Palm print identification anti-fraud method based on double-branch self-supervised learning

Abstract

The invention relates to the technical field of data analysis, in particular to a palmprint recognition anti-fraud method based on double-branch self-supervised learning, and the method comprises the steps: constructing a double-branch self-supervised reconstruction architecture, determining a first to-be-processed image based on a high-frequency self-supervised branch, and determining a second to-be-processed image based on a chromaticity regularization self-supervised branch. Dividing the two images into patch blocks which are not overlapped with each other, independently generating random mask matrixes with different spatial distributions, determining visible areas, inputting the patch blocks of the visible areas into a shared encoder for feature extraction, outputting latent variable features, and splicing the latent variable features with learnable mask marks to obtain a patch matrix; and respectively sending to a high-frequency decoder and a chroma decoder, determining total loss, carrying out physical decoupling on illumination and material attributes, and carrying out fine adjustment on the architecture to finish convergence of the architecture. According to the method, the palm print features with discrimination are extracted through the shared encoder, and chromaticity distribution consistency constraints are introduced, so that effective decoupling of ambient light and real material attributes is realized.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a palmprint recognition anti-fraud method based on dual-branch self-supervised learning. Background Technology

[0002] Palmprint recognition, as a non-contact and highly convenient biometric identification technology, has been widely applied in payment and security fields. Current mainstream methods are based on deep learning, using CNNs or Vision Transformers to train classification models on large-scale labeled data. However, these supervised methods have significant limitations in real-world cross-media scenarios: different acquisition devices and attack media can introduce color gamut shifts, causing models to easily overfit specific color features, leading to decreased generalization performance under varying lighting conditions or new attack media. Simultaneously, ambient light interference and material properties are highly coupled in imaging, and traditional methods struggle to physically decouple them. Intense lighting can mask subtle color and texture differences between genuine and fake palmprints, affecting model stability. Furthermore, the demand for real-time deployment of edge devices is increasingly prominent, and existing complex models struggle to meet the requirements of lightweight design and low latency. Although self-supervised learning performs exceptionally well in the vision domain, lightweight self-supervised architectures for palmprint anti-fraud are still lacking, and most methods fail to fully leverage high-frequency texture and inherent material features in their modeling.

[0003] Chinese Patent Publication No. CN110458662A discloses an anti-fraud risk control method and device. The method includes: collecting user and enterprise information data, including: obtaining user information and enterprise information input by the user; recognizing information in the document image uploaded by the user through OCR; obtaining enterprise information data from the enterprise's business platform; and obtaining enterprise information data on the consortium blockchain through a preset interface; verifying the user information in the user and enterprise information data; after the user information verification is passed, verifying the enterprise information in the user and enterprise information data; and after the enterprise information verification is passed, performing anti-fraud risk assessments of the actual controller, legal person, and enterprise respectively in a trusted execution environment built on the SGX multi-party secure computation server based on the user and enterprise information data. This invention improves the data source, ensures the authenticity and security of the data, and improves the accuracy of anti-fraud risk control.

[0004] Chinese Patent Publication No. CN114841705A discloses a scene-based fraud detection method, comprising: pre-training multiple independent anti-fraud sub-models; determining the user's network environment after the user goes online, and acquiring the user's call records, SMS records, and web browsing records within a pre-set time period, and capturing the user's traffic data to form a first subsample; determining the temporal correlation between the data in the first subsample, and determining the input dataset based on the coarse labeling results; determining at least some anti-fraud sub-models from the multiple independent anti-fraud sub-models based on the benchmark dataset with the highest similarity, and determining the weights of at least some anti-fraud sub-models based on the benchmark dataset with the highest similarity; and combining at least some anti-fraud sub-models, using the input dataset as input, to determine the user's fraud risk. This application's method configures the weights of sub-models based on the temporal correlation of fraud-related events, greatly improving the independence of the models and the flexibility of configuration.

[0005] However, the following problems still exist in the existing technology. Existing high-performance anti-fraud models are usually complex in structure and computationally expensive, making it difficult to meet the stringent real-time response requirements of mobile terminals or embedded handheld devices. Summary of the Invention

[0006] To address this issue, the present invention provides a palmprint recognition anti-fraud method based on dual-branch self-supervised learning, which solves the problem that existing high-performance anti-fraud models are usually complex in structure and have high computational cost, making it difficult to meet the strict requirements of real-time response for mobile terminals or embedded palm print devices.

[0007] To achieve the above objectives, this invention provides a palmprint recognition anti-fraud method based on dual-branch self-supervised learning, comprising: A dual-branch self-supervised reconstruction architecture is constructed, including a shared encoder, a high-frequency decoder, and a chroma decoder. Domain-sensitive features are decoupled based on two parallel reconstruction branches, including a high-frequency self-supervised branch and a chroma regularization self-supervised branch. Input the original palm print image, determine the first image to be processed based on the high-frequency self-supervised branch, and determine the second image to be processed based on the chroma regularization self-supervised branch; The first image to be processed and the second image to be processed are divided into non-overlapping patch blocks, and each patch block is independently generated with a random mask matrix with a different spatial distribution to determine the visible area. The patch blocks of each visible region are input into the shared encoder for feature extraction to capture the global dependencies between image patches and output latent variable features; Each latent variable feature is concatenated with a learnable mask label and fed into the high-frequency decoder and the chroma decoder respectively to calculate the first loss and the second loss, determine the total loss, and physically decouple the illumination and material properties. The high-frequency decoder and the chroma decoder are removed, and the shared encoder that has learned the physical decoupling features is retained. The output is global average pooled at the output end of the shared encoder and then connected to a linear classification layer to output the authenticity judgment result. A small number of original palm print images with authenticity labels are used to fine-tune the architecture to complete the convergence of the architecture and perform palm print recognition anti-fraud.

[0008] Furthermore, the process of determining the first image to be processed includes: The original palmprint image is converted to the frequency domain using a two-dimensional discrete Fourier transform, and a random radius is defined. The random radius is sampled from a preset range to construct a circular high-pass filter; The first image to be processed is obtained by filtering out low-frequency components and retaining high-frequency spatial information at different scales.

[0009] Furthermore, the process of determining the second image to be processed includes: The original palm print image is determined to be a three-dimensional matrix containing three color channels: red, green, and blue. Randomly select a color channel and set its value to 0 to obtain the second image to be processed.

[0010] Furthermore, the process of determining the visible area includes: A predetermined proportion of patch blocks in each of the aforementioned random mask matrices are determined to be visible blocks; The set of visible blocks is determined to be a visible region.

[0011] Furthermore, the global dependencies between image patches are captured through the Vision Transformer architecture based on a global attention mechanism.

[0012] Furthermore, the latent variable features include high-frequency texture and low-frequency chromaticity.

[0013] Furthermore, the process of calculating the first loss includes: Calculate texture-guided loss based on gradient vector cosine similarity; The mean square error loss between the reconstructed pixel values ​​and the original pixel values ​​of the masked region in the first image to be processed is determined as the pixel-level reconstruction loss. The weighted sum of the texture-guided loss and the pixel-level reconstruction loss is determined to be the first loss.

[0014] Furthermore, the process of calculating the second loss includes, The mean square error loss between the reconstructed pixel values ​​and the original pixel values ​​of the masked region in the second image to be processed is determined as the color pixel-level reconstruction loss. Two sets of log-chromaticity ratio maps are constructed based on the channels retained and the channels discarded in the second image to be processed. Each of the aforementioned mapping maps and the target mapping map is spatially flattened and then ordered in ascending order to obtain the corresponding empirical distribution sequence. The difference between each of the empirical distribution sequences is calculated using distance calculation to obtain the chromaticity distribution loss; The weighted sum of the color pixel reconstruction loss and the chromaticity distribution loss is determined to be the second loss.

[0015] Furthermore, the process of determining the total loss includes: The weighted sum of the first loss and the second loss is determined to be the total loss.

[0016] Furthermore, the process of physically decoupling lighting and material properties includes: Calculate the logarithm of the channel ratio to suppress ambient light interference and achieve separation of lighting and material properties; In the channel ratio calculation, the ambient light component is canceled out, and only the relative reflectivity information reflecting the material properties is retained.

[0017] Compared with existing technologies, this invention constructs a dual-branch self-supervised reconstruction architecture. A first image to be processed is determined based on a high-frequency self-supervised branch, and a second image is determined based on a chroma regularization self-supervised branch. The two images are divided into non-overlapping patch blocks, each generating its own spatially distinct random mask matrix to determine the visible region. The patch blocks of each visible region are input into a shared encoder for feature extraction, outputting latent variable features. These features are concatenated with learnable mask markers and fed into the high-frequency decoder and chroma decoder respectively to determine the total loss. Physical decoupling of illumination and material properties is achieved, and the architecture is fine-tuned to achieve convergence. This invention extracts discriminative palmprint features through a shared encoder and introduces chroma distribution consistency constraints to effectively decouple ambient lighting from real material properties.

[0018] In particular, by designing a lightweight architecture based on dual-branch self-supervised learning, this invention decouples domain-sensitive features based on two parallel reconstruction branches. In existing technologies, most rely on single feature modalities (such as pure texture analysis or image classification) or complex domain adaptive networks. These methods often fail to deeply integrate with the imaging physics mechanism and systematically separate and suppress key domain interference factors such as illumination changes. Furthermore, due to the lack of structured modeling of the illumination effect during image formation, existing models struggle to effectively decouple domain-sensitive features in cross-domain scenarios, resulting in limited extraction of essential features and decreased generalization performance. Based on this, this invention considers using high-frequency self-supervised branches and chromaticity regularization self-supervised branches to perform multiple partitions on the original palmprint image, providing data and theoretical basis for subsequent determination of the random mask matrix and visible area. This achieves effective decoupling of illumination interference and material properties, improving the generalization ability and deployment efficiency of the palmprint anti-fraud system.

[0019] In particular, by using raw palmprint images as input without manual labeling, a self-supervised reconstruction architecture is used to achieve physical alignment and fitting of ambient lighting and material properties. This self-supervised approach extracts stable essential features from the image itself, reducing reliance on complex and expensive datasets with genuine and fake labels for data collection. In existing technologies, the recognition of raw palmprints heavily relies on manually labeled genuine and fake tags. This supervised learning paradigm not only faces extremely high human and time costs associated with data collection and labeling, but its performance is also easily limited by the quality and scale of the labeling. Furthermore, such methods often fail to deeply model the underlying physical processes of palmprint image formation (such as the interaction between lighting and surface materials), resulting in insufficient robustness of the learned features to changes in the imaging environment (especially non-uniform lighting) and limited generalization ability. Therefore, this invention considers designing a self-supervised reconstruction architecture to extract inherent patterns from the data itself. It utilizes only the structural and contextual information inherent in the raw palmprint image as a supervisory signal. During the reconstruction of the raw image, it implicitly decomposes and understands the key physical factors affecting the image's appearance, reducing human intervention and improving the generalization ability and deployment efficiency of the palmprint anti-fraud system.

[0020] In particular, this invention enables efficient and accurate anti-fraud prediction in real-world data collection scenarios through model training or fine-tuning, while significantly improving system real-time performance. By completely eliminating the computationally intensive decoder during the inference phase and only requiring the operation of a lightweight encoder and classification head, the overall computational load is significantly reduced and inference latency is significantly shortened. This allows for efficient deployment and real-time response on resource-constrained edge devices (such as palm payment terminals and embedded vision modules) while ensuring high accuracy. In practice, the final determination of anti-fraud methods often relies on complex multi-stage models or architectures containing heavy decoders. Compared to real-world scenarios where high accuracy and real-time performance are difficult to achieve simultaneously, resulting in high computational load, high latency, and inefficient deployment on resource-constrained edge devices, this invention not only avoids the problems of high computational overhead and high latency but also demonstrates stability in real-world scenarios such as cross-media and complex lighting conditions, improving the generalization ability and deployment efficiency of the palmprint anti-fraud system. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating the steps of the palmprint recognition anti-fraud method based on dual-branch self-supervised learning according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the steps for calculating the first loss according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the steps for calculating the second loss according to an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0023] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0024] Please see Figure 1 , Figure 1 This is a schematic diagram illustrating the steps of the palmprint recognition anti-fraud method based on dual-branch self-supervised learning according to an embodiment of the present invention. The palmprint recognition anti-fraud method based on dual-branch self-supervised learning of the present invention includes: Step S1: Construct a dual-branch self-supervised reconstruction architecture, including a shared encoder, a high-frequency decoder, and a chroma decoder. Decouple domain-sensitive features based on two parallel reconstruction branches, including a high-frequency self-supervised branch and a chroma regularization self-supervised branch. Step S2: Input the original palm print image, determine the first image to be processed based on the high-frequency self-supervised branch, and determine the second image to be processed based on the chroma regularization self-supervised branch; Step S3: Divide the first image to be processed and the second image to be processed into non-overlapping patch blocks, and generate random mask matrices with different spatial distributions for each block to determine the visible area. Step S4: Input the patch blocks of each visible region into the shared encoder for feature extraction to capture the global dependency relationship between image patches and output latent variable features; Step S5: Concatenate each latent variable feature with a learnable mask mark, and send them to the high-frequency decoder and the chroma decoder respectively to calculate the first loss and the second loss, determine the total loss, and physically decouple the illumination and material properties. Step S6: Remove the high-frequency decoder and the chroma decoder, retain the shared encoder that has learned the physical decoupling features, perform global average pooling on the output of the shared encoder and then connect it to the linear classification layer to output the authenticity discrimination result. Use a small number of original palm print images with authenticity labels to fine-tune the architecture to complete the convergence of the architecture and perform palm print recognition anti-fraud.

[0025] Specifically, there is no limitation on the method of obtaining the original palm print image. For example, in this embodiment, it can be open source data on the Internet or a palm print image authorized by the user. Any reasonable method will suffice, and will not be elaborated further.

[0026] Specifically, after pre-training, this invention achieves the transformation from self-supervised learning to downstream classification tasks through task redirection. First, the first and second decoders performing the reconstruction task are removed, retaining only the shared encoder that has learned the physical decoupling features. The output of the shared encoder is then subjected to global average pooling and connected to a linear classification layer to output the true / false judgment result. The network was then fine-tuned using original palmprint images with both real and fake labels. Since the encoder had already mastered robust material and texture features through the mask reconstruction task during pre-training, the fine-tuning stage only required a small amount of labeled data and a low learning rate to achieve rapid convergence. The final deployed model only contained a shared encoder and classification head. By completely eliminating the computational overhead of the decoder, the inference speed on edge devices was significantly improved, meeting the requirements for real-time anti-fraud.

[0027] Specifically, this invention designs a lightweight architecture based on dual-branch self-supervised learning. It decouples domain-sensitive features using two parallel reconstruction branches. Existing technologies mostly rely on single feature modalities (such as pure texture analysis or image classification) or complex domain adaptive networks. These methods often fail to deeply integrate with imaging physics mechanisms, systematically separating and suppressing key domain interference factors such as illumination changes. Furthermore, due to the lack of structured modeling of the illumination effect during image formation, existing models struggle to effectively decouple domain-sensitive features in cross-domain scenarios, resulting in limited extraction of essential features and decreased generalization performance. Therefore, this invention considers using a high-frequency self-supervised branch and a chromaticity regularization self-supervised branch to perform multiple partitions of the original palmprint image. This provides data and a theoretical basis for subsequently determining the random mask matrix and visible areas, achieving effective decoupling of illumination interference and material properties, and improving the generalization ability and deployment efficiency of the palmprint anti-fraud system.

[0028] Specifically, the process of determining the first image to be processed includes: The original palmprint image is converted to the frequency domain using a two-dimensional discrete Fourier transform, and a random radius is defined. The random radius is sampled from a preset range to construct a circular high-pass filter; The first image to be processed is obtained by filtering out low-frequency components and retaining high-frequency spatial information at different scales.

[0029] Specifically, the first image to be processed is determined according to formula (1). (1) In formula (1), This represents the first image to be processed. Indicates Fourier transform, This represents the inverse Fourier transform. Indicates a random radius of High-pass filtering operation, Represents a random radius. The preset range is , where x represents the original palm print image.

[0030] Specifically, the process of determining the second image to be processed includes: The original palm print image is determined to be a three-dimensional matrix containing three color channels: red, green, and blue. Randomly select a color channel and set its value to 0 to obtain the second image to be processed.

[0031] Specifically, the chromaticity regularization self-supervised branch simulates the color gamut shift of different acquisition devices and decouples the influence of ambient light. The second image to be processed is determined according to formula (2). (2) In formula (2), This represents the second image to be processed. This indicates element-wise multiplication. This indicates that the mask is discarded in a single channel.

[0032] Specifically, this invention uses raw palmprint images as input without manual labeling. A self-supervised reconstruction architecture is employed to achieve physical alignment and fitting of ambient lighting and material properties. By mining stable essential features from the image itself through self-supervision, it reduces reliance on complex and expensive datasets with genuine and fake labels for data collection. In existing technologies, the recognition of raw palmprints heavily depends on manually labeled genuine and fake tags. This supervised learning paradigm not only faces extremely high human and time costs associated with data collection and labeling, but its performance is also easily limited by the quality and scale of the labeling. Furthermore, such methods often fail to deeply model the underlying physical processes of palmprint image formation (such as the interaction between lighting and surface materials), resulting in insufficient robustness of the learned features to changes in the imaging environment (especially non-uniform lighting) and limited generalization ability. Therefore, this invention considers designing a self-supervised reconstruction architecture to mine inherent patterns from the data itself. It utilizes only the structural and contextual information inherent in the raw palmprint image as a supervisory signal. During the reconstruction of the raw image, it implicitly decomposes and understands the key physical factors affecting the image's appearance, reducing human intervention and improving the generalization ability and deployment efficiency of the palmprint anti-fraud system.

[0033] Specifically, the process of determining the visible area includes: A predetermined proportion of patch blocks in each of the aforementioned random mask matrices are determined to be visible blocks; The set of visible blocks is determined to be a visible region.

[0034] Specifically, the preset ratio is 50%, which is designed to balance the difficulty of reconstruction and semantic extraction: too low a mask ratio will cause the network to reconstruct through simple local pixel interpolation, and will not be able to explore deep material features; while the 50% ratio forces the shared encoder to use the global attention mechanism of Vision Transformer to infer the missing high-frequency texture and low-frequency chromaticity patterns from the spatially discontinuous visible patches.

[0035] Specifically, the Vision Transformer architecture, based on a global attention mechanism, captures the global dependencies between image patches.

[0036] Specifically, the Vision Transformer architecture parameters are set as follows: Linear projection layer: Divides the input image (size 224x224x3) into 16x16 patch blocks, and obtains a total of 196 feature vectors; The shared encoder embedding dimension is set to 192. It contains 12 Transformer coding blocks, each with 3 attention heads. During the self-supervised pre-training phase, a fixed mask ratio of 50% is used for both the high-frequency branch and the chroma branch, so that the encoder can learn semantic features from a very small number of visible patches. The high-frequency decoder and the chroma decoder are set up in the same way, with the embedding dimension set to 512 and the depth set to 8 layers. The decoder does not participate in the final anti-counterfeiting inference work and will not increase the inference overhead.

[0037] Specifically, the high-frequency decoder is responsible for predicting the high-frequency residuals of the reconstructed original image to enhance the model's perception of edge structures; the chromaticity decoder is responsible for reconstructing the color pixels of the original image to guide the model to learn cross-domain stable material chromaticity distribution.

[0038] Specifically, latent variable features include high-frequency texture and low-frequency chromaticity.

[0039] Please see Figure 2 , Figure 2 This is a schematic diagram illustrating the steps for calculating the first loss according to an embodiment of the present invention. Specifically, the process of calculating the first loss includes: Calculate texture-guided loss based on gradient vector cosine similarity; The mean square error loss between the reconstructed pixel values ​​and the original pixel values ​​of the masked region in the first image to be processed is determined as the pixel-level reconstruction loss. The weighted sum of the texture-guided loss and the pixel-level reconstruction loss is determined to be the first loss.

[0040] Specifically, the masked region refers to the set of image patch blocks that are marked as invisible in a random mask matrix.

[0041] Specifically, the steps for loss of mean squared error are as follows: Determine the square of the difference between each reconstructed pixel value and its corresponding original pixel value in the masked region; The mean squared error loss is determined by the ratio of the sum of the squares to the total number of pixels in the masked region.

[0042] Specifically, the first loss function is represented by formula (3). (3) In formula (3), Denotes the first loss function. Indicates texture-guided loss, Indicates pixel-level reconstruction loss. The weighting coefficients represent the texture-guided loss. The weight coefficients represent the pixel-level reconstruction loss.

[0043] Specifically, according to formula (4), the texture-guided loss is expressed as follows: (4) In formula (4), Indicates the use of The spatial gradient vector extracted by the operator, Indicates the target high-frequency residual, Indicates the prediction of high-frequency residuals. This represents the cosine similarity operator between two vectors.

[0044] Please see Figure 3 , Figure 3 This is a schematic diagram illustrating the steps for calculating the second loss according to an embodiment of the present invention. Specifically, the process of calculating the second loss includes: The mean square error loss between the reconstructed pixel values ​​and the original pixel values ​​of the masked region in the second image to be processed is determined as the color pixel-level reconstruction loss. Two sets of log-chromaticity ratio maps are constructed based on the channels retained and the channels discarded in the second image to be processed. Each of the aforementioned mapping maps and the target mapping map is spatially flattened and then ordered in ascending order to obtain the corresponding empirical distribution sequence. The difference between each of the empirical distribution sequences is calculated using distance calculation to obtain the chromaticity distribution loss; The weighted sum of the color pixel reconstruction loss and the chromaticity distribution loss is determined to be the second loss.

[0045] Specifically, the target mapping map refers to the logarithmic chromaticity ratio mapping map calculated from the original ground truth image (i.e., the original palm print image without channel discarding).

[0046] Specifically, the second loss function is represented by formula (5). (5) In formula (5), This represents the second loss function. Indicates the color pixel reconstruction loss. Indicates the loss of chromaticity distribution. This represents the second weighting coefficient preset for the color pixel reconstruction loss. This represents the second weighting coefficient preset for the chromaticity distribution loss.

[0047] Specifically, the process of constructing the chromaticity distribution loss is as follows: Let the number of channels retained in the second image to be processed be... The discarded channel is Based on the imaging physics model, two sets of logarithmic chromaticity ratio mappings are constructed. ; To enable the model to learn the statistical distribution of materials rather than absolute pixel values, this invention modifies the predicted mapping map. Mapping diagram with target Perform spatial flattening and then sort in ascending order. Obtain the corresponding empirical distribution sequence ; The difference between sorted sequences is calculated using distance to obtain the chromaticity distribution loss. The distance is expressed by formula (6). (6) Specifically, by performing an ascending sort operation and constructing a chromaticity distribution loss, the aim is to guide the model from "absolute pixel comparison" to learning "material intrinsic distribution characteristics". Since single-point pixel comparison values ​​are easily affected by sensor noise and fluctuations in the acquisition environment, by using the sorted empirical distribution sequence as a supervision signal, the statistical regularity of the material in chromaticity distribution is effectively extracted. This not only suppresses random noise interference caused by hardware differences, but also forces the shared encoder to capture the deep discrimination regularity of genuine and counterfeit materials at the physical property level, thereby achieving robust anti-counterfeiting discrimination in cross-media scenarios.

[0048] Specifically, the process of determining the total loss includes: The weighted sum of the first loss and the second loss is determined to be the total loss.

[0049] Specifically, the total loss function is represented by formula (7). (7) In formula (7), Represents the total loss function. This represents the weighting coefficients of the first loss function. This represents the weighting coefficients of the second loss function.

[0050] Specifically, the process of physically decoupling lighting and material properties includes: Calculate the logarithm of the channel ratio to suppress ambient light interference and achieve separation of lighting and material properties; In the channel ratio calculation, the ambient light component is canceled out, and only the relative reflectivity information reflecting the material properties is retained.

[0051] Specifically, there are no restrictions on the specific method for calculating the channel ratio. In practice, it can be calculated based on Retinex theory. According to classical Retinex theory, the pixel value at any channel of an image can be represented as the ambient light intensity. Surface reflectivity The product of, i.e. This simplified model reveals that the illumination component acts on the object's reflectivity as a multiplicative factor during the imaging process.

[0052] It is understandable that, at the same location in the same image, the ambient light intensity can be approximated. The multiplicative effect of each color channel exhibits local spatial consistency, while the reflectance characteristics... The ratio varies depending on the channel, and is calculated according to formula (8). (8) In formula (8), This represents the pixel response values ​​of different color channels at the same location in space in the original palm print image; Indicates the palm print on the surface of the object in the channel Reflectance component at a specific wavelength; This represents a logarithmic operation, converting the multiplicative imaging model into an additive model; Among them, the ambient light intensity is the same in both the numerator and denominator. The illumination was removed, thus eliminating the influence of light on image features and retaining only the relative reflectivity information that reflects material properties.

[0053] Specifically, this invention enables efficient and accurate anti-fraud prediction in real-world data acquisition scenarios through model training or fine-tuning, while significantly improving system real-time performance. By completely eliminating the computationally intensive decoder during the inference phase and only requiring the operation of a lightweight encoder and classification head, the overall computational load is drastically reduced and inference latency is significantly shortened. This ensures high accuracy while achieving efficient deployment and real-time response on resource-constrained edge devices (such as palm print payment terminals and embedded vision modules). In practice, the final determination of anti-fraud methods often relies on complex multi-stage models or architectures containing heavy decoders. Compared to real-world scenarios where high accuracy and real-time performance are difficult to achieve simultaneously, resulting in high computational load, high latency, and inefficient deployment on resource-constrained edge devices, this invention not only avoids the problems of high computational overhead and high latency but also demonstrates stability in real-world scenarios such as cross-media and complex lighting conditions, improving the generalization ability and deployment efficiency of the palmprint anti-fraud system.

[0054] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0055] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A palmprint recognition anti-fraud method based on dual-branch self-supervised learning, characterized in that, include: A dual-branch self-supervised reconstruction architecture is constructed, including a shared encoder, a high-frequency decoder, and a chroma decoder. Domain-sensitive features are decoupled based on two parallel reconstruction branches, including a high-frequency self-supervised branch and a chroma regularization self-supervised branch. Input the original palm print image, determine the first image to be processed based on the high-frequency self-supervised branch, and determine the second image to be processed based on the chroma regularization self-supervised branch; The first image to be processed and the second image to be processed are divided into non-overlapping patch blocks, and each patch block is independently generated with a random mask matrix with a different spatial distribution to determine the visible area. The patch blocks of each visible region are input into the shared encoder for feature extraction to capture the global dependencies between image patches and output latent variable features; Each latent variable feature is concatenated with a learnable mask label and fed into the high-frequency decoder and the chroma decoder respectively to calculate the first loss and the second loss, determine the total loss, and physically decouple the illumination and material properties. The high-frequency decoder and the chroma decoder are removed, and the shared encoder that has learned the physical decoupling features is retained. The output is global average pooled at the output end of the shared encoder and then connected to a linear classification layer to output the authenticity judgment result. A small number of original palm print images with authenticity labels are used to fine-tune the architecture to complete the convergence of the architecture and perform palm print recognition anti-fraud.

2. The palmprint recognition anti-fraud method based on dual-branch self-supervised learning according to claim 1, characterized in that, The process of determining the first image to be processed includes: The original palmprint image is converted to the frequency domain using a two-dimensional discrete Fourier transform, and a random radius is defined. The random radius is sampled from a preset range to construct a circular high-pass filter; The first image to be processed is obtained by filtering out low-frequency components and retaining high-frequency spatial information at different scales.

3. The palmprint recognition anti-fraud method based on dual-branch self-supervised learning according to claim 1, characterized in that, The process of determining the second image to be processed includes: The original palm print image is determined to be a three-dimensional matrix containing three color channels: red, green, and blue. Randomly select a color channel and set its value to 0 to obtain the second image to be processed.

4. The palmprint recognition anti-fraud method based on dual-branch self-supervised learning according to claim 1, characterized in that, The process of determining the visible area includes: A predetermined proportion of patch blocks in each of the aforementioned random mask matrices are determined to be visible blocks; The set of visible blocks is determined to be a visible region.

5. The palmprint recognition anti-fraud method based on dual-branch self-supervised learning according to claim 1, characterized in that, The Vision Transformer architecture, based on a global attention mechanism, captures global dependencies between image patches.

6. The palmprint recognition anti-fraud method based on dual-branch self-supervised learning according to claim 1, characterized in that, The latent variable features include high-frequency texture and low-frequency chromaticity.

7. The palmprint recognition anti-fraud method based on dual-branch self-supervised learning according to claim 1, characterized in that, The process of calculating the first loss includes: Calculate texture-guided loss based on gradient vector cosine similarity; The mean square error loss between the reconstructed pixel values ​​and the original pixel values ​​of the masked region in the first image to be processed is determined as the pixel-level reconstruction loss. The weighted sum of the texture-guided loss and the pixel-level reconstruction loss is determined to be the first loss.

8. The palmprint recognition anti-fraud method based on dual-branch self-supervised learning according to claim 1, characterized in that, The process of calculating the second loss includes: The mean square error loss between the reconstructed pixel values ​​and the original pixel values ​​of the masked region in the second image to be processed is determined as the color pixel-level reconstruction loss. Two sets of log-chromaticity ratio maps are constructed based on the channels retained and the channels discarded in the second image to be processed. Each of the aforementioned mapping maps and the target mapping map is spatially flattened and then ordered in ascending order to obtain the corresponding empirical distribution sequence. The difference between each of the empirical distribution sequences is calculated using distance calculation to obtain the chromaticity distribution loss; The weighted sum of the color pixel reconstruction loss and the chromaticity distribution loss is determined to be the second loss.

9. The palmprint recognition anti-fraud method based on dual-branch self-supervised learning according to claim 1, characterized in that, The process of determining the total loss includes: The weighted sum of the first loss and the second loss is determined to be the total loss.

10. The palmprint recognition anti-fraud method based on dual-branch self-supervised learning according to claim 1, characterized in that, The process of physically decoupling lighting and material properties includes: Calculate the logarithm of the channel ratio to suppress ambient light interference and achieve separation of lighting and material properties; In the channel ratio calculation, the ambient light component is canceled out, and only the relative reflectivity information reflecting the material properties is retained.

Images