Attribute decomposition-based unconditional bifocal face image generation method and system

Abstract

The invention provides an unconditional dual-light face image generation method and system based on attribute decomposition, and the method comprises the steps: 1, obtaining training data: carrying out the data collection through a thermal infrared camera and a visible light camera, and processing an infrared image through a quantification method, so as to enable the infrared image to be aligned with a visible light image; step 2, a training stage of the generative network: a training sample is a face image pair with matched identities, an identity feature extractor performs pre-training on a data set, and a generator recombines identity features acquired by the identity feature extractor and style features extracted by an encoder to reconstruct an image; step 3, a stage of generating a bifocal face image; any two images with different spectrums are selected as training samples, and pre-training is carried out on features extracted by an identity feature extractor through a random identity sampler. The method has the beneficial effects that the style and identity characteristics of the visible light image and the style characteristics of the thermal infrared image are extracted respectively, and consistent retention of identity information in a cross-modal mode is realized.

Description

Technical Field

[0001] This invention relates to the field of dual-light face image generation technology, and in particular to an unconditional dual-light face image generation method and system based on attribute decomposition. Background Technology

[0002] Cross-modal face recognition, from thermal infrared to visible light, has practical applications and significant potential for improvement in fields such as identity verification, security monitoring, and human-computer interaction, making it a worthy area of ​​research. However, a challenge facing cross-modal face recognition is the lack of a large number of dual-light face images for identity matching available for training. Existing dual-light face datasets are relatively small, containing fewer than 10,000 images. Therefore, synthesizing a large number of thermal infrared and visible light matched face images through generative networks is an important research direction.

[0003] Generative models can generally be divided into conditional generative models and unconditional generative models.

[0004] Conditional two-light generation involves generating a corresponding face image in another modality given a face image in one modality (e.g., infrared or visible light). The goal is to generate a target modality image that matches the input image in terms of identity consistency, structural alignment, and visual realism. Pix2Pix is ​​a representative work on conditional generative models. Pix2Pix consists of a generator and a discriminator. The generator uses an encoder-decoder structure with skip connections to preserve detailed information from the input image, while the discriminator focuses on the local structure of the image to improve texture realism. Using paired infrared and visible light images, Pix2Pix can generate matching two-light face images.

[0005] Unconditional generative models refer to the simultaneous generation of a pair of structurally and identity-identifying visible and infrared face images without relying on any external input conditions (such as infrared or visible light images). The core of this method lies in generating bimodal image pairs from random noise or latent variables in a single operation using a generative model, without requiring original images or pairing labels. This falls under the category of unconditional, cross-modal, dual-output face image generation. The most representative work includes variational autoencoders, whose core idea is to learn the distribution of latent variables in the data and sample from it to generate new data. After training, images can be directly sampled from a standard normal distribution to unconditionally generate modal pairs. The unconditional dual-light face generation model built using variational autoencoders is a cross-modal generation scheme that is stable to train, structurally clear, and has low data requirements, making it suitable for widespread application in the field of face recognition.

[0006] Conditional generative modeling methods heavily rely on the quality of face data obtained through thermal infrared and visible light dual-light matching. The lack of training data limits the model's training effectiveness and generalization ability to some extent. Unconditional generative model enhancement methods can train generative models without paired training data, enabling end-to-end unconditional bimodal face image generation and simplifying practical applications. However, current unconditional generative modeling techniques still have many challenges.

[0007] Existing methods often encode the entire image as a single latent variable, failing to effectively separate style attributes (such as illumination and spectral modality) from identity attributes (such as facial structure). This results in the inability to freely combine modal styles in the generated image, poor identity preservation capabilities, and a tendency for identity information to be interfered with or lost.

[0008] While current unconditional multimodal generation frameworks employ optimization methods and loss functions that can synthesize visually realistic dual-light faces, their latent variable encoding mixes style and identity information, making it difficult to achieve strict feature control and style switching, and lacking structural decoupling processing.

[0009] In the generation stage, existing dual-light image generation technology still relies on training data for the identity information of the generated images. The number of identity features is still limited by the number of faces in the training set, which fails to expand the number of identities in the dataset.

[0010] Unconditional generative models lack high-performance pre-trained identity encoders and rely entirely on end-to-end training, resulting in limited performance in preserving facial identities and large errors in identity feature reconstruction. This may lead to a decrease in identity recognition rate and make it difficult to meet the needs of subsequent recognition work.

[0011] Summary of the Invention To address the problems in the prior art, this invention provides a method and system for generating unconditional dual-light face images based on attribute decomposition, comprising the following steps:

[0012] Step 1, Acquisition of training data: Data is collected using a thermal infrared camera and a visible light camera, and the infrared images are processed by quantization to align them with the visible light images;

[0013] Step 2, the training phase of the generator network: The training samples are face image pairs that match identities. The identity encoder is pre-trained on the dataset. During training, the generator recombines the identity features obtained by the identity encoder and the style features extracted by the encoder to reconstruct the image.

[0014] Step 3, the generation stage of dual-light face images; the training samples are any two images with different spectra, and a random identity encoder is generated. The random identity encoder is pre-trained on the features extracted by the identity encoder. During generation, the generation round is specified. In each round, according to the style of the given input image, combined with the identity features of the random identity encoder, matching image pairs with different identities are generated.

[0015] As a further improvement of the present invention, in step 1, the data collection conditions include: changes in lighting, facial expressions, and angle changes. Changes in lighting include normal lighting, dim lighting, and darkness. Facial expressions include normal faces, different facial expressions, wearing glasses, and wearing masks. Angle changes include changes in the head's position at different angles. As a further improvement of the present invention, step 2 specifically includes:

[0016] Step S1: Use one encoder to extract style features and facial identity features from visible light face images, and use another encoder to extract style features from thermal infrared face images.

[0017] Step S2: Input the style features and facial identity features extracted in step S1 into the generator, and output the three-channel image generated by the generator.

[0018] As a further improvement of the present invention, in step S1, the style encoder projects the image sample onto the corresponding style feature space, and the identity encoder projects the image sample onto the identity space; the mesh encoder includes a visible light style feature encoder and a thermal infrared style feature encoder, the visible light style feature encoder learns the image style of the visible light image, and the thermal infrared style feature encoder learns the image style of the thermal infrared image. As a further improvement of the present invention, the specific implementation of the style encoder is as follows:

[0019] C32-C64-C128-C256-C256-C256-F,

[0020] Among them, C K The generator represents K 3×3 Conv-BN-LeakyReLU layers, and F represents a fully connected layer; the specific implementation of the generator is as follows:

[0021] D128-Res-D128-Res-D64-Res-D64-Res-D32-Res-D32-Res-C32,

[0022] Among them, D K A ConvT-BN-LeakyReLU layer representing K convolutional kernels. As a further improvement of this invention, the loss function used in the training of this unconditional dual-light face image generation method includes the image loss function L. L1 Orthogonal loss function L ortIdentity Preservation Loss Function L id and the generative adversarial loss function L adv The specific formula is as follows:

[0023] L = L L1 +αL ort +βL id +γL adv (1)

[0024] Wherein, the image loss function L L1 This is used to measure the pixel difference between the generated image pair and the real image. When training with paired dual-light face images, through attribute decoupling and image reconstruction of the network, the output image pair is close to the original image pair in pixels. The formula is as follows:

[0025]

[0026] in, and These represent the generated thermal infrared image and visible light image, respectively. T and I V These represent the original thermal infrared image and the original visible light image, respectively.

[0027] Orthogonal loss function L ort It decouples the identity attributes and image style attributes contained in a face image by making the visible light style features, infrared style features and identity features mutually orthogonal;

[0028] Identity Preservation Loss Function L id It is used to constrain the generator to fully preserve the identity information attributes in the face when reconstructing the image;

[0029] Generative adversarial loss function L adv Used to increase the stability of the training process;

[0030] The unconditional dual-light face image generation method fixes the encoder during generator training and fixes the generator during encoder training.

[0031] As a further improvement of the present invention, in step 3, the random identity encoder is implemented based on a variational autoencoder network. For any input sample X, output, and corresponding latent vector z, the variational autoencoder learns the mapping from the input sample X to the distribution P(z|X), where the distribution P(z|X) consists of the mean μ and variance σ. 2 This means that the generator learns the mapping from random sampling of the distribution P(z|X) to the output. As a further improvement of the invention, step 3 further includes:

[0032] Step a1, Random Identity Encoder Training: The training samples of the random identity encoder come from the identity features extracted from the visible light image in the identity encoder. After training, the decoding module in the variational autoencoder is retained to obtain a random identity encoder. Given a noise or randomly sampled from the normal distribution, the noise is input into the random identity encoder to obtain a random identity feature vector that satisfies the distribution law of face identity features.

[0033] Step a2, dual-light face generation: After the unconditional dual-light face generation network is constructed, any two thermal infrared and visible light images are input to provide style features. Random identity features are then obtained from the noise. Finally, the generator recombines these two pairs of features to generate a pair of face images with the same identity.

[0034] As a further improvement of the present invention, the unconditional dual-light face image generation method employs a high-efficiency lightweight identity encoder, and the specific operation of the MFM in the high-efficiency lightweight feature encoder is as follows:

[0035] MFM(x) i =max(x i x i+N ), i = 1, 2, ..., N (4)

[0036] Where x is the input feature map with dimensions 2N×H×W, and the corresponding output dimension is N×H×W. N represents the number of input channels, H represents the height, and W represents the width.

[0037] The present invention also discloses an unconditional dual-light face image generation system based on attribute decomposition, comprising: a memory, a processor, and a computer program stored in the memory, wherein the computer program is configured to implement the steps of the method described in the present invention when invoked by the processor.

[0038] The present invention also discloses an unconditional dual-light face image generation system based on attribute decomposition, comprising: a memory, a processor, and a computer program stored in the memory, wherein the computer program is configured to implement the steps of the method described in the present invention when invoked by the processor.

[0039] The beneficial effects of this invention are: it enables the generation of a large number of randomly identified face images using a dual-light face image generation network based on face attribute decomposition. Specifically:

[0040] (1) This invention deconstructs the representation of face images into two key features: image style features and identity features. By designing a multi-channel encoder structure, the style and identity features of visible light images and the style features of thermal infrared images are extracted respectively, so as to achieve consistent preservation of identity information across modalities.

[0041] (2) This invention designs a loss function based on feature decoupling. In order to effectively separate style and identity information, a style and identity decoupling loss function is proposed to guide the encoder to learn mutually independent feature representations, while enhancing the stability of the training process and the identity control capability of the generated image.

[0042] (3) The present invention designs a random identity encoder to replace the real identity feature extractor, generates a controllable identity vector from noise, and realizes cross-modal identity preservation generation. This mechanism can automatically generate infrared-visible image pairs with consistent identities but different image styles, effectively expand the identity sample space and improve the model generalization ability.

[0043] (4) This invention designs a high-performance identity encoder and adopts a pre-training strategy. A high-performance identity encoder with optimized structure is introduced, which adopts the maximum feature mapping activation function and improves its initial recognition ability through a pre-training strategy, thereby enhancing the identity preservation ability of the entire generative model and the accuracy of the generated images. Attached Figure Description

[0044] Figure 1 This is a diagram of the unconditional dual-light face image generation network structure based on face attribute decomposition of the present invention.

[0045] Figure 2 This is the random identity encoder structure of the present invention. Detailed Implementation

[0046] To address the problems in the prior art, this invention provides an unconditional dual-light face image generation method based on attribute decomposition. It designs a network that synthesizes a large number of random dual-light identity matching face image pairs from noise to assist in the training of cross-modal face recognition models.

[0047] This invention discloses an unconditional dual-light face image generation method based on attribute decomposition, with potential applications including access control, security monitoring, identity verification, person classification, and identity screening. The method involves generating a large number of virtual dual-light face image pairs with diverse identities before training face recognition models using dual-light face images, thereby improving the accuracy of face recognition.

[0048] This invention discloses an unconditional two-light face image generation method based on attribute decomposition, comprising the following steps:

[0049] Step 1, Acquisition of training data: Data is collected using a high-resolution thermal infrared camera and a visible light camera, and the infrared images are converted to 8 bits and aligned with the visible light images using a quantization method; the lighting changes in this dataset are set with three different lighting conditions: normal lighting, shadow, and darkness; the facial changes are set with three conditions: normal face, different expressions, wearing glasses, and wearing a mask; and the angle changes include the changes in the head at different angles.

[0050] Step 2, the training phase of the generator network: The training samples are face image pairs that match identities. The identity encoder is pre-trained on the dataset. During training, the generator recombines the identity features obtained by the identity encoder and the style features extracted by the encoder to reconstruct the image.

[0051] Step 3, the generation stage of dual-light face images; the training samples are any two images with different spectra, and the random identity encoder is used to pre-train on the features extracted by the identity encoder. During generation, the generation round is specified. In each round, according to the style of the given input image, the identity features of the random identity encoder are combined to generate matching image pairs with different identities.

[0052] 1. A face image generation method with style and identity decoupling

[0053] This invention proposes a face image generation method based on the decoupling of style features and identity features, which is particularly suitable for generating bimodal (infrared and visible light) face images. This method divides the face image generation process into two complementary parts: image style feature extraction and identity feature extraction, thereby achieving controllable image synthesis and style transfer.

[0054] The specific network structure is as follows: Figure 1 As shown, the network uses one encoder to extract style features and facial identity features from visible light images, and another encoder to extract style features from thermal infrared images. The style encoder projects image samples onto the corresponding style feature space. The visible light style feature encoder and the thermal infrared style feature encoder learn the image styles of the visible light image and the thermal infrared image, respectively, while the identity encoder projects image samples onto the identity space. The generator takes identity features and image style features as input and outputs a generated three-channel image, primarily mapping the joint distribution of paired identity features and style features back to the image, thus completing the decoupling of facial attributes and image reconstruction. More specifically, the style encoder is implemented as follows:

[0055] C32-C64-C 128-C256-C256-C256-F,

[0056] Where C KF represents K 3×3 Conv-BN-LeakyReLU layers (standard network layers with convolution, normalization, and activation functions), and F represents a fully connected layer.

[0057] The generator is implemented as follows:

[0058] D128-Res-D128-Res-D64-Res-D64-Res-D32-Res-D32-Res-

[0059] C32,

[0060] Where D K A ConvT-BN-LeakyReLU layer (deconvolutional normalized activation layer) representing K convolutional kernels.

[0061] 2. Loss function based on feature decoupling

[0062] The loss function used in the training of the model proposed in this invention consists of four parts, namely the image loss function L. L1 Orthogonal loss function L ort Identity Preservation Loss Function L id and the generative adversarial loss function L adv , Figure 1 The diagram shows the location of some loss functions within the network; the specific formulas are as follows:

[0063] L = L L1 +αL ort +βL id +γL adv (1)

[0064] First, the image loss function L L1 This is used to measure the pixel difference between the generated image pair and the real image. When training with paired two-light face images, through attribute decoupling and image reconstruction of the network, the output image pair should be as close as possible to the original image pair in pixels. The formula is expressed as follows:

[0065]

[0066] in and These represent the generated thermal infrared image and visible light image, respectively. T and I V These represent the original thermal infrared image and the original visible light image, respectively.

[0067] Orthogonal loss function L ortThe main objective is to decouple the identity and image style attributes inherent in a face image by making visible light style features, infrared style features, and identity features orthogonal to each other. Specifically, this algorithm uses the normalized inner product of the two features to represent similarity, and forces them to gradually become orthogonal by minimizing the absolute value of the inner product.

[0068] Identity Preservation Loss Function L id The main purpose is to constrain the generator to fully preserve the identity information attributes in the face when reconstructing the image.

[0069] Generative adversarial loss function L adv It is mainly used to increase the stability of the training process. The unconditional generative network based on attribute decomposition proposed in this invention fixes the encoder when training the generator, and fixes the generator when training the encoder.

[0070] 3. Random Identity Encoder

[0071] This invention employs a random identity encoder to generate random identity features from noise. In the network generation stage, the random identity encoder replaces the identity encoder, thereby replacing the original identity features of the dual-light image with random identity features and generating a pair of random thermal infrared and visible light image pairs with the same identity.

[0072] The random identity encoder is implemented based on a variational autoencoder network. Figure 2 The principle is illustrated. For any input X, output, and corresponding latent vector z, the encoder learns the mapping from sample X to distribution P(z|X), which consists of mean μ and variance σ. 2 This means that the generator learns a mapping from random sampling of the distribution P(z|X) to the output. First, the distribution P(z|X) is constrained to approximate a normal distribution, with its mean and variance approximating 0 and 1 respectively. Second, for the reconstruction of the output as input, it should approximate the input itself. Finally, to ensure the differentiability of the neural network, random sampling of the distribution P(z|X) is achieved by adding noise ∈ from the normal distribution. The specific formula is as follows:

[0073] Z=μ+σ⊙ ∈ (3)

[0074] Where ⊙ represents the Hadamard product, which is the digit-by-digit multiplication of vectors.

[0075] The training samples for the random identity encoder are derived from identity features extracted from visible light images in the identity encoder. After training, retaining the decoding module in the variational autoencoder yields a random identity encoder. Given noise or random samples from a normal distribution, inputting the noise into the random identity encoder produces a random identity feature vector that follows the distribution pattern of facial identity features. After the unconditional dual-light face generation network is constructed, the input can be any two thermal infrared and visible light images providing style features. Random identity features are then extracted from the noise, and finally, the generator recombines these two pairs of features to generate a pair of face images with the same identity.

[0076] 4. Highly efficient and lightweight feature extractor

[0077] To improve image identity feature extraction, the identity encoder in this invention employs a pre-trained, efficient, and lightweight identity feature encoder network as its basic structure to enhance the accuracy and robustness of identity feature extraction. The core of this efficient and lightweight feature extractor lies in introducing the Maximum Feature Mapping (MFM) activation function to replace the traditional ReLU activation function, thereby achieving channel-level feature selection and redundancy suppression. The core idea of ​​MFM is to retain the feature response with the largest value among multiple channels at each location, thus achieving a dual function of non-linear activation and channel selection. The specific operation of MFM is as follows:

[0078] MFM(x) i =max(x i x i+N ), i = 1, 2, ..., N (4)

[0079] x is the input feature map with dimensions 2N×H×W, and the corresponding output dimension is N×H×W. N represents the number of input channels, H represents the height, and W represents the width.

[0080] According to the formula, only the maximum value of the two sets of features is retained at each position, compressing the channel dimension while preserving highly discriminative features. Furthermore, this invention uses a feature extractor pre-trained on a visible light face database as the backbone network for encoding identity information. This structure effectively enhances the representation of individual identity attributes, maintaining good discriminative performance even in the presence of image noise, occlusion, or inaccurate labeling.

[0081] The present invention also discloses an unconditional dual-light face image generation system based on attribute decomposition, comprising: a memory, a processor, and a computer program stored in the memory, wherein the computer program is configured to implement the steps of the method described in the present invention when invoked by the processor.

[0082] The specific innovations of this invention are as follows:

[0083] (1) The available features of a face image are reconstructed into two parts: image style features and face identity features. The thermal infrared domain, visible light domain, and lighting conditions can be regarded as the style features of the image, while each person's unique facial attributes represent identity features. Based on the idea of ​​attribute decomposition, this invention uses different encoders to extract the style features of the visible light image, the face identity features of the visible light image, and the style features of the thermal infrared image, which can make the generated infrared image and the visible light image have higher consistency in terms of identity preservation.

[0084] (2) During training, the present invention uses an additional loss function and designs a corresponding loss function to decouple the style feature and identity feature vectors. Finally, the encoder is used to reconstruct a new dual-light face image from the style feature and identity feature, so that the style and identity encoding can be decoupled from each other and the stability of the training process is enhanced.

[0085] (3) The present invention uses a random identity sampler to generate random identity features from noise. In the generation stage of the network, the random identity sampler is used instead of the identity feature extractor, thereby replacing the original identity features of the dual-light image with random identity features and generating a pair of random thermal infrared and visible light image pairs with the same identity, increasing the number of identities in the dataset.

[0086] (4) A higher performance identity encoder was designed, and the network was improved by pre-training to achieve better initial performance, thereby improving the generation quality of the dual-light face image generation network and increasing the accuracy of the results.

[0087] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for generating unconditional two-light face images based on attribute decomposition, characterized in that, Includes the following steps: Step 1, Acquisition of training data: Data is collected using a thermal infrared camera and a visible light camera, and the infrared images are processed by quantization to align them with the visible light images; Step 2, the training phase of the generator network: The training samples are face image pairs that match identities. The identity encoder is pre-trained on the dataset. During training, the generator recombines the identity features obtained by the identity encoder and the style features extracted by the encoder to reconstruct the image. Step 3, the generation stage of dual-light face images; the training samples are any two images with different spectra, and a random identity encoder is generated. The random identity encoder is pre-trained on the features extracted by the identity encoder. During generation, the generation round is specified. In each round, according to the style of the given input image, combined with the identity features of the random identity encoder, matching image pairs with different identities are generated.

2. The unconditional dual-light face image generation method according to claim 1, characterized in that, In step 1, the data collection conditions include: lighting changes, facial changes, and angle changes. Lighting changes include normal lighting, dim lighting, and darkness. Facial changes include normal face, different expressions, wearing glasses, and wearing a mask. Angle changes include changes in the head position at different angles.

3. The unconditional dual-light face image generation method according to claim 1, characterized in that, Step 2 specifically includes: Step S1: Use one encoder to extract style features and facial identity features from visible light face images, and use another encoder to extract style features from thermal infrared face images. Step S2: Input the style features and facial identity features extracted in step S1 into the generator, and output the three-channel image generated by the generator.

4. The unconditional dual-light face image generation method according to claim 3, characterized in that, In step S1, the style encoder projects the image sample into the corresponding style feature space, and the identity encoder projects the image sample into the identity space; the mesh encoder includes a visible light style feature encoder and a thermal infrared style feature encoder. The visible light style feature encoder learns the image style of the visible light image, and the thermal infrared style feature encoder learns the image style of the thermal infrared image.

5. The unconditional dual-light face image generation method according to claim 4, characterized in that, The specific implementation of the style encoder is as follows: C32-C64-C128-C256-C256-C256-F, Among them, C K The generator represents K 3×3 Conv-BN-LeakyReLU layers, and F represents a fully connected layer; the specific implementation of the generator is as follows: D128-Res-D128-Res-D64-Res-D64-Res-D32-Res-D32-Res-C32, Among them, D K A ConvT-BN-LeakyReLU layer representing K convolutional kernels.

6. The unconditional dual-light face image generation method according to claim 1, characterized in that, The unconditional two-light face image generation method uses the image loss function L during training. L1 Orthogonal loss function L ort Identity Preservation Loss Function L id and the generative adversarial loss function L adv The specific formula is as follows: L=L L1 +αL ort +βL id +γL adv (1) Wherein, the image loss function L L1 This is used to measure the pixel difference between the generated image pair and the real image. When training with paired dual-light face images, through attribute decoupling and image reconstruction of the network, the output image pair is close to the original image pair in pixels. The formula is as follows: in, and These represent the generated thermal infrared image and visible light image, respectively. T and I V These represent the original thermal infrared image and the original visible light image, respectively. Orthogonal loss function L ort It decouples the identity attributes and image style attributes contained in a face image by making the visible light style features, infrared style features and identity features mutually orthogonal; Identity Preservation Loss Function L id It is used to constrain the generator to fully preserve the identity information attributes in the face when reconstructing the image; Generative adversarial loss function L adv Used to increase the stability of the training process; The unconditional dual-light face image generation method fixes the encoder during generator training and fixes the generator during encoder training.

7. The unconditional dual-light face image generation method according to claim 1, characterized in that, In step 3, the random identity encoder is implemented based on a variational autoencoder network. For any input sample X, output, and corresponding latent vector z, the variational autoencoder learns the mapping from the input sample X to the distribution P(z|X), which consists of the mean μ and variance σ. 2 This means that the generator learns the mapping from random sampling of the distribution P(z|X) to the output.

8. The unconditional dual-light face image generation method according to claim 7, characterized in that, Step 3 also includes: Step a1, Random Identity Encoder Training: The training samples of the random identity encoder come from the identity features extracted from the visible light image in the identity encoder. After training, the decoding module in the variational autoencoder is retained to obtain a random identity encoder. Given a noise or randomly sampled from the normal distribution, the noise is input into the random identity encoder to obtain a random identity feature vector that satisfies the distribution law of face identity features. Step a2, dual-light face generation: After the unconditional dual-light face generation network is constructed, any two thermal infrared and visible light images are input to provide style features. Random identity features are then obtained from the noise. Finally, the generator recombines these two pairs of features to generate a pair of face images with the same identity.

9. The unconditional dual-light face image generation method according to claim 1, characterized in that, The unconditional dual-light face image generation method employs a high-efficiency, lightweight identity encoder. The specific operation of the MFM in this high-efficiency, lightweight feature encoder is as follows: MFM(x) i =max(x i ,x i+N ),i=1,2,...,N (4) Where x is the input feature map with dimensions 2N×H×W, and the corresponding output dimension is N×H×W. N represents the number of input channels, H represents the height, and W represents the width.

10. An unconditional two-light face image generation system based on attribute decomposition, characterized in that, include: A memory, a processor, and a computer program stored on the memory, the computer program being configured to implement the steps of the method of any one of claims 1-9 when invoked by the processor.

Images