The last two terms indicate that G B Attempting to generate a realistic prototype image of y in domain A. and Then try to distinguish And the real domain A prototype image x rp ;

[0055] S22: Constructing the identity preservation loss function for the model. To preserve the identity information of the original face images as much as possible during cross-style prototype reconstruction, an identity preservation loss function, L, is designed using the identity label information of the training dataset. id Used to guide the training of G and D id and L id The expression is as follows:

[0056]

[0057] In the formula and D id and The i-th element, the loss function L id The first two items are used to train D. id and This forces them to correctly predict the identity labels of face images in their respective domains; specifically, given an input image y (or x) from domain B (or domain A), Aimed at accurately predicting the identity tags of images (or Loss function L id The last two terms are used to train G to generate identity-preserving face prototype images, so that the domain identity discriminator classifies them as having the same identity label as the input image.

[0058] S23: Construct a prototype model to learn the recurrent loss function, in order to prevent the encoder-decoder subnetwork G in G from being recurrent. A and G B Generate random prototype images unrelated to the input image, and design a prototype learning recurrent loss function, namely L. cyc Unlike the cycle consistency loss function in Cycle-GAN, which forces the input image to remain consistent with the original image after two forward and backward mappings, L... cyc The aim is to use a prototype learning recurrent network to revert an input image to its prototype in the source domain after two forward and backward mappings. and L cycThe expression is as follows:

[0059]

[0060] In the formula ||·||1 represents the L1 norm;

[0061] S24: Construct the final loss function of the model. By combining the three loss functions mentioned above, the final loss function L of the model and its optimization are shown below:

[0062]

[0063] In the formula, λ and μ are equilibrium hyperparameters, representing L, respectively. id and L cyc The weight.

[0064] S3: Input the training set of face images with identity labels from domain A and domain B into the model for model training, and build a test set to view the face prototype reconstruction effect of the trained model on the test face images.

[0065] S31: Given a training set of labeled face images from domains A and B, where each training image x in domain A follows a P-order inequality. dataA Distribution and use For each training image y in domain B, the annotation is performed, and the image follows the P-order. dataB Distribution and use Add annotations. (or ) represents the identity label of x (or y). (or This reflects whether x (or y) contains facial changes. Taking x as an example, if x contains any facial changes, then... otherwise Select from the training set and Face samples were used to construct real-world prototype corpora for domains A and B, respectively. Each image x in the prototype corpus for domain A... rp Obey P realA Distribution, each image y in the prototype corpus of domain B rp Obey P realB Next, the training set, the prototype corpus of domain A, and the prototype corpus of domain B are input into the model and the model is trained according to the loss function L.

[0066] S32: Build a test set and check the trained model's performance on test face images x from domain A. t With test face image y from domain B t The effect of cross-style face prototype reconstruction.

[0067] The specific implementation method is as follows:

[0068] S1: Experimental Dataset Configuration. This invention focuses on cross-style face prototyping using face images of different styles. The datasets we used are the CASIA NIR-VIS v2.0 dataset, which includes near-infrared and visible light images, and the CUGSF dataset, which includes sketches and real photographs. CASIA NIR-VIS v2.0 is currently the largest publicly available dataset spanning the near-infrared and visible light domains. It collected 17,580 images from 725 volunteers, with each volunteer having 5-50 near-infrared face images and 1-22 visible light face images. The images in this dataset contain various facial variations, including pose, expression, lighting, and occlusion (such as wearing glasses). In the experiments, we selected 360 volunteers, containing approximately 6,100 near-infrared images and 2,500 visible light images, as the training set, and another 358 volunteers' near-infrared and visible light images as the test set. The CUGSF dataset is a commonly used public dataset spanning sketch and photograph domains. It contains 1194 photographs from 1194 volunteers in the FERET dataset, along with 1194 corresponding sketches drawn by artists from those photographs. 200 sketches and 200 photographs from 200 volunteers were randomly selected for evaluation. Then, facial images with expressions and poses were borrowed from five subsets of FERET (bd, bf, bg, bj, and bk) to augment the photograph set, resulting in each volunteer having one sketch and six real photographs. In the experiments, 100 volunteers (600 photographs and 100 sketches) were randomly selected as the training set, and images from the remaining 100 volunteers were selected as the test set. During model training and testing, all images from both datasets were first converted to 128×128 pixel grayscale images.

[0069] S2: Model implementation details and parameter settings. Model generator G, such as... Figure 1 As shown.

[0070] Model discriminator D, such as Figure 2 As shown, the model discriminator like Figure 3 As shown. For D and The model uses CASIA-Net as the backbone network and adds an additional fully connected (FC) layer, such as... Figure 5 As shown. Each convolutional layer in the network undergoes batch normalization (BN) and exponential linear unit (ELU) activation. D (or Output an N dA +1 dimensional vector, where the first N d Wei Bei D id (or The first dimension is used to predict the identity label of a face image in domain B (or domain A), while the last dimension is reserved for D. gan (or This is used to score and thus distinguish between real and fake prototype images in domain B (or domain A).

[0071] For the encoder Enc of G A and Enc B The model uses a Lightened-CNN pre-trained on the MS-Celeb-1M dataset as the encoding network, such as... Figure 6 As shown. And for the decoder Dec in G... A and Dec B The model uses inverse CASIA-Net as the decoding network and introduces batch normalization and exponential linear units after each deconvolution layer, such as... Figure 7 As shown. G is transmitted via encoder Enc A Encode a 256-dimensional identity feature P for an input image x from domain A. x Meanwhile, through encoder Enc B Encode a 256-dimensional identity feature P for the input image y from domain B. y Subsequently, P x It is concatenated with a 50-dimensional random noise vector z1 and input into the decoder Dec. A Generate the prototype image of x in domain B. P y It is concatenated with another 50-dimensional random noise vector z2 and input into the decoder Dec. B Generate the prototype image of y in domain A. at last, It will be fed back to its symmetric encoder-decoder subnetwork G B Generate image To approximate the true prototype x in domain A rp ,at the same time It will also be fed back to its symmetric encoder-decoder subnetwork G A Generate image To approximate the true prototype of y in domain B rp The entire prototype learning cycle can be referenced. Figure 4 .

[0072] The balancing hyperparameters λ and μ of the loss function L are set to 2 and 0.1, respectively. Based on the loss function L, we optimize the model using the Mini-Batch Gradient Descent algorithm, with a batch size of 5. All G, D, and The network weights are initialized from a zero-centered normal distribution with a standard deviation of 0.02. We use Adam as the optimizer, with momentum and learning rate set to 0.5 and 0.0002, respectively.

[0073] S3: Model Testing Results. After the model is trained, we test the generator G in the model against random test face images x from domain A. t With random test face images y from domain B t Perform cross-style face prototype reconstruction. Figure 8 This is a prototype reconstruction example between the near-infrared domain (domain A) and the visible light domain (domain B). Figure 9 This is a prototype reconstruction example between the visible light domain (domain B) and the near-infrared domain (domain A). Figure 10 This is an example of reconstructing the prototype between the sketch domain (domain A) and the photograph domain (domain B). Figure 11 This is an example of reconstructing the prototypes between the photographic domain (domain B) and the sketch domain (domain A). Figures 8-11 As can be observed, the model successfully reconstructs cross-style facial prototypes across near-infrared and visible light domains, as well as sketches and photographs, while accurately preserving the identity information of the input images. Specifically, the model can reconstruct identity-preserving facial prototype images in the visible light domain based on contaminated facial images with facial expressions, head postures, or those showing glasses in the near-infrared domain, and can also reconstruct standard frontal photographs based on sketches. Conversely, the model can also perform well in reconstructing prototypes from the visible light domain to the near-infrared domain and from the photographic domain to the sketch domain.

[0074] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for cross-style face prototype reconstruction based on identity preservation, characterized in that, Includes the following steps: S1: Establish a cross-style face prototype reconstruction model based on GAN. The model includes a generator and two multi-task discriminators. The generator consists of two symmetric encoder-decoder subnetworks to generate face prototype images of two styles, domain A and domain B. The two multi-task discriminators are applied to domain A and domain B respectively to predict the identity label of the image in the domain and determine whether the image is a real prototype. S2: Establish the model's loss function to guide the model's training process, ultimately enabling the generator to retain the original identity features of the image while converting the style of face images from different domains. S3: Input the training set of face images with identity labels from domain A and domain B into the model for model training, and build a test set to view the face prototype reconstruction effect of the trained model on the test face images. The specific process of step S1 is as follows: S11: Building Generators : It contains two symmetric encoder-decoder subnetworks, namely and Each sub-network contains an encoder and a decoder; firstly, via encoder For input images from domain A Encode a 256-dimensional identity feature ,at the same time via encoder For input images from domain B Encode a 256-dimensional identity feature ; Subsequently, With a 50-dimensional random noise vector Concatenate and input to decoder Generated in Prototype image in domain B , With another 50-dimensional random noise vector Concatenate and input to decoder Generated in Prototype image in domain A ;at last, It will be fed back to its symmetric encoder-decoder subnetwork. Generate image To approach The real prototype in domain A , at the same time It will also be fed back to its symmetric encoder-decoder subnetwork. Generate image To approach The real prototype in domain B ; S12: Construct the discriminator and : It is a multi-task discriminator, in which It is an identity-related sub-discriminator used to predict the identity of a face image in domain B; Output a A dimensional vector, where the position of the maximum value corresponds to the identity label; It is an adversarial discriminator used to distinguish between real prototype images and real prototype images in domain B. The generated virtual prototype image; similar to , It is also a multi-task discriminator consisting of two sub-discriminators. and ; Output a A dimensional vector is used to predict the identity of an input face image in domain A. Used to distinguish between real and fake prototype images in domain A; The specific process of step S2 is as follows: S21: Construct the prototype adversarial loss function for the model. Design a prototype adversarial loss function, i.e. Used to guide training , and , The expression is as follows: ； In the formula , , The expected value of the distribution function, and the loss function. The first two items indicate Attempting to generate realistic Prototype image in domain B ,and Then try to distinguish And real domain B prototype images loss function The last two items indicate Attempting to generate realistic Prototype image in domain A ,and Then try to distinguish And real domain A prototype image ; S22: Construct the identity preservation loss function for the model. Design an identity preservation loss function using the identity label information from the training dataset, i.e. Used to guide training , and , The expression is as follows: ； In the formula and They represent and The elements, loss function The first two are used for training. and This forces them to correctly predict the identity labels of face images in their respective domains; specifically, given an input image from domain B or domain A. or , or Aimed at accurately predicting the identity tags of images or loss function The last two items are used for training. Generate an identity-preserving prototype image of the face, so that the domain identity discriminator classifies it as having the same identity label as the input image; S23: Construct a prototype learning recurrent loss function. Design a prototype learning recurrent loss function, i.e. Unlike the cycle consistency loss function in Cycle-GAN, which forces the input image to remain consistent with the original image after two forward and backward mappings, The aim is to use a prototype learning recurrent network to revert an input image to its prototype in the source domain after two forward and backward mappings. and , The expression is as follows: ； In the formula , , Represents the L1 norm; S24: Construct the final loss function of the model. By combining the three loss functions mentioned above, the final loss function of the model is... The optimizations are shown below: ； In the formula and To balance the hyperparameters, respectively representing and The weight.

2. The cross-style face prototype reconstruction method based on identity preservation according to claim 1, characterized in that, The specific process of step S3 is as follows: S31: Given a training set of face images with identity labels from domain A and domain B, where each training image from domain A... obey Distribution and use Label each training image in domain B. obey Distribution and use Add annotations. or express or identity tags or reflect or Does it include facial changes; selected from the training set and Face samples were used to construct real-world prototype corpora for domains A and B, respectively. Each image in the prototype corpus for domain A... obey Distribution, each image in the prototype corpus of domain B obey Next, the training set, the prototype corpus of domain A, and the prototype corpus of domain B are input into the model according to the loss function. Perform model training; S32: Build a test set and check the trained model's performance on test face images from domain A. With test face images from domain B The effect of cross-style face prototype reconstruction.

3. The cross-style face prototype reconstruction method based on identity preservation according to claim 2, characterized in that, In step S31, if Including any facial changes, then ,otherwise .

Images