Anti-imaging interference cross-domain biological texture feature recognition method

Abstract

The present application relates to the technical field of biometric recognition, and particularly relates to a cross-domain biological texture feature recognition method resisting imaging interference, comprising the following steps: constructing an image dataset; constructing an identity and style feature decoupling network ISFDNet; constructing an image-level low-frequency disturbance module LFDM; constructing a collaborative supervision loss function, the collaborative supervision loss function comprising an identity supervision loss, a decoupling supervision loss and a style invariance loss; training the constructed ISFDNet using the training set; and using the trained ISFDNet to perform biological texture feature recognition. The present application proposes a general decoupling framework for biological texture features without target domain data, which explicitly separates identity and style features, and the model can extract pure identity features that are not affected by collection equipment, illumination and environmental noise, effectively overcoming the performance decline problem caused by cross-domain, and maintaining high recognition accuracy in complex scenes such as cross-device and cross-spectrum.

Description

Technical Field

[0001] This invention relates to the field of biometric recognition technology, and in particular to a cross-domain biometric texture feature recognition method that is resistant to imaging interference. Background Technology

[0002] Biometric texture recognition is a key technology for identity authentication that utilizes the inherent physiological texture structures of the human body (such as fingerprints, irises, palm prints, and faces). Compared to traditional password authentication, biometric texture features have inherent advantages such as high uniqueness, difficulty in forgery, and portability, and have been widely used in fields such as financial payments, border security checks, smart terminal unlocking, and access control. With the rapid development of deep learning technology, feature extraction methods based on convolutional neural networks have gradually replaced traditional manual coding methods, significantly improving recognition accuracy in single, controlled environments.

[0003] However, in real-world open application scenarios, such as mobile financial payments, smart community access control, or airport security checks, biometric texture recognition systems face severe technical challenges due to the heterogeneity of imaging equipment and the variability of the physical environment. Acquisition devices from different manufacturers (such as smartphone cameras, dedicated texture feature collectors, and security monitoring probes) employ different models of photosensitive chips and optical lens modules. Due to the differences in the photoelectric conversion characteristics of photosensitive elements and the optical curvature of lenses, the biometric texture images of the same user acquired exhibit significant physical differences in resolution, focal length, distortion, and signal-to-noise ratio. For example, sensor noise (Gaussian noise, shot noise) in low-cost cameras can mask the subtle textures of palm prints, while wide-angle lenses can introduce nonlinear geometric deformations.

[0004] Furthermore, changes in the physical conditions of the imaging environment are also a key factor leading to a decline in recognition performance. In the acquisition environment, variations in lighting intensity, lighting angle, background interference, and different acquisition postures result in completely different texture clarity, lighting uniformity, and color tone in the biometric texture feature images of the same user. These issues cause models that perform well under specific conditions to experience a decline in recognition performance when faced with novel sensors or unknown acquisition scenarios, severely impacting the system's robustness and generalization ability.

[0005] Existing technologies primarily employ two solutions: domain adaptation and data augmentation. Domain adaptation methods typically require pre-acquiring large amounts of data from the target scene (such as a specific camera model) for fine-tuning, which is difficult to implement in scenarios involving privacy protection or the access of unknown new devices. Existing methods often employ strategies such as data augmentation (e.g., generating stylized images) or feature alignment. However, these methods mostly operate only at the data level or merely constrain the overall distribution of features, failing to address the information obfuscation problem at the underlying mechanism of feature generation.

[0006] From the perspective of image formation, any biological texture image is actually a coupling of "stable identity information" and "variable style information." Identity information refers to the inherent physiological characteristics of an organism, such as its texture structure, direction, and bifurcation points. This is the fundamental basis for identity identification and should be kept constant. Style information, on the other hand, refers to the visual representation introduced by external factors such as imaging device characteristics, lighting tone, blur level, and noise.

[0007] Existing deep learning models typically extract "hybrid features" where identity and style are highly entangled. Due to the lack of an explicit decoupling mechanism, these models often incorrectly encode style changes as identity differences. When the application scenario changes (i.e., style changes), this entanglement leads to increased intra-class distance in the feature space, resulting in recognition errors. To fundamentally solve this feature entanglement problem, it is necessary to delve into the mechanism of feature generation. If image features can be explicitly decomposed, stripping away style interference that changes with the environment and retaining only the essential features related to identity, true cross-domain robust recognition can be achieved. In the field of biometric texture feature recognition, there is currently a lack of a universal technical solution that can achieve complete separation of identity and style at both the feature and image levels without requiring target domain data. Therefore, developing a cross-domain recognition method based on a decoupling mechanism is of great significance for improving the practicality of biometric recognition systems in multi-source heterogeneous environments. Summary of the Invention

[0008] The purpose of this invention is to provide a cross-domain biological texture feature recognition method that is resistant to imaging interference. It addresses the technical problem of unstable extraction of biological texture image features caused by heterogeneous hardware of acquisition devices (such as sensor noise) and variable imaging environment (such as light intensity and angle). By simulating low-frequency interference in optical imaging and using a feature decoupling mechanism, the inherent physiological structural features of the organism (i.e., identity features) that are not affected by imaging conditions are separated from the mixed image signal containing image style information, thereby achieving high-precision cross-domain identity authentication.

[0009] This invention is implemented as follows: This invention provides a cross-domain biological texture feature recognition method resistant to imaging interference, comprising the following steps:

[0010] S101: Construct a biological texture image dataset, and divide the image dataset into a training set and a test set in a 1:1 ratio; wherein, the biological texture includes fingerprint, iris, palm print, and face texture;

[0011] S102: Construct the identity and style feature decoupling network ISFDNet. The network ISFDNet adopts an encoder-decoder architecture. The encoder is used to decouple the input image and output the corresponding identity feature vector and style feature vector. The decoder is used to reconstruct the identity feature vector and style feature vector output by the encoder into a hybrid feature.

[0012] S103: Construct an image-level low-frequency perturbation module LFDM, which is used to perturb the low-frequency components of the original images in the training set to generate stylized images;

[0013] S104: Construct a collaborative supervision loss function, which includes identity supervision loss, decoupling supervision loss, and style invariance loss;

[0014] S105: The constructed ISFDNet is trained using the training set and the image-level low-frequency perturbation module LFDM. The original images in the training set and the stylized images generated by the image-level low-frequency perturbation module LFDM are paired and input into ISFDNet. The network parameters are optimized through the collaborative supervision loss function to obtain the trained ISFDNet.

[0015] S106: The image to be recognized is input into the trained ISFDNet. The decoupled identity feature vector is extracted, normalized, and then matched with the database template features to complete the identity authentication.

[0016] More preferably, in step S102, the encoder includes a backbone network and a feature decoupling module FDM. The backbone network is used to extract mixed features from the input image, and the feature decoupling module FDM is used to decouple the mixed features into identity feature vectors and style feature vectors. The backbone network can adopt any existing biometric recognition network structure. The feature decoupling module FDM includes a spatial awareness separation module SASM, a channel attention module, and a fully connected layer.

[0017] The decoder includes a spatial awareness separation module (SASM), a channel attention module, a convolutional layer, a batch normalization layer, and a ReLU activation function layer.

[0018] More preferably, in step S102, the spatial awareness separation module SASM in the encoder incorporates convolutional layers and a softmax function. It receives the mixed features output from the backbone network, generates a dual-channel attention logic graph through the convolutional layers, and uses the softmax function to generate complementary identity masks in the channel dimension. and style mask Identity mask and style mask in spatial location The sum of the weights on each feature is 1; the identity mask and style mask are multiplied element-wise with the hybrid features to obtain the initially separated identity features and style features.

[0019] More preferably, in step S103, the specific processing procedure of the image-level low-frequency perturbation module LFDM is as follows: the input image is transformed to the frequency domain using a two-dimensional discrete Fourier transform; the frequency domain is decomposed into low-frequency components and high-frequency components using a low-pass filter mask; the high-frequency components are kept unchanged, the phase spectrum of the low-frequency components is kept unchanged, and the amplitude spectrum of the low-frequency components is processed by a composite perturbation function to obtain the perturbed low-frequency components; the perturbed low-frequency components are merged with the original high-frequency components, and reconstructed into a stylized image through inverse Fourier transform.

[0020] More preferably, the composite perturbation function includes a combination of three strategies: global amplitude scaling, adaptive noise injection, and smooth illumination field simulation. Global amplitude scaling multiplies the low-frequency amplitude by a random sampling scaling factor to simulate brightness and contrast variations. Adaptive noise injection calculates the noise standard deviation based on the mean of the low-frequency amplitude, generates a Gaussian noise field, and superimposes it onto the amplitude spectrum to simulate sensor noise. Smooth illumination field simulation generates a smooth Gaussian random field as an illumination interference map, which is applied to the amplitude spectrum in a multiplicative manner to simulate non-uniform illumination variations.

[0021] More preferably, in step S104, the identity supervision loss includes cross-entropy loss and triplet loss, the decoupling supervision loss includes reconstruction loss, loop loss, orthogonal loss and identity adversarial loss, and the style invariance loss includes identity consistency loss.

[0022] More preferably, in step S104, the collaborative supervision loss function The calculation formula is: ;

[0023] in, , , , , , Losses related to identity supervision Reconstruction losses Cyclic loss Orthogonal loss Identity confrontation loss and identity consistency loss The weighting coefficients.

[0024] More preferably, in step S105, during the training process, the original image and the stylized image are input into ISFDNet in pairs, and the original image is processed by the backbone network to obtain the hybrid features of the original image. Hybrid features of stylized images , and The image is then fed into the Feature Decoupling Module (FDM), which outputs the identity feature vector of the original image. and style feature vector and the identity feature vector of stylized images and style feature vector ;Will and Combined, input into the decoder, to obtain reconstructed hybrid features , and Combined, input to decoder, to obtain stylized reconstructed hybrid features ;Will The input feature decoupling module (FDM) outputs a cyclic identity feature vector. ; and Used for identity supervision loss and style invariance loss. Used to reconstruct losses, and Used for cyclic loss, and Used for orthogonal loss and Used to combat identity-related losses.

[0025] Compared with the prior art, the present invention has the following advantages:

[0026] (1) This invention proposes a general decoupling framework for biological texture features that does not require target domain data. By explicitly separating identity and style features, the model can extract pure identity features that are not affected by the acquisition device, lighting and environmental noise, effectively overcoming the performance degradation problem caused by cross-domain, and maintaining high recognition accuracy in complex scenarios such as cross-device and cross-spectrum.

[0027] (2) This invention designs a Spatial Aware Separation Module (SASM) at the feature layer to adaptively generate complementary masks for spatial feature separation; at the same time, it designs a Low Frequency Disturbance Module (LFDM) at the image layer to simulate real style changes using Fourier frequency domain characteristics. This synergistic effect between the feature layer and the image layer fundamentally solves the problem of highly entangled identity information and style information, ensuring thorough decoupling;

[0028] (3) The decoupling framework proposed in this invention has good versatility. It is not only applicable to the recognition of various biological texture features such as palm prints, fingerprints, and veins, but can also be seamlessly integrated into the existing mainstream recognition backbone network as a general plug-in. Without changing the original network structure, it significantly improves the cross-domain recognition performance of the existing model. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a flowchart illustrating the method for decoupling identity and style features in an embodiment of the present invention.

[0031] Figure 2 This is a flowchart of image-level low-frequency perturbation in an embodiment of the present invention;

[0032] Figure 3 This is a collaborative strategy graph of the loss function in an embodiment of the present invention;

[0033] Figure 4 This is a graph showing the ROC curve of the PolyU dataset in an in-domain experiment according to an embodiment of the present invention;

[0034] Figure 5 This is a graph showing the ROC curve of the Tongji dataset in the domain experiment of this invention embodiment;

[0035] Figure 6 This is a graph showing the ROC curve of the IITD dataset in the domain experiment of this invention embodiment;

[0036] Figure 7 This is a graph showing the ROC curve of the Red dataset in an in-domain experiment according to an embodiment of the present invention.

[0037] Figure 8 This is a graph showing the ROC curve of the Green dataset in the domain experiment of this invention embodiment;

[0038] Figure 9 This is a graph showing the ROC curve of the Blue dataset in the domain experiment of this invention embodiment;

[0039] Figure 10 This is a ROC curve of the NIR dataset in the domain experiment of this invention embodiment;

[0040] Figure 11 This is a visualization analysis diagram of the feature mask generated by the spatial perception separation module in this embodiment of the invention;

[0041] Figure 12 This is a graph showing the change in mask quality index during the training process in an embodiment of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art. The terms "comprising" and similar expressions used herein mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, but does not exclude other elements or objects.

[0043] Example 1: This example provides a cross-domain biological texture feature recognition method that is resistant to imaging interference, such as... Figure 1 As shown, it includes the following steps:

[0044] S101: Construct a biological texture image dataset, and divide the image dataset into a training set and a test set in a 1:1 ratio; wherein, the biological texture includes fingerprint, iris, palm print, and face texture;

[0045] S102: Construct the identity and style feature decoupling network ISFDNet. The network ISFDNet adopts an encoder-decoder architecture. The encoder is used to decouple the input image and output the corresponding identity feature vector and style feature vector. The decoder is used to reconstruct the identity feature vector and style feature vector output by the encoder into a hybrid feature.

[0046] The encoder includes a backbone network and a feature decoupling module (FDM). The backbone network is used to extract mixed features from the input image, and the feature decoupling module (FDM) is used to decouple the mixed features into identity feature vectors and style feature vectors. The backbone network can adopt any existing biometric recognition network structure. The feature decoupling module (FDM) includes a spatial awareness separation module (SASM), a channel attention module, and a fully connected layer.

[0047] The decoder includes a spatial awareness separation module (SASM), a channel attention module, a convolutional layer, a batch normalization layer, and a ReLU activation function layer.

[0048] The spatially aware separation module (SASM) in the encoder incorporates convolutional layers and a softmax function. It receives mixed features from the backbone network output, generates a dual-channel attention logic graph through the convolutional layers, and uses the softmax function to generate complementary identity masks in the channel dimension. and style mask Identity mask and style mask in spatial location The sum of the weights on each feature is 1; the identity mask and style mask are multiplied element-wise with the hybrid features to obtain the initially separated identity features and style features.

[0049] S103: Construct an image-level low-frequency perturbation module LFDM, which is used to perturb the low-frequency components of the original images in the training set to generate stylized images;

[0050] The specific processing steps of the image-level low-frequency perturbation module LFDM are as follows: the input image is transformed to the frequency domain using a two-dimensional discrete Fourier transform; the frequency domain is decomposed into low-frequency and high-frequency components using a low-pass filter mask; the high-frequency components are kept unchanged, the phase spectrum of the low-frequency components is kept unchanged, and the amplitude spectrum of the low-frequency components is processed by a composite perturbation function to obtain the perturbed low-frequency components; the perturbed low-frequency components are merged with the original high-frequency components, and the stylized image is reconstructed by inverse Fourier transform.

[0051] The composite perturbation function includes a combination of three strategies: global amplitude scaling, adaptive noise injection, and smoothed illumination field simulation. Specifically, global amplitude scaling multiplies the low-frequency amplitude by a random sampling scaling factor to simulate changes in brightness and contrast; adaptive noise injection calculates the noise standard deviation based on the mean of the low-frequency amplitude, generates a Gaussian noise field, and superimposes it onto the amplitude spectrum to simulate sensor noise; and smoothed illumination field simulation generates a smoothed Gaussian random field as an illumination interference map, which is applied to the amplitude spectrum in a multiplicative manner to simulate non-uniform illumination changes.

[0052] S104: Construct a collaborative supervision loss function, which includes identity supervision loss, decoupling supervision loss, and style invariance loss; identity supervision loss includes cross-entropy loss and triplet loss, decoupling supervision loss includes reconstruction loss, loop loss, orthogonal loss, and identity adversarial loss, and style invariance loss includes identity consistency loss.

[0053] The collaborative supervision loss function The calculation formula is: ;

[0054] in, , , , , , Losses related to identity supervision Reconstruction losses Cyclic loss Orthogonal loss Identity confrontation loss and identity consistency loss The weighting coefficients.

[0055] S105: The constructed ISFDNet is trained using the training set and the image-level low-frequency perturbation module LFDM. The original images in the training set and the stylized images generated by the image-level low-frequency perturbation module LFDM are paired and input into ISFDNet. The network parameters are optimized through the collaborative supervision loss function to obtain the trained ISFDNet.

[0056] During training, the original image and the stylized image are input into ISFDNet in pairs. The original image is processed by the backbone network to obtain the hybrid features of the original image. Hybrid features of stylized images , and The image is then fed into the Feature Decoupling Module (FDM), which outputs the identity feature vector of the original image. and style feature vector and the identity feature vector of stylized images and style feature vector ;Will and Combined, input into the decoder, to obtain reconstructed hybrid features , and Combined, input to decoder, to obtain stylized reconstructed hybrid features ;Will The input feature decoupling module (FDM) outputs a cyclic identity feature vector. ; and Used for identity supervision loss and style invariance loss. Used to reconstruct losses, and Used for cyclic loss, and Used for orthogonal loss and Used to combat identity-related losses.

[0057] S106: The image to be recognized is input into the trained ISFDNet. The decoupled identity feature vector is extracted, normalized, and then matched with the database template features to complete the identity authentication.

[0058] Example 2: Based on Example 1, this example refines the above method, such as... Figure 1 As shown, it includes the following steps:

[0059] S101: In this embodiment, the selected biometric object is palm print. Four common palm print datasets are used to construct an image dataset, which is then divided into a training set and a test set. The ratio of the training set to the test set used in this embodiment is 1:1. Preprocessing and data augmentation operations are performed on the image data in the training set; specifically, the image size is uniformly adjusted to 128. 128; Data augmentation operations include random cropping and scaling (scale 0.8-1.0 to simulate different distances between the lens and the hand during shooting), random perspective transformation (distortion scale 0.15 to simulate different hand tilts and angles during shooting), and random rotation (within 10 degrees to simulate slight differences in hand orientation during shooting); finally, the image is normalized to have a mean of 0 and a standard deviation of 1.

[0060] S102: Construct the identity and style feature decoupling network ISFDNet, which adopts an encoder-decoder architecture;

[0061] The encoder decouples the input image and outputs corresponding identity feature vectors and style feature vectors. The encoder includes a backbone network and a feature decoupling module (FDM). The backbone network extracts hybrid features from the input image. The Feature Decoupling Module (FDM) is used to decouple mixed features. The decoupling is performed on identity feature vectors and style feature vectors. The backbone network can adopt any existing biometric recognition network structure, such as the coordinate-aware contrast competition network CO3Net and the comprehensive competition network CCNet. The feature decoupling module FDM includes a spatial awareness separation module SASM, two channel attention modules, and four fully connected layers.

[0062] Specifically, the Spatial Aware Separation Module (SASM) in the encoder is configured as a pixel-confidence-based signal filter. Due to differences in acquisition devices and ambient lighting, the amount of biological texture feature information contained in different pixel locations in an image varies significantly. The SASM receives mixed features obtained from the backbone network as input, and these mixed features are output as a dual-channel attention logic graph through two convolutional layers. Where C, H, and W represent the number of channels, height, and width, respectively, and the number of channels C=2. These two channels correspond to the attention of identity features and style features, respectively. To generate two complementary and interdependent masks, the Softmax function is applied along the channel dimension of the attention logic graph:

[0063] ;

[0064] in This represents a dual-channel attention logic diagram. This represents the normalized exponential function operation. and These represent the identity mask and style mask, respectively. The identity mask spatially activates pixel regions containing clear textures, bifurcation points, and other stable physiological structures, while the style mask activates pixel regions containing non-biological interference such as shadows and sensor noise. Based on the normalization properties of Softmax, at each spatial location... Above, the attention weights of the identity mask and style mask satisfy... , and Mixed features with the original Element-by-element multiplication generates preliminary separated identity features. and style characteristics These two groups of initial separation features and The input features are processed into two independent parallel branches, each containing a channel attention module and a fully connected layer. The channel attention module assigns channel-dimensional weights to the input features, learning the importance weights of each channel to enhance the response to key feature channels while suppressing irrelevant or redundant channel signals. The feature map after channel attention weighting is then input into the fully connected layer, which flattens the feature map and maps it to a low-dimensional latent space, ultimately generating a highly decoupled identity feature vector. With style feature vector .

[0065] The decoder is used to reconstruct the identity feature vector and style feature vector output by the encoder into a hybrid feature. It does not directly participate in the identity recognition process, but verifies the decoupling effect of the encoder through the reconstructed hybrid feature, ensuring that the separated identity features are pure and the style features are free from identity interference. The decoder includes a spatially aware separation module (SASM), two channel attention modules, two convolutional layers, two batch normalization layers, and a ReLU activation function layer.

[0066] The identity feature vector output by the encoder and style feature vector The features are mapped and processed by the Spatial Aware Separation Module (SASM) in the decoder. (It should be clarified that although the structure of this module is the same as that of the SASM in the encoder, its function in the decoder is different. The SASM in the encoder aims to "separate" features based on pixel confidence, while the SASM in the decoder aims to perform "spatial allocation and guidance," predicting the distribution of identity and style information in the image space, thereby guiding the network to accurately fill the texture detail areas with identity information and restore the style information to the lighting or background areas.) After SASM processing, the two features are respectively fed into the channel attention module for channel-dimensional weighting. Then, they are fused into preliminary reconstructed features by element-wise addition. Finally, the reconstructed features are processed by a convolutional operation unit consisting of convolutional layers, batch normalization layers, and ReLU activation function layers to further restore the local details and texture structure of the image, ultimately outputting the reconstructed hybrid features.

[0067] S103: Construct an image-level low-frequency perturbation module (LFDM) to generate stylized images. The LFDM is located at the network input (i.e., the very front of the entire ISFDNet network). During the training phase, it perturbs the low-frequency components of the original images in the training set to generate stylized images as augmented training samples. Before the original images in the training set enter the ISFDNet encoder structure, they are processed by the LFDM to generate stylized images, which are then input into the encoder in pairs with the original images. Figure 2 As shown, the image-level low-frequency perturbation module LFDM can generate stylized variants, providing targeted stylized training samples. Working in conjunction with the feature-level spatial awareness separation module SASM, it forces the network to learn stable identity representation encoding under style changes.

[0068] Specifically, in the optical physical model of biometric imaging, the image signal can be decomposed into two components: high-frequency components and low-frequency components. The high-frequency components are the texture details such as main lines and wrinkles in the image. This information originates from the physiological structure of the organism's surface and belongs to the identity feature information that should remain unchanged. The low-frequency components are the overall brightness distribution, hue, and contrast of the image. This information mainly depends on the camera's photosensitive chip and the acquisition environment and belongs to the imaging interference information that changes with the environment. Based on this principle, the image-level low-frequency perturbation module LFDM simulates style changes without damaging the identity feature information by perturbing the low-frequency content while retaining the high-frequency structure.

[0069] The operation process is as follows: ① Given a size of The original image ,in The image's height and width are given respectively, and they are transformed to the frequency domain using a two-dimensional discrete Fourier transform: ;in Represents the original image Spatial domain coordinates, For row coordinates, For column coordinates, This represents the frequency domain after the two-dimensional discrete Fourier transform. Represents the normalized frequency in the vertical direction. Represents the normalized frequency in the horizontal direction. This represents the normalized spatial frequency projection. Represents pi (π). Represents the imaginary unit, satisfying , The base of the natural logarithm. Represents row coordinates The value range is from 0 to column coordinates The value range is from 0 to ;

[0070] ②Then a circular low-pass filter mask is applied. The frequency domain is decomposed into low-frequency and high-frequency components, defined as follows:

[0071] ;

[0072] in The preset radius for the low-pass filter is set to [value]. ,in and These are the height and width of the image. Indicates the current frequency domain point To the Spectrum Center The Euclidean distance (i.e., the center of the low-pass filter) is used to obtain the low-frequency component through this mask. and high frequency components :

[0073] ;

[0074] ;

[0075] in This represents a circular low-pass filter mask. This indicates element-wise multiplication. This represents the frequency domain after the two-dimensional discrete Fourier transform.

[0076] ③ After an image undergoes a two-dimensional discrete Fourier transform to the frequency domain, it decomposes into a phase spectrum and an amplitude spectrum. The phase spectrum encodes the image's feature structure and spatial arrangement, such as the direction of palm print lines, the location of bifurcation points, and the spatial distribution of edge contours—these are the core elements of identity recognition. The amplitude spectrum, on the other hand, reflects the image's contrast and intensity, such as overall brightness, local differences in brightness and darkness, and the strength of noise—these are style interferences affected by device / lighting. Therefore, we keep the high-frequency components unchanged and the phase spectrum of the low-frequency components unchanged. We then apply a composite perturbation function to the amplitude spectrum of the low-frequency components to obtain the perturbed low-frequency component amplitude spectrum. We merge the perturbed low-frequency component amplitude spectrum with the original low-frequency component phase spectrum to obtain the perturbed low-frequency component. Finally, we merge the perturbed low-frequency component with the original high-frequency components to form new frequency domain data, which is then reconstructed into a stylized image through an inverse Fourier transform.

[0077] Furthermore, the image-level low-frequency perturbation module LFDM uses a composite perturbation function to perturb the amplitude spectrum of the low-frequency components to simulate style interference observed in real-world environments. Specifically, this composite perturbation function integrates three simulation strategies, including global amplitude scaling, adaptive noise injection, and smooth illumination field simulation.

[0078] ① The global amplitude scaling strategy simulates the differences in overall image brightness and contrast caused by different acquisition devices and acquisition environments. The model is as follows:

[0079] ;

[0080] in For random sampling scaling factor, when At that time, it simulates a "low light / shadow style" imaging environment with insufficient light, underexposure, or low contrast. At that time, it simulates a "high-brightness" imaging environment characterized by strong light, overexposure, or high contrast. Settings The range can cover common imaging illumination fluctuations, while avoiding damage to high-frequency texture structures containing key identity information due to factors that are too small (resulting in a completely black image) or too large (resulting in pixel saturation and overflow). This represents the amplitude of the low-frequency components of the image. This is the result of applying a global amplitude scaling strategy to the low-frequency component amplitude spectrum of the image.

[0081] ② An adaptive noise injection strategy is used to simulate the impact of sensor-generated noise on image quality. First, the noise standard deviation is adaptively calculated based on the mean of the non-zero values ​​in the amplitude spectrum: ;

[0082] in, It is the noise standard deviation. It is the mean of the non-zero values ​​in the amplitude spectrum. It is the noise scaling factor of random sampling. The lower limit is used to simulate high-end acquisition equipment with high signal-to-noise ratios, introducing only a very small amount of thermal noise. The upper limit simulates the significant electronic noise produced by low-cost cameras or at high ISO sensitivity settings. (Settings) The upper limit is set to prevent excessive noise amplitude from completely masking the frequency domain information, thus ensuring that sufficient signals to reconstruct the basic structure of the image are preserved while introducing style interference. A Gaussian noise field is then generated. And superimposed on the amplitude spectrum, Indicates a normal distribution. This represents the noise variance, while ensuring the non-negativity of the amplitude values ​​through the following methods:

[0083] ;

[0084] in, This indicates the operation of retrieving the maximum value. Represents a Gaussian noise field. It is related to the amplitude of the low-frequency component. Non-negative real numbers with consistent dimension are the amplitudes of low-frequency components. The result after one perturbation Indicates and Non-negative real numbers with consistent dimension are the amplitudes of low-frequency components. The result after two perturbations is also the result after the adaptive noise injection strategy.

[0085] ③ The smooth lighting field simulation strategy aims to simulate the uneven light spots and shadow interference formed on the skin surface by side lighting, backlighting, or point light sources using a multiplicative smoothing field. First, a standard Gaussian random field is generated, which is then smoothed and normalized using a Gaussian filter to obtain a continuously varying surface. Subsequently, the light intensity factor was randomly sampled. To generate a light field :

[0086] ;

[0087] Represents a smooth Gaussian random field. Represents the light intensity factor, when When the value is positive, the generated illumination field acts on the amplitude spectrum, producing local gain, simulating the reflected or bright areas under point light source illumination. When a negative value is taken, the generated illumination field will experience local attenuation, simulating shadow areas caused by occlusion or backlighting, thus limiting the amplitude to... Within this range, it ensures that the amplitude of illumination changes is within a reasonable physical fluctuation range, which increases the diversity of samples and avoids information loss due to excessive darkening or pixel overflow due to overexposure, resulting in multiplicative textures. It was used as an illumination interference map in the amplitude spectrum:

[0088] ;

[0089] in, This indicates the operation of retrieving the maximum value. Indicates and Non-negative real numbers with consistent dimension are the amplitudes of low-frequency components. The result after two perturbations It is a multiplicative texture. This represents the result after applying a smoothed illumination field simulation strategy. By combining these strategies, the Image-Level Low-Frequency Perturbation Module (LFDM) generates training samples with diverse styles while maintaining consistent identity information.

[0090] S104: Construct a collaborative supervised loss function. By integrating multiple loss functions, the feature space is constrained from complementary dimensions, such as... Figure 3 As shown, these objectives collectively guide the identity and style feature decoupling network ISFDNet to learn clean, independent, and style-invariant identity representations; the collaborative supervision loss function comprises three core components: identity supervision loss to ensure discriminativeness, decoupling supervision loss to enhance feature separation, and style invariance loss to improve cross-domain and cross-sensor generalization capabilities.

[0091] ① Identity supervision loss ensures that the generated identity representation has significant identifiable features. Identity supervision loss includes cross-entropy loss and triplet loss.

[0092] Cross-entropy loss The main objective is to minimize the difference between the predicted sample and the true label, making the predicted value as close as possible to the true label; the cross-entropy loss function is as follows: ;in, The cross-entropy loss function is used, where N and M represent the number of identities and the number of samples for each identity, respectively. and These are the label and predicted probability of the i-th sample in the c-th identity, respectively.

[0093] Cross-entropy loss reduces intra-class tightness but cannot guarantee clear inter-class separation. In contrast, triplet loss... Simultaneously constraining intra-class tightness and inter-class separation, making features of the same class closer together and features of different classes further apart, the mathematical expression for triplet loss is: ;in, For triple loss, This indicates the operation of retrieving the maximum value. This indicates a minimum value operation, where A is the anchor sample, P is the positive sample, N is the negative sample, and margin is the default distance threshold, which is set to 2 in this embodiment. It represents Euclidean distance.

[0094] In summary, identity surveillance losses as follows:

[0095] ;

[0096] in, and Cross-entropy loss and triplet loss The weight.

[0097] ②Decoupling supervision loss includes four types of loss functions: reconstruction loss, recurrence loss, orthogonal loss, and identity adversarial loss;

[0098] I. Reconstruction Losses This prevents the loss of key information in the original image during decoupling. Through the decoder, the separated identity and style features are recombined to reconstruct the hybrid features, which are then compared with the original hybrid features extracted by the backbone network during forward propagation.

[0099] ;

[0100] in, To rebuild the losses, These are hybrid features extracted from the backbone network. It is a hybrid feature reconstructed by the decoder using identity features and style features. Represents cosine similarity. Describing the L1 norm, and These are the weighting coefficients for L1 loss and cosine similarity, respectively. This process can quantify the information loss during decoupling and recombination, ensuring that key information is not discarded during decoupling.

[0101] II. Cyclic Loss To verify and ensure a clear and thorough decoupling of identity and style factors, the identity feature vector of the original image is combined with the style feature vector of the stylized image. This fused representation is then passed through a decoder to obtain the stylized reconstruction hybrid features. In theory, stylized reconstruction can reconstruct hybrid features. It should be able to preserve the original identity information of the subject under style perturbation. Accordingly, this invention addresses the stylization reconstruction of hybrid features. Re-encode, and make the cyclically encoded cyclic identity feature vector By applying the consistency objective constraint, it is aligned with the identity feature vector of the original image. Highly matched. The calculation formula is as follows:

[0102] ;

[0103] in, Indicates cyclical loss. This represents the identity feature vector of the original image. Represents a cyclic identity feature vector. This represents the L2 norm.

[0104] III. Next, design the orthogonal loss function. By enforcing the mathematical independence between identity features and style features, this orthogonal loss function consists of pairwise orthogonal terms. and batch of orthogonal items constitute:

[0105] ;

[0106] Specifically, It is an orthogonal loss function. To approve orthogonal terms, These are paired orthogonal terms used to ensure that the identity features and style features of a single sample are mutually orthogonal. Their expression is:

[0107] ;

[0108] in and Representing the identity feature vector and style feature vector of the original image, Describing the L2 norm, This represents taking the expected value. Furthermore, it introduces... Batch constraints constrain the independence of the feature distribution across the entire batch by calculating the covariance matrix of all identity and style features within that batch.

[0109] ;

[0110] in and Representing the identity feature vector and style feature vector of the original image, This indicates that the identity feature vector of the original image is transposed. Describing the L2 norm, Indicates taking the expected value. This represents the square of the van Hilbert-Schmidt norm.

[0111] IV. Although orthogonalization forcibly separates identity and style features, residual information in the style branch may still be used for identity recognition. Therefore, this embodiment employs a dedicated adversarial training mechanism to further refine the style representation, ensuring it contains no identity cues and strictly maintains non-discriminatoryness. The same identity adversarial loss as the formula above is used during training. By combining a gradient inversion layer with style features to predict a user’s identity, this adversarial process forces style features to discard potential identity cues, thereby achieving complete decoupling of identity and style.

[0112] ③ Style invariance loss is achieved through identity consistency loss To achieve this, identity consistency loss ensures that identity representation remains invariant under diverse style perturbations. This loss function minimizes the angular distance between the original image and the stylized image in terms of identity features, ensuring that the two sets of features remain aligned in the feature space. The formula is shown below:

[0113] ;

[0114] in, This indicates a loss of identity consistency. The identity feature vector representing the original image. Represents the identity feature vector of a stylized image. This represents the L2 norm.

[0115] Final loss function As shown below:

[0116] ;

[0117] in, , , , , , Losses related to identity supervision Reconstruction losses Cyclic loss Orthogonal loss Identity confrontation loss and identity consistency loss The weighting coefficients.

[0118] S105: The constructed identity and style feature decoupling network ISFDNet is trained using the training set; during training, the original image and the stylized image generated by the image-level low-frequency perturbation module LFDM are input into the encoder in pairs, and the original image is processed by the backbone network to obtain the hybrid features of the original image. and stylized images , and The image is then fed into the Feature Decoupling Module (FDM), which outputs the identity feature vector of the original image. and style feature vector and the identity feature vector of stylized images and style feature vector ;Will and Combined, input into the decoder, to obtain reconstructed hybrid features , and Combined, input to decoder, to obtain stylized reconstructed hybrid features ;Will The input feature decoupling module (FDM) outputs a cyclic identity feature vector. ; and Used for identity loss and style invariance loss. Used to reconstruct losses, and Used for cyclic loss, and Used for orthogonal loss and Used for identity adversarial loss; the network parameters are optimized using a collaborative supervised loss function so that the network learns style-invariant identity representations.

[0119] S106: Feature recognition is performed using the trained Identity and Style Feature Decoupling Network (ISFDNet). During testing or practical application, the image to be recognized is input into the trained ISFDNet encoder, and the identity feature vector is extracted by the Feature Decoupling Module (FDM). After L2 normalization of the vector, the identity authentication is completed by matching the cosine similarity or Euclidean distance with the template features in the database.

[0120] Example 3: Building upon Example 2, the effectiveness of the method described in this invention in cross-domain biological texture feature recognition tasks is verified through experiments. To objectively evaluate the performance of the identity and style feature decoupling network ISFDNet proposed in this invention, palm prints are used as representative biological texture features in this example. Four public datasets were selected for testing: the Hong Kong Polytechnic University palm print dataset PolyU, the Tongji University palm print dataset Tongji, the Indian Institute of Technology Delhi palm print dataset IITD, and the Multi-Spectral palm print dataset Multi-Spectral.

[0121] It should be clarified that using equal error rate (EER) and such Figures 4-10The Receiver Operating Characteristic (ROC) curve shown is used as an evaluation metric. The Receiver Emission Rate (EER), a key validation metric, corresponds to the point where the False Acceptance Rate (FAR) and False Rejection Rate (FRR) are equal. A lower EER value indicates higher recognition accuracy. The ROC curve is a widely used tool for evaluating recognition performance, typically plotted with the False Acceptance Rate (FAR) on the x-axis and the True Acceptance Rate (GAR) on the y-axis. Here, the False Rejection Rate (FRR) = 1 - GAR: the True Acceptance Rate (GAR) measures the proportion of correctly accepted true matches, while the False Acceptance Rate (FAR) quantifies the proportion of incorrectly accepted spoofed matches. The closer the ROC curve is to the upper left corner, the better the recognition performance. The Multi-Spectral palmprint dataset includes four bands: red, green, blue, and near-infrared (NIR). The red, green, and blue bands fall within the visible light range and primarily focus on capturing the rich textural details of the palm surface, such as main lines and irregular folds. The near-infrared band, on the other hand, has strong tissue penetration, enabling further revelation of subcutaneous venous structures and is less affected by surface dirt or lighting conditions. These four bands were selected as the basis for experiments to simulate extreme cross-domain imaging scenarios: images in different bands exhibit significant "style" differences in grayscale distribution, texture clarity, and imaging mechanisms. By testing on such cross-spectral data, this invention can objectively verify whether it can effectively decouple "spectral style" from "biological identity," thereby demonstrating the robustness and generalization ability of the method in dealing with changes in imaging spectra.

[0122] Table 1. Equal error rate (%) across datasets where the source dataset is Tongji.

[0123]

[0124] Table 2. Equal error rates (%) across datasets where the source datasets are PolyU and IITD.

[0125]

[0126] Red, Blue, Green, and NIR represent the red band, blue band, green band, and near-infrared band, respectively. CO3Net, CCNet, SACNet, MSPHNet, and PalmRSS represent the coordinate-aware contrastive competitive neural network, the comprehensive competitive network, the scale-aware competitive network, the multi-scale parallel hybrid network, and the single-source domain generalization network, respectively. ISFDNet-CO3Net is a hybrid model formed with CO3Net as the backbone network, and ISFDNet-CCNet is a hybrid model formed with CCNet as the backbone network. PolyU, Tongji, IITD, and Multi-Spectral are the palmprint datasets from Hong Kong Polytechnic University, Tongji University, the Indian Institute of Technology Delhi, and the multispectral palmprint dataset, respectively. This invention uses CO3Net and CCNet as the backbone networks of the identity and style feature decoupling network ISFDNet. As shown in Tables 1 and 2, this invention performs benchmark tests against CO3Net, CCNet, SACNet, MSPHNet, and PalmRSS through cross-database evaluation. In both ISFDNet variants, the model supported by CCNet typically achieves a lower equal error rate, indicating that CCNet, as a stronger backbone network, can provide more discriminative basic features for decoupling and produce better performance after decoupling.

[0127] Notably, ISFDNet effectively enhances the domain generalization ability of the backbone network: both CCNet and CO3Net exhibit reduced cross-database error rates after incorporating the decoupling framework. When using Tongji as the source domain and Green as the target domain, the CO3Net error rate in the ISFDNet identity and style feature decoupling network is reduced by 83%, demonstrating that ISFDNet, as a shared feature decoupler, can be easily migrated to existing architectures and supports the generation of more robust identity embeddings. Cross-database results strongly demonstrate the identity-style decoupling paradigm—by explicitly separating identity stabilization factors and domain style factors at the feature layer, ISFDNet can learn clearer and more robust identity encodings, thereby improving the generalization ability of cross-domain biological texture features.

[0128] Table 3. Equal error rates (%) for source spectra of NIR and Red across the spectrum.

[0129]

[0130] Table 4. Equal error rates (%) for source spectra of Green and Blue across the spectrum.

[0131]

[0132] Tables 3 and 4 report the cross-spectral generalization performance of the Identity and Style Feature Decoupling Network (ISFDNet) on multispectral datasets, and compare it with CO3Net, CCNet, SACNet, MSPHNet, LSFM, and PalmRSS. ISFDNet performs exceptionally well in most cross-spectral settings, significantly outperforming the original backbone network and baseline models. When using CCNet as the backbone, the model achieves an equal error rate of 0% across various configurations, demonstrating outstanding cross-spectral performance. It is noteworthy that performance decreases when the source or target domain involves the near-infrared spectrum, although overall performance still improves. This difference stems from the fundamental difference between near-infrared and other spectra: near-infrared light can penetrate the skin and reveal subcutaneous veins. Therefore, the transfer from near-infrared to visible light is a cross-modal problem, not style transfer.

[0133] Figures 4-10 ROC curves of the identity and style feature decoupling network ISFDNet and the contrastive method were plotted on multiple datasets. Figures 4 to 10 These are the PolyU, Tongji, IITD, Red, Green, Blue, and NIR datasets. The performance trends remain consistent across different sensors and acquisition settings. The ISFDNet curve, representing the decoupling of identity and style features, is close to the top left corner, indicating that this model maintains an extremely high true acceptance rate (GAR) across all false acceptance rate (FAR) intervals.

[0134] Table 5. Intra-domain error rates (%) for the three datasets

[0135]

[0136] Table 6. Intra-domain error rates (%) under the Multi-Spectral dataset

[0137]

[0138] PalmCode, CompCode, OrdinalCode, FusionCode, RLOC, BOCV, E-BOCV, HOC, DOC, DRCC, 2TCC, and MTCC are, in order, palmprint codes, contested codes, ordinal codes, fusion codes, robust linear directional codes, binary directional co-occurrence vector codes, binary directional co-occurrence vector extended codes, half-directional codes, bidirectional codes, discriminative robust contested codes, second-order texture extended codes, and multi-order texture extended codes. DHPN, PalmNet, DHN, and CompNet are, in order, deep hashing palm vein networks, palmprint networks, deep hashing networks, and contested networks. Table 5 reports the in-domain results of traditional and deep methods on multiple datasets. On the PolyU and Tongji datasets, both ISFDNet variants of the identity and style feature decoupling network achieved 0% error rate, reaching state-of-the-art performance. In the more challenging IITD benchmark, the ISFDNet identity and style feature decoupling network also demonstrated high competitiveness. These results confirm the optimization quality and low variance of the models. As summarized in Table 6, the Identity and Style Feature Decoupling Network (ISFDNet) also achieved outstanding results in evaluations within the multispectral dataset domain. Both CCNet and PalmRSS versions of ISFDNet achieved an equal error rate of 0%, demonstrating stability and robustness under different spectral conditions. This performance remained constant under varying sessions and lighting conditions.

[0139] Figure 11 This demonstrates the visual separation effect of the Spatial Aware Separation Module (SASM) on the input samples. The identity mask and style mask, generated through the Softmax mechanism, force the sum of the weights at each location to be 1, thus exhibiting a high degree of complementarity: high identity saliency (dark red) corresponds to low style saliency (light blue), and vice versa. Overlap is measured by calculating the element-wise minimum of the two masks to assess co-activation at the same spatial location; a large area of ​​light purple indicates clear attribution and effective decoupling. Coverage reflects the overall feature utilization by taking the maximum value between the masks; a predominantly dark green coloring indicates extensive spatial use with minimal waste. Mask dissimilarity measures dominance by calculating the absolute difference between masks point-by-point; darker dark orange areas indicate clearer and more explicit assignment of identity and style. Figure 12 The changes in mask overlap and coverage during training were recorded: as training progressed, overlap decreased while coverage increased, demonstrating that the collaborative supervision strategy can continuously enhance and stabilize feature decoupling capabilities.

[0140] Table 7 Ablation Experiment

[0141]

[0142] To verify the effectiveness and necessity of each component of the collaborative supervision strategy, ablation experiments were conducted in this embodiment. As shown in Table 7, This represents reconstruction loss and cyclical loss. Indicates orthogonal loss. Indicates identity as a means of combating loss. The table shows the identity consistency loss, with √ and × indicating active and inactive status, respectively. Experiments were conducted across databases (trained in PolyU as the source domain and tested in the unseen target domains Tongji and IITD). The baseline model, trained using only the identity loss, achieved equal error rates of 2.4334% and 6.7391% on Tongji and IITD, respectively. Introducing reconstruction and cyclic losses to the baseline model significantly reduced the equal error rates in both target domains. This demonstrates that ensuring the integrity and effectiveness of information during decoupling is fundamental to the entire framework. A key finding is that using any orthogonal loss alone (identity adversarial loss or identity consistency loss) leads to a decrease in generalization performance compared to the baseline. This highlights the inherent correlation between constraints—if feature effectiveness is not ensured through reconstruction and cyclic losses, applying orthogonality, adversarial separation, or consistency constraints alone may result in weak discriminative solutions, thus impairing recognition accuracy. When all loss functions work together, the identity and style feature decoupling network ISFDNet achieves optimal performance, with equal error rates of 1.8074% and 4.2986% on PolyU and IITD, respectively. Notably, on the more challenging IITD dataset, the equal error rate decreases by approximately 36% compared to the baseline. Ablation experiments demonstrate that although individual constraints are ineffective when acting alone, they are indispensable in the ensemble framework. Reconstruction loss and cyclic loss lay the foundation for effective decoupling, while the remaining objectives optimize the feature space along complementary dimensions, collectively achieving optimal generalization performance. Therefore, collaborative supervision strategies are crucial for achieving robust biological texture feature recognition.

[0143] While embodiments of the present invention have been described in detail above, it will be apparent to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations fall within the scope and spirit of the invention as set forth in the claims. Furthermore, the invention described herein may have other embodiments and can be implemented or carried out in various ways.

Claims

1. A method for cross-domain biological texture feature recognition resistant to imaging interference, characterized in that, Includes the following steps: S101: Construct a biological texture image dataset, and divide the image dataset into a training set and a test set in a 1:1 ratio; wherein, the biological texture includes fingerprint, iris, palm print, and face texture; S102: Construct the identity and style feature decoupling network ISFDNet. The network ISFDNet adopts an encoder-decoder architecture. The encoder is used to decouple the input image and output the corresponding identity feature vector and style feature vector. The decoder is used to reconstruct the identity feature vector and style feature vector output by the encoder into a hybrid feature. S103: Construct an image-level low-frequency perturbation module LFDM, which is used to perturb the low-frequency components of the original images in the training set to generate stylized images; S104: Construct a collaborative supervision loss function, which includes identity supervision loss, decoupling supervision loss, and style invariance loss; S105: The constructed ISFDNet is trained using the training set and the image-level low-frequency perturbation module LFDM. The original images in the training set and the stylized images generated by the image-level low-frequency perturbation module LFDM are paired and input into ISFDNet. The network parameters are optimized through the collaborative supervision loss function to obtain the trained ISFDNet. S106: The image to be recognized is input into the trained ISFDNet, the decoupled identity feature vector is extracted, and after normalization, it is matched with the database template features to complete the identity authentication. In step S102, the encoder includes a backbone network and a feature decoupling module FDM. The backbone network is used to extract mixed features from the input image, and the feature decoupling module FDM is used to decouple the mixed features into identity feature vectors and style feature vectors. The backbone network can adopt a biometric recognition network structure. The feature decoupling module FDM includes a spatial awareness separation module SASM, a channel attention module, and a fully connected layer. The decoder includes a spatial awareness separation module (SASM), a channel attention module, a convolutional layer, a batch normalization layer, and a ReLU activation function layer. In step S102, the spatially aware separation module (SASM) in the encoder incorporates convolutional layers and a softmax function. It receives the mixed features output from the backbone network, generates a dual-channel attention logic graph through the convolutional layers, and uses the softmax function to generate complementary identity masks in the channel dimension. and style mask Identity mask and style mask in spatial location The sum of the weights on each feature is 1; the identity mask and style mask are multiplied element-wise with the hybrid feature to obtain the initially separated identity feature and style feature; The initially separated identity features and style features are fed into two independent parallel branches. Each branch has a channel attention module and a fully connected layer. The channel attention module is used to weight the input features according to the channel dimension. By learning the importance weight of each channel, it enhances the response to key feature channels while suppressing irrelevant or redundant channel signals. The feature map after channel attention weighting is then fed into the fully connected layer. The fully connected layer flattens the feature map and maps it to a low-dimensional latent space, ultimately generating highly decoupled identity feature vectors and style feature vectors. In step S104, the identity supervision loss includes cross-entropy loss and triplet loss, the decoupling supervision loss includes reconstruction loss, loop loss, orthogonal loss and identity adversarial loss, and the style invariance loss includes identity consistency loss.

2. The method for cross-domain biological texture feature recognition against imaging interference according to claim 1, characterized in that, In step S103, the specific processing procedure of the image-level low-frequency perturbation module LFDM is as follows: the input image is transformed to the frequency domain using a two-dimensional discrete Fourier transform; the frequency domain is decomposed into low-frequency components and high-frequency components using a low-pass filter mask; the high-frequency components are kept unchanged, the phase spectrum of the low-frequency components is kept unchanged, and the amplitude spectrum of the low-frequency components is processed by a composite perturbation function to obtain the perturbed low-frequency components; the perturbed low-frequency components are merged with the original high-frequency components, and reconstructed into a stylized image through inverse Fourier transform.

3. The method for cross-domain biological texture feature recognition against imaging interference according to claim 2, characterized in that, The composite perturbation function includes a combination of three strategies: global amplitude scaling, adaptive noise injection, and smoothed illumination field simulation. Specifically, global amplitude scaling multiplies the low-frequency amplitude by a random sampling scaling factor to simulate changes in brightness and contrast; adaptive noise injection calculates the noise standard deviation based on the mean of the low-frequency amplitude, generates a Gaussian noise field, and superimposes it onto the amplitude spectrum to simulate sensor noise; and smoothed illumination field simulation generates a smoothed Gaussian random field as an illumination interference map, which is applied to the amplitude spectrum in a multiplicative manner to simulate non-uniform illumination changes.

4. The method for cross-domain biological texture feature recognition against imaging interference according to claim 3, characterized in that, In step S104, the collaborative supervision loss function The calculation formula is: ; in, , , , , , Losses related to identity supervision Reconstruction losses Cyclic loss Orthogonal loss Identity confrontation loss and identity consistency loss The weighting coefficients.

5. The method for cross-domain biological texture feature recognition against imaging interference according to claim 4, characterized in that, In step S105, during training, the original image and the stylized image are input into ISFDNet in pairs. The original image is processed by the backbone network to obtain the hybrid features of the original image. Hybrid features of stylized images , and The image is then fed into the Feature Decoupling Module (FDM), which outputs the identity feature vector of the original image. and style feature vector and the identity feature vector of stylized images and style feature vector ;Will and Combined, input into the decoder, to obtain reconstructed hybrid features , and Combined, input to decoder, to obtain stylized reconstructed hybrid features ;Will The input feature decoupling module (FDM) outputs a cyclic identity feature vector. ; and Used for identity supervision loss and style invariance loss. Used to reconstruct losses, and Used for cyclic loss, and Used for orthogonal loss and Used to combat identity-related losses.

Images