Certificate authentication method and device, equipment, medium and program product

Abstract

The application provides a certificate authentication method and device, equipment, medium and program product, which can be applied to the fields of artificial intelligence and financial technology. The method comprises the following steps: collecting a multi-band image of a target certificate, and preprocessing the multi-band image to obtain a multispectral image; processing the multispectral image based on a pre-trained spectral texture attention fusion model to obtain spectral texture fusion features; performing physical attribute inference on the multispectral image through a pre-trained physical information neural network model to obtain physical attribute features; constructing a counterfeit prototype based on a pre-trained meta-learning task model to obtain counterfeit prototype features; and performing identification processing on the spectral texture fusion features, the physical attribute features and the counterfeit prototype features through a pre-trained multi-modal authentication large model to obtain an identification result of the target certificate.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and fintech, and more specifically to a method, apparatus, device, medium, and program product for authenticating credentials. Background Technology

[0002] As important carriers of the financial system, bank drafts, certificates, and cash are directly related to the safety of funds and market order. Therefore, the authentication of these documents is of great significance in preventing financial risks. Strengthening the ability to identify counterfeit documents can build a solid financial security defense and ensure the compliant conduct of transactions.

[0003] At present, multispectral imaging has become a commonly used anti-counterfeiting technology. It achieves authenticity determination through multi-band imaging and feature analysis. Existing anti-counterfeiting schemes mostly use simple methods such as channel stitching and decision layer fusion for feature processing. However, the granularity of multispectral feature fusion is too large, and simple fusion cannot fully explore the complementary information across the spectrum. At the same time, multi-band image fusion cannot take into account both fidelity and detailed information. Existing deep learning models cannot effectively perceive the differences in spectral physical properties, resulting in insufficient accuracy of anti-counterfeiting identification. Summary of the Invention

[0004] In view of the above problems, embodiments of this application provide a method, apparatus, device, medium, and program product for credential authentication.

[0005] According to a first aspect of this application, a method for authenticating a credential is provided, comprising: acquiring a multi-band image of a target credential and preprocessing the multi-band image to obtain a multispectral image; processing the multispectral image based on a pre-trained spectral texture attention fusion model to obtain spectral texture fusion features; performing physical attribute inference on the multispectral image using a pre-trained physical information neural network model to obtain physical attribute features; constructing a counterfeit prototype on the multispectral image based on a pre-trained meta-learning task model to obtain counterfeit prototype features; and performing authentication processing on the spectral texture fusion features, the physical attribute features, and the counterfeit prototype features using a pre-trained multimodal authentication model to obtain the authentication result of the target credential.

[0006] According to an embodiment of this application, the meta-learning task model includes a decoupling layer, a feature extraction layer, and a prototype calculation layer. The pre-trained meta-learning task model constructs a prototype from the multispectral image to obtain prototype features, including: performing convolution and gradient inversion on the multispectral image based on the decoupling layer to obtain decoupling features; performing partial convolution residual processing on the decoupling features based on the feature extraction layer to obtain microscopic physical features; adding the decoupling features and the microscopic physical features using a skip connection method to obtain microscopic fusion features; and calculating the prototype features corresponding to the microscopic fusion features using an embedding function based on the prototype calculation layer.

[0007] According to an embodiment of this application, the training process of the meta-learning task model includes: constructing a task pool based on a learning task of the imitation type, wherein the learning task includes a support set and a query set; sampling a target task based on the task pool, sampling samples from the support set of the target task, and calculating a loss function value based on the initialization parameters of the meta-learning task model and the sampled support set, performing an inner loop gradient update based on the loss function value to obtain the inner loop update parameters of the target task; calculating a meta-loss based on the query set of the target task and the inner loop update parameters, and performing an outer loop backpropagation update based on the meta-loss to update the initialization parameters; dynamically adjusting the weights of the global parameters updated by the outer loop according to the task characteristics to obtain task-adaptive weighted parameters; and responding to updating the learning task of the imitation type, obtaining updated samples based on the updated learning task, and performing model parameter gradient updates based on the updated samples and the task-adaptive weighted parameters.

[0008] According to an embodiment of this application, the spectral texture attention fusion model includes a spectral unmixing network, a spatial texture feature extraction network, a spectral feature extraction network, and an attention fusion network. The process of processing the multispectral image based on the pre-trained spectral texture attention fusion model to obtain spectral texture fusion features includes: extracting spectral and spatial information of the multispectral image through the spectral unmixing network; concatenating the spectral information of the multispectral image to obtain spectral features; concatenating the spatial information of the multispectral image to obtain spatial features; extracting high- and low-frequency features from the spatial features based on the spatial texture feature extraction network to obtain enhanced texture features; performing multi-scale feature fusion on the spectral features based on the spectral feature extraction network to obtain enhanced spectral features; and fusing the enhanced texture features and the enhanced spectral features based on the attention fusion network to generate the spectral texture fusion features.

[0009] According to an embodiment of this application, the step of extracting high and low frequency features from the spatial features based on the spatial texture feature extraction network to obtain enhanced texture features includes: performing wavelet transform on the spatial features to obtain low-frequency components and high-frequency components; performing Fourier transform on the low-frequency components, performing convolution and double pooling on the transformed low-frequency components to obtain low-frequency fusion components, and performing inverse Fourier transform on the low-frequency fusion components to obtain enhanced low-frequency features; concatenating the high-frequency components, performing convolution adjustment on the concatenated high-frequency components, and adding the adjusted high-frequency components to the spatial features to obtain spatial high-frequency features; performing jump fusion on the spatial high-frequency features through a U-shaped network to obtain high-frequency fusion features, and performing wavelet transform on the high-frequency fusion features to obtain enhanced high-frequency features; and performing inverse wavelet transform on the enhanced high-frequency features and the enhanced low-frequency features to obtain the enhanced texture features.

[0010] According to an embodiment of this application, the step of performing multi-scale feature fusion on the spectral features based on the spectral feature extraction network to obtain enhanced spectral features includes: performing convolution processing on the spectral features to extract local spectral features; performing global average pooling processing on the local spectral features to obtain a global context vector; performing difference weight activation on the global context vector and the local spectral features based on weight parameters to obtain activated features; segmenting the activated features based on channel dimensions to obtain multiple sets of spectral features; performing multi-scale feature mapping processing on the multiple sets of spectral features according to preset multiple strides to obtain multi-scale features; concatenating and fusing the multi-scale features to obtain multi-scale fused features; performing convolution and activation processing on the multi-scale fused features to obtain a first activated feature; performing convolution and activation processing on the activated features to obtain a second activated feature; multiplying the first activated feature and the second activated feature pixel by pixel to obtain channel attention features, and performing convolution processing on the channel attention features to obtain the enhanced spectral features.

[0011] According to an embodiment of this application, the step of fusing the enhanced texture features and the enhanced spectral features based on the attention fusion network to generate the spectral texture fusion feature includes: performing self-attention extraction on the enhanced texture features to obtain texture self-attention features; performing self-attention extraction on the enhanced spectral features to obtain spectral self-attention features; performing cross-attention extraction on the enhanced spectral features and the enhanced texture features to obtain cross-attention features; concatenating the cross-attention features, the texture self-attention features, and the spectral self-attention features, and performing multilayer perception and convolution processing on the concatenated attention features to obtain enhanced attention features; and fusing the enhanced attention features, the enhanced spectral features, and the enhanced texture features to generate the spectral texture fusion feature.

[0012] According to an embodiment of this application, the physical information neural network model includes a convolutional layer, a residual module, and a multi-task dense prediction module; the step of performing physical attribute inference on the multispectral image using the pre-trained physical information neural network model to obtain physical attribute features includes: extracting features from the multispectral image based on the convolutional layer to obtain shallow features; processing the shallow features through the residual module to obtain deep features; and performing physical attribute prediction based on the deep features through the multi-task dense prediction module to obtain the physical attribute features.

[0013] According to an embodiment of this application, the loss function of the physical information neural network model includes data loss of physical attributes and physical consistency loss; wherein, the physical attributes include paper optical attributes, ink optical attributes, printing pressure characteristics, and fluorescence attenuation characteristics; the physical consistency loss includes energy conservation constraints, wavelength constraints, and optical transmission theory constraints.

[0014] According to an embodiment of this application, the multimodal anti-counterfeiting model includes a sliding window transformation network, a physical network, a convolutional network, a cross-modal interaction layer, and a multi-task prediction layer. The multimodal anti-counterfeiting model, trained in advance, performs authentication processing on the spectral texture fusion features, the physical attribute features, and the counterfeit prototype features to obtain the authentication result of the target credential. This includes: inputting the spectral texture fusion features into the sliding window transformation network to extract visual features; inputting the physical attribute features into the physical network to extract physical features; wherein the physical network includes a multilayer perceptron, a residual connection structure, and a one-dimensional convolutional unit; inputting the counterfeit prototype features into the convolutional network to extract counterfeit category features; fusing the visual features, physical features, and counterfeit category features through the cross-modal interaction layer based on a cross-attention mechanism to generate joint features; and performing authenticity prediction on the joint features through the multi-task prediction layer to obtain the authentication result of the target credential.

[0015] According to an embodiment of this application, the preprocessing of the multi-band image includes: image registration of the multi-band image; performing illumination unevenness correction on the registered multi-band image based on the band information of the multi-band image to obtain a multi-band corrected image; inputting the multi-band corrected image into a pre-trained image reconstruction model for the corresponding band for super-resolution reconstruction to obtain a multi-band reconstructed image; and performing image enhancement on the multi-band reconstructed image to obtain the multispectral image.

[0016] According to an embodiment of this application, the image reconstruction model includes a shallow feature extraction layer, a deep feature extraction layer, and a reconstruction layer. The step of inputting the multi-band corrected image into a pre-trained image reconstruction model for the corresponding band for super-resolution reconstruction includes: extracting shallow features from the corrected image of the current band based on the shallow feature extraction layer of the current band to obtain target shallow features; performing differential modulation and residual fusion on the target shallow features based on the deep feature extraction layer of the current band to generate target deep features; wherein the deep feature extraction layer includes a local feature enhancement network, a differential feature modulation network, a multi-scale feature aggregation network, a channel attention enhancement network, a residual feature extraction network, and a residual fusion network; and performing convolutional reconstruction on the target deep features through the reconstruction layer of the current band to obtain the reconstructed image of the current band.

[0017] According to an embodiment of this application, the method further includes: calling a real-time authentication interface to obtain the authentication result of the target credential, and generating an abnormal area heat map and physical attribute report based on the authentication result; and calculating the transaction risk level based on the authentication result and transaction information, and performing a risk warning operation if the transaction risk level is higher than a preset risk threshold.

[0018] According to an embodiment of this application, the method further includes: in response to a user modifying the authentication result of the target credential, obtaining the corrected authentication result and storing it in a false alarm database; and based on the false alarm database and a timed training strategy, using continuous learning technology to incrementally train the credential authentication model; wherein the credential authentication model includes one or more of the following: the spectral texture attention fusion model, the physical information neural network model, the meta-learning task model, and the multimodal authentication big model.

[0019] According to a second aspect of this application, a certificate authentication device is provided, comprising: an image acquisition and preprocessing module for acquiring multi-band images of a target certificate and preprocessing the multi-band images to obtain a multispectral image; an attention fusion module for processing the multispectral image based on a pre-trained spectral texture attention fusion model to obtain spectral texture fusion features; a physical attribute reasoning module for performing physical attribute reasoning on the multispectral image using a pre-trained physical information neural network model to obtain physical attribute features; a prototype construction module for constructing a counterfeit prototype on the multispectral image based on a pre-trained meta-learning task model to obtain counterfeit prototype features; and a multimodal authentication module for performing authentication processing on the spectral texture fusion features, the physical attribute features, and the counterfeit prototype features using a pre-trained multimodal authentication large model to obtain the authentication result of the target certificate.

[0020] According to a third aspect of this application, an electronic device is provided, comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0021] According to a fourth aspect of this application, a computer-readable storage medium is also provided, on which a computer program or instructions are stored, wherein the computer program or instructions, when executed by a processor, implement the steps of the above-described method.

[0022] According to a fifth aspect of this application, a computer program product is also provided, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0023] In the embodiments of this application, high-quality basic data is provided for subsequent spectral image analysis through multi-band image acquisition and preprocessing. A spectral texture attention fusion model is used to finely fuse multispectral information, deeply mine cross-band complementary features, and extract spectral texture fusion features, balancing image fidelity and detail. A physical information neural network combines spectral physical characteristics to perform inference, extract physical attribute features, and accurately capture the physical attributes of the voucher. A counterfeit prototype is constructed based on a meta-learning task model, adapting to various counterfeiting methods. A multimodal counterfeit detection model integrates multiple feature categories for comprehensive discrimination, significantly improving the accuracy, stability, and generalization ability of voucher counterfeiting detection, effectively preventing various counterfeiting risks, and ensuring the security of financial transactions. Attached Figure Description

[0024] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0025] Figure 1 The illustrations depict application scenarios of the credential authentication method, apparatus, device, medium, and program products according to embodiments of this application.

[0026] Figure 2 A flowchart illustrating a credential authentication method according to an embodiment of this application is shown in the schematic diagram.

[0027] Figure 3 The schematic diagram illustrates the structure of a multispectral image fusion intelligent anti-counterfeiting system according to an embodiment of this application;

[0028] Figure 4 This illustration schematically shows a flowchart of the spectral texture attention fusion method for credential authentication according to an embodiment of this application;

[0029] Figure 5A This illustration schematically shows a structural diagram of an image reconstruction model for a credential authentication method according to an embodiment of this application;

[0030] Figure 5B The diagram illustrates the network connection of an image reconstruction model for a credential authentication method according to an embodiment of this application.

[0031] Figure 6 This illustration schematically shows a flowchart of the spectral texture attention fusion method for credential authentication according to an embodiment of this application;

[0032] Figure 7A This diagram schematically illustrates the structure of a spectral texture attention fusion model for a credential authentication method according to an embodiment of this application.

[0033] Figure 7B The diagram illustrates the network connection of the spectral texture attention fusion model of the credential authentication method according to an embodiment of this application.

[0034] Figure 8A This illustration schematically shows the structure of a physical information neural network model of a credential authentication method according to an embodiment of this application;

[0035] Figure 8B This illustration schematically shows the network connection diagram of the physical information neural network model of the credential authentication method according to an embodiment of this application;

[0036] Figure 9A This illustration schematically shows the structure of the meta-learning task model of the credential authentication method according to an embodiment of this application;

[0037] Figure 9B The diagram illustrates the network connection of the meta-learning task model of the credential authentication method according to an embodiment of this application.

[0038] Figure 10A This schematically illustrates the structure of a multimodal authentication model for a credential authentication method according to an embodiment of this application.

[0039] Figure 10B The diagram illustrates the network connection of a multimodal authentication model for a credential authentication method according to an embodiment of this application.

[0040] Figure 11 This schematic diagram illustrates the structural block diagram of a credential authentication device according to an embodiment of this application;

[0041] Figure 12 A block diagram schematically illustrates an electronic device suitable for implementing a credential authentication method according to an embodiment of this application. Detailed Implementation

[0042] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0043] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0044] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0045] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0046] As used in this paper, the term "model" refers to a model that learns the relationship between inputs and outputs from training data, enabling it to generate corresponding outputs for a given input after training. Model generation can be based on machine learning techniques. Deep learning is a machine learning algorithm that processes inputs and provides corresponding outputs using multiple layers of processing units. A neural network model is an example of a deep learning-based model. In this paper, "model" may also be referred to as a "machine learning model," "learning model," "machine learning network," or "learning network," and these terms are used interchangeably.

[0047] It's important to note that the term "neural network" can refer to a machine learning network based on deep learning. A neural network processes input and provides corresponding output, typically consisting of an input layer, an output layer, and one or more hidden layers between them. Neural networks used in deep learning applications often include many hidden layers, increasing the network's depth. The layers of a neural network are connected sequentially, so that the output of the previous layer serves as the input to the next layer. The input layer receives the input to the neural network, while the output layer's output becomes the final output. Each layer of a neural network includes one or more nodes (also called processing nodes or neurons), each processing the input from the layer above.

[0048] It should be understood that machine learning generally includes three phases: training, testing, and application (also known as inference). In the training phase, a given model is trained using a large amount of training data, iteratively updating parameter values ​​until the model can consistently generate inferences that meet the expected goals from the training data. Through training, the model can be considered to have learned the relationship between inputs and outputs (also known as the input-output mapping) from the training data. The parameter values ​​of the trained model are determined. In the testing phase, test inputs are applied to the trained model to test whether it can provide the correct output, thus determining the model's performance. In the application phase, the model can be used to process actual inputs based on the trained parameter values ​​to determine the corresponding output.

[0049] In one or more embodiments described herein, the term "large model" can refer to a deep learning model with a large number of model parameters, which can include hundreds of millions, tens of billions, hundreds of billions, trillions, or even more than ten trillion model parameters.

[0050] In the field of authentication of bank drafts, certificates, and cash, the pressure of authentication has increased significantly, leading to the widespread application of multispectral imaging technology. Current technologies primarily involve acquiring spectral images across multiple bands and extracting anti-counterfeiting features for authenticity verification. Currently, multispectral authentication algorithms are mainly divided into two categories: one is a matching method based on manual features, which extracts fluorescence regions and infrared absorption features in specific bands and compares them with standard templates to identify authenticity; the other is a classification method based on deep learning, which uses multi-band images as multi-channel input to train a convolutional neural network (CNN) model for authenticity verification.

[0051] However, the above two technical solutions have obvious drawbacks, specifically:

[0052] The granularity of multispectral feature fusion is too large: existing methods are mostly simple channel splicing or decision-level fusion, lacking fine interaction modeling of spectral and spatial dimensions. Furthermore, there are complex nonlinear correlations between physical anti-counterfeiting features in different bands, and simple fusion is insufficient to fully explore the complementary information across the spectrum.

[0053] Lack of embedded physical prior knowledge: Existing deep learning models mainly learn the spectrum-appearance mapping relationship from data, failing to fully utilize prior knowledge such as physical characteristics and optical properties. The reflection / transmission characteristics of paper at different wavelengths obey physical optical laws (such as Lambert reflection and Kubelka-Munk theory). There are fundamental differences between genuine and counterfeit banknotes at the physical property level, and purely data-driven models struggle to capture these subtle differences at the physical level, limiting their generalization ability to new counterfeiting methods.

[0054] The model has a weak ability to adapt to new imitation methods: After the existing model is trained, the parameters are fixed. When faced with new imitation technologies (such as the use of special materials or new printing processes), a large number of new samples need to be collected for retraining, which is costly and slow to respond.

[0055] The balance between fidelity and detail in multi-band image fusion is insufficient: Multispectral image fusion technology can fuse image information of the same target acquired in multiple bands such as visible light, infrared, and ultraviolet at the pixel level, feature level, or decision level to generate enhanced images containing richer spectral information. However, there is an inherent trade-off between single infrared and visible light fusion. Grayscale fusion retains structural details but loses spectral fidelity, while color fusion maintains spectral authenticity but sacrifices detail consistency. For document authentication, it is necessary to retain printing details (texture, dot structure) and maintain the authenticity of spectral anti-counterfeiting features, which existing methods struggle to achieve simultaneously.

[0056] Limited cross-scenario generalization ability: Multi-band images of different ticket types, versions, and printing batches have domain differences, but existing methods lack effective modeling of domain-invariant features, resulting in decreased model performance and low accuracy in counterfeit detection in new scenarios.

[0057] This application provides a method for authenticating a document, comprising: acquiring multi-band images of the target document and preprocessing the multi-band images to obtain multispectral images; processing the multispectral images based on a pre-trained spectral texture attention fusion model to obtain spectral texture fusion features; performing physical attribute inference on the multispectral images using a pre-trained physical information neural network model to obtain physical attribute features; constructing a counterfeit prototype on the multispectral images based on a pre-trained meta-learning task model to obtain counterfeit prototype features; and performing identification processing on the spectral texture fusion features, physical attribute features, and counterfeit prototype features using a pre-trained multimodal authentication model to obtain the authentication result of the target document. In this application's embodiments, multi-band image acquisition and preprocessing provide high-quality basic data for subsequent spectral image analysis. The spectral texture attention fusion model is used to finely fuse multispectral information, deeply mine cross-band complementary features, and extract spectral texture fusion features, balancing image fidelity and detail information. The physical information neural network combines spectral physical characteristics to perform inference, extract physical attribute features, and accurately capture the physical attributes of the document. A counterfeit prototype is constructed based on the meta-learning task model, adapting to various counterfeiting methods. By integrating multiple features into a multimodal authentication model, the accuracy, stability, and generalization ability of authentication credentials are significantly improved, effectively preventing various counterfeiting risks and ensuring the security of financial transactions.

[0058] The document authentication method and apparatus of this application can be used in the fields of artificial intelligence and fintech, as well as in any other field. The application fields of the document authentication method and apparatus of this application are not limited.

[0059] In the technical solution of this application, the user information (including but not limited to user personal information, user credential information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0060] In scenarios where personal information is used for automated decision-making, the methods, devices, and systems provided in this application all provide users with corresponding operation entry points for users to choose to agree to or reject the automated decision results; if the user chooses to reject, the process enters the expert decision-making process.

[0061] Figure 1 The illustrations depict application scenarios of the credential authentication method, apparatus, device, medium, and program products according to embodiments of this application.

[0062] like Figure 1 As shown, application scenario 100 according to an embodiment of this application may include a first terminal device 101, a second terminal device 102, a database 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the database 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables. For example, a user can use the first terminal device 101 and the second terminal device 102 to interact with the server 105 and / or the database 103 through the network 104 to receive or send information, etc.

[0063] The first terminal device 101 and the second terminal device 102 can be electronic devices such as smartphones, wearable devices, personal computers, intelligent voice interaction devices, smart home appliances, intelligent vehicles, in-vehicle terminals, aircraft, unmanned vending terminals, and extended reality devices. Extended reality devices can include virtual reality devices, augmented reality devices, and mixed reality devices. A client application for the target application can be installed and run on the terminal device. This target application can include, but is not limited to, financial transaction applications, payment applications, shopping applications, web browser applications, search applications, instant messaging tools, email clients, and social media platform software (these are just examples). Furthermore, this application embodiment does not limit the form of the target application, including but not limited to applications, mini-programs, etc., installed on the terminal device, and can also be in web page form.

[0064] Server 105 can be a server providing various services, such as a backend management server supporting websites browsed by users using the first terminal device 101 and the second terminal device 102 (this is just an example). The backend management server can analyze and process received user requests and other data, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services such as cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks, and big data. The server can be a backend server for the aforementioned target application, used to provide backend services to the clients of the target application.

[0065] Database 103 is a professional storage system for storing and managing data. It can store various types of data related to the target application, such as user account information, business transaction records, and content resource data. It supports structured, semi-structured, or unstructured data storage and has management capabilities such as adding, deleting, modifying, querying, backing up, and restoring data. In this application scenario, database 103 can be connected to server 105 via a communication link. Server 104 can retrieve the required data from database 103 for processing based on requests from the first terminal device 101 and the second terminal device 102. It can also synchronously store new data generated by the operations of the first terminal device 101 and the second terminal device 102 into database 103, thereby achieving data persistence and efficient retrieval.

[0066] It should be noted that the credential authentication method provided in this application embodiment can generally be executed by server 105 and / or terminal devices 101-102. Accordingly, the credential authentication device provided in this application embodiment can generally be set in server 105 and / or terminal devices 101-102.

[0067] It should be understood that Figure 1 The number of terminal devices, networks, databases, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, databases, and servers can be included.

[0068] Figure 2 A flowchart illustrating a credential authentication method according to an embodiment of this application is shown. Figure 2 As shown, the credential authentication method 200 according to the embodiments of this application may include steps S210 to S250.

[0069] In step S210, multi-band images of the target document are acquired and preprocessed to obtain a multispectral image.

[0070] The target document refers to various paper or plastic documents such as bills, certificates, receipts, and cash that possess anti-counterfeiting attributes and require verification of authenticity. Multi-band images are single-band original images of the document obtained by photographing it using different wavelength light sources, including images in a single spectral dimension such as ultraviolet, infrared, visible light, side light, and transmitted light.

[0071] By aligning a multispectral imaging device with the target document, reflectance images in different bands such as visible light, infrared, and ultraviolet are simultaneously acquired. For example, multispectral imaging devices are deployed at bank counters, self-service ticket acceptance terminals, document verification terminals, cash processing terminals, and cash sorting equipment to acquire multi-band images corresponding to the target document. These multi-band images include visible light images, infrared transmission / reflection images, ultraviolet fluorescence images, side-light images, and transmitted light images. Visible light images represent reflectance images under white light illumination, used to extract macroscopic textures, printing quality, color distribution, and other characteristics. Infrared transmission / reflection images represent images illuminated by 850nm-940nm infrared light, used to detect infrared-absorbing inks, paper thickness uniformity, and watermark layers. Ultraviolet fluorescence images represent images illuminated by 365nm ultraviolet light, acquiring the luminescence characteristics of fluorescent fibers and fluorescent inks. Side-light images represent images illuminated at a 45-degree angle, enhancing the shadow information of printing pressure marks and intaglio raised areas. Transmitted light images represent images illuminated by backlight, used to detect watermarks, security thread embedding depth, paper uniformity, etc.

[0072] The original images of the band have problems such as uneven illumination, imaging shift, and low resolution. Therefore, it is necessary to extract the spectral image corresponding to the original image of the band through preprocessing. Preprocessing includes image registration, uneven illumination correction, super-resolution reconstruction, and image enhancement. Preprocessing can eliminate the imaging position deviation of the band, correct the uneven illumination problem of the image, improve the clarity of image details, and enhance the anti-counterfeiting spectral characteristics of ink and paper.

[0073] The system acquires multi-band images of different wavelengths, including ultraviolet, visible, and infrared, for the same target credential. After preprocessing, the spatial pixels of each band image are precisely aligned, and the band channel data is stored independently. Each band image corresponds to its respective spectral image, and the information from each band is not merged. Spectral images of any or all bands can be retrieved for credential authentication. Furthermore, the system can structure multispectral image data to form a structured data package containing multispectral images, device status data, and acquisition timestamps for subsequent image retrieval and processing.

[0074] In step S220, the multispectral image is processed based on the pre-trained spectral texture attention fusion model to obtain spectral texture fusion features.

[0075] A pre-trained spectral-texture attention fusion model is used to process multispectral images and output spectral-texture fusion features. This model adaptively fuses spectral and texture information across different bands through a cross-spectral attention mechanism, and performs multi-scale processing on both types of information using two different methods. The model training phase relies on a massive amount of diverse and multispectral samples for pre-training, solving for the weight parameters of the fusion module. The model introduces a cross-spectral attention mechanism, enabling adaptive fusion of spectral and texture information across different bands, and performs multi-scale feature extraction on both types of information through a dual-path structure, achieving refined fusion of multispectral features.

[0076] The spectral texture attention fusion model does not independently complete the identification and judgment. The output fusion feature spectral texture fusion feature will be input into the multimodal anti-counterfeiting model for collaborative operation. The spectral texture attention fusion model can be pre-trained using multispectral samples first, and then the spectral texture attention fusion model and the multimodal anti-counterfeiting model can be jointly optimized and trained.

[0077] The spectral texture attention fusion model employs a spectral-spatial unmixing approach to extract features from both types of data, then performs cross-domain feature fusion to achieve joint learning of spatial and spectral correlations. The model uses a weighted sum of reconstruction loss, spatial gradient loss, and spectral gradient loss as its overall loss function. Subsequent joint training phases incorporate these losses into the overall loss function for fine-tuning of parameters. Specifically, the reconstruction loss is composed of the norm loss between the multispectral original image and the generated image, and the perceptual loss. The spatial gradient loss is the sum of the squared differences in the horizontal and vertical gradients of the two images across all bands, divided by N, where N is the total number of bands. The spectral gradient loss is the sum of the squared differences in the spectral gradients of the two images across all bands, divided by N. The gradient is calculated by taking the pixel difference between adjacent coordinate points in the horizontal, vertical, or spectral dimensions and the current coordinate point.

[0078] In step S230, physical attribute inference is performed on the multispectral image using a pre-trained physical information neural network model to obtain physical attribute features.

[0079] By simulating the physical reflection properties of paper and ink under multispectral conditions using a neurophysics engine, prior physical knowledge (such as light reflection / transmission models, ink absorption characteristics, and paper scattering characteristics) is embedded into a Physics-Informed Neural Networks (PINNs) model, enabling the model to simulate and infer physical optical processes. Based on the pre-trained PINN model, physical property inference is performed on multispectral images to obtain physical property features.

[0080] During inference, the model strictly adheres to the physical optical laws embedded during training (such as the conservation of spectral reflectance, the law of ink light absorption, and the physical laws of paper diffuse reflection). It interprets the input multispectral image at a physical level, rather than simply extracting visual features. The final output physical property features are not traditional image pixel features, but directly correspond to the core physical optical parameters of paper and ink under multispectral conditions, including spectral reflectance, ink absorptivity, paper scattering coefficient, refractive index, and optical attenuation coefficient under multispectral conditions. The entire inference process is constrained by physical laws, ensuring that the output features fully conform to real physical optical characteristics, without physical deviation, and accurately represent the essential physical properties of paper and ink.

[0081] The system interprets the input multispectral information at a physical level, rather than simply extracting visual features. The final output physical property features are not traditional image pixel features, but rather directly correspond to the core physical optical parameters of paper and ink under multispectral conditions, including spectral reflectance, ink absorptivity, paper scattering coefficient, refractive index, and optical attenuation coefficient under multispectral conditions. The entire inference process is constrained by physical laws, ensuring that the output features fully conform to real physical optical properties, without physical deviation, and accurately characterize the essential physical properties of paper and ink.

[0082] According to embodiments of this application, the loss function of the physical information neural network model includes data loss of physical attributes and physical consistency loss; wherein, physical attributes include paper optical attributes, ink optical attributes, printing pressure characteristics and fluorescence attenuation characteristics; physical consistency loss includes energy conservation constraints, wavelength constraints and optical transmission theory constraints.

[0083] For the loss function of the physical information neural network model, a learnable physical parameter layer is designed to model the following physical properties:

[0084] Optical properties of paper: reflectivity, transmittance, scattering coefficient, and absorption coefficient of paper substrate at different wavelengths. The optical response of paper is modeled using the Kubelka-Munk theory.

[0085] Optical properties of inks: absorption coefficient, fluorescence quantum yield, and infrared attenuation coefficient of different colored inks (four-color printing ink, fluorescent ink, infrared absorbing ink).

[0086] Printing pressure characteristics: Local paper deformation caused by gravure / relief printing affects the shadow distribution under side lighting, which is modeled as a change in surface normal vector.

[0087] Fluorescence decay characteristics: The fluorescence intensity decay curve after ultraviolet excitation over time is used to distinguish between genuine and counterfeit fluorescent materials.

[0088] After acquiring physical property information such as paper optical properties, ink optical properties, printing pressure characteristics, and fluorescence attenuation characteristics, the measured data consistent with the actual physical process are calculated through partial differential equations and normalized into predicted value form, which is used as the true value of the actual physical property characteristics for the initial training of the network and the calculation of data term loss. At the same time, the physical consistency loss (physical information error term) under the PINNs framework is introduced, embedding physical prior knowledge into the loss function to constrain the physical parameters output by the network to meet known physical laws, including: energy conservation constraint (reflectivity + absorptivity + transmittance ≤ 1), wavelength constraint, i.e., fluorescence Stokes displacement constraint (emitted light wavelength > excitation light wavelength), and optical transmission theory constraint, i.e., Kubelka-Munk theory constraint (absorption-dissipation ratio is linearly related to concentration).

[0089] Training can be performed using PINNs, with the goal of inversely inferring physical parameters from input multispectral images, thereby achieving the identification of abnormal physical properties. For example, counterfeit credentials may use ordinary paper instead of anti-counterfeiting paper, and their absorption-scattering characteristic curves differ from those of genuine paper; the attenuation characteristics of counterfeit fluorescent ink are different from those of genuine ink.

[0090] The physical information neural network model is trained on real multispectral image data. Its loss function consists of two parts: physical attribute data loss and physical consistency loss. First, the physical attribute data loss between the model's predicted paper optical properties, ink optical properties, printing pressure characteristics, and fluorescence attenuation characteristics and the actual labeled data is calculated. Then, the physical consistency loss is constructed by combining energy conservation constraints, wavelength constraints, and optical transmission theory constraints. Model optimization is achieved by jointly minimizing these two types of losses. The training aims to infer physical parameters from multispectral images. By explicitly modeling optical, fluorescence, and mechanical physical processes, physical laws are embedded into the network as hard constraints. This not only allows the model to accurately learn the various physical characteristics of real samples, ensuring that the output results conform to objective physical laws, but also improves the model's interpretability and generalization ability. It enables the model to effectively identify conventional counterfeit documents based on differences in absorption-scattering characteristic curves, fluorescence attenuation, and other physical indicators, while also recognizing various novel counterfeiting methods never seen before.

[0091] In the embodiments of this application, the physical information neural network model sets a dual loss function, combining data loss and physical consistency loss for training, while introducing multiple optical constraints. The physical information neural network model embeds prior physical knowledge, not only fitting physical property features based on sample data, but also strictly adhering to physical laws such as energy conservation and optical transmission, making the model's inference results physically interpretable. This enables the model to effectively identify counterfeit documents made using ordinary paper and ink, avoiding the problem of purely data-driven models being easily misled by new counterfeiting methods, and significantly improving the reliability of physical dimension authentication.

[0092] In step S240, based on the pre-trained meta-learning task model, a prototype for imitation is constructed from the multispectral image to obtain the prototype features.

[0093] The multispectral image is input into the trained meta-learning task model. The model extracts the image's depth features and, based on the task knowledge and representation rules learned by the model, simulates and generates a replica prototype corresponding to the replica category, outputting the features of the constructed replica prototype.

[0094] It is worth noting that Model-Agnostic Meta-Learning (MAML) is a machine learning paradigm that "learns to learn." It enables models to quickly adapt to new tasks through cross-task training, requiring only a small number of samples to update model parameters. Meta-learning is used to achieve rapid adaptation to new imitation methods with a small number of samples.

[0095] According to embodiments of this application, the training process of the meta-learning task model includes: constructing a task pool based on a learning task of the imitation type, wherein the learning task includes a support set and a query set; sampling a target task based on the task pool, sampling samples from the support set of the target task, calculating a loss function value based on the initialization parameters of the meta-learning task model and the sampled support set, performing an inner loop gradient update based on the loss function value to obtain the inner loop update parameters of the target task; calculating a meta-loss based on the query set and the inner loop update parameters of the target task, and performing an outer loop backpropagation update based on the meta-loss to update the initialization parameters; dynamically adjusting the weights of the global parameters updated by the outer loop according to the task characteristics to obtain task-adaptive weighted parameters; and responding to the update of the learning task of the imitation type, obtaining update samples based on the updated learning task, and performing gradient updates of the model parameters based on the task-adaptive weighted parameters according to the update samples.

[0096] In the model training phase of the meta-learning task, a metric learning branch is enabled. This branch performs prototype classification and metric learning solely for model training. Euclidean distance is used as the distance calculation method to calculate the Euclidean distance between the feature vector of the sample to be detected and the prototypes of each category. Then, a normalized exponential function is used to derive the classification probability, thus completing the prototype training. After model training is complete, prototype features for each category are obtained. In actual use, the metric learning branch is removed, and the prototype features are directly output.

[0097] Considering the possibility of insufficient available data when using the model in subsequent iterations, a meta-learning approach is adopted to enable the model to quickly adapt to new scenarios while avoiding knowledge forgetting. The overall training process involves first building a task library, then conducting meta-training, which is divided into two stages: an inner loop and an outer loop. After each batch of training for all tasks is completed in the inner loop, backpropagation is performed using updated parameter copies to update the overall parameters, which is the outer loop. After the parameters are updated, adjustable weights are used to dynamically adjust the weights based on task characteristics, allowing training to continue as new imitation samples emerge.

[0098] For example, the training process of the meta-learning task model is as follows: different imitation types (such as laser printing imitation, inkjet printing imitation, photocopying imitation, offset printing imitation, hologram imitation, etc.) are treated as different learning tasks. A task pool is constructed according to the learning tasks of imitation types. Each learning task includes a support set (a small number of labeled samples) and a query set (validation samples). During the meta-training phase, the model is trained on a large number of imitation type tasks to learn and initialize parameters, enabling it to quickly adapt to new tasks through a small number of gradient updates.

[0099] The meta-training process is as follows:

[0100] Sample a batch of target tasks {Ti} from the task pool.

[0101] For each target task Ti, sample K samples from the support set, where K is an integer not less than 1.

[0102] Based on the current initialization parameters θ, the loss L is calculated on the support set. Ti(fθ) That is, the loss function value, and perform inner loop gradient updates: θ ' i = θ - α∇L Ti(fθ) α is the inner loop learning rate, θ ' i These are the parameters updated for the inner loop.

[0103] Evaluate the updated parameter θ on the query set. ' i The meta-loss is the query set loss L. Ti ' (fθ) The sum of L Ti ' (fθ) This indicates the calculation of loss on the query set. The meta-loss is calculated, and backpropagation in the outer loop is performed to update the model's global initialization parameters. The initialization parameters θ are updated through backpropagation.

[0104] Introducing a task-adaptive module into the meta-learning framework, the attention weights of the feature extraction layer are dynamically adjusted according to the characteristics of the current task (such as image distribution and imitation type), realizing task-adaptive feature reweighting, determining task-adaptive weighting parameters, and enabling the model to quickly focus on the key discriminative features of the current task.

[0105] After meta-training, for newly added imitation type samples, a small number of updated samples are used to form a new task support set. Based on the task-adaptive weighted parameters, a small number of gradient updates are performed to achieve rapid adaptation to the new imitation type without forgetting. The few-sample adaptation process is as follows: when a new imitation method appears, only a small number of samples (e.g., 5-10) need to be collected as a support set. Based on the initialization parameters obtained from meta-learning, a small number of gradient updates (usually 5-10 steps) are performed to obtain a meta-learning task model suitable for the new task.

[0106] In the embodiments of this application, a task pool is constructed to divide the support set and query set. Meta-learning training is completed by combining inner and outer loop double-layer gradient updates, and then adaptive parameters are obtained by dynamically adjusting the weights. This training method fits the needs of few-shot learning, can quickly adapt to different imitation types of tasks, and significantly improves the model update response speed. At the same time, it supports iterative optimization of parameters based on newly added imitation samples, and can adapt to new imitation techniques with a small number of new samples, strengthening the model's adaptability to new imitation methods and reducing the cost of massive sample collection and training.

[0107] In step S250, the pre-trained multimodal anti-counterfeiting model is used to identify the spectral texture fusion features, physical attribute features, and counterfeit prototype features to obtain the identification result of the target credential.

[0108] The multimodal counterfeit detection model is a large-scale deep learning model that can simultaneously fuse and analyze various types of data, such as images, spectra, and physical properties, specifically designed for identifying genuine and counterfeit samples. The multimodal counterfeit detection model comprises a sliding window transformation network, a physical network, a convolutional network, a cross-modal interaction layer, and a multi-task prediction layer. The overall training is divided into two phases. The first phase involves pre-training the large model using a massive amount of real-world samples of tickets, certificates, and cash. Self-supervised pre-training is performed on the sliding window transformation network, and this phase focuses solely on training the sliding window transformation network. A masked multispectral modeling task is used, randomly obscuring local areas of certain spectral bands, allowing the model to reconstruct the obscured spectral information and learn general multispectral features. The second phase is multi-task joint training, integrating all branches of the physical network, convolutional network, cross-modal interaction layer, and multi-task prediction layer for training, while simultaneously optimizing multiple task objectives. The system employs focus loss for true / false binary classification, optimizes spectral texture attention fusion loss, uses mean square error regression loss to calculate physical attribute regression loss, and includes imitation type classification loss and uncertainty estimation loss.

[0109] It is worth noting that after the independent training of the spectral texture attention fusion model, the physical information neural network model, the meta-learning task model, and the multimodal anti-spoofing model is completed in sequence, the models are networked together as a whole, and multi-task joint fine-tuning is carried out based on labeled samples to achieve collaborative adaptation of the models.

[0110] The spectral texture fusion features, physical attribute features, and counterfeit prototype features corresponding to the target voucher are extracted. These three types of features are then uniformly input into the trained multimodal authentication model. Based on the multimodal discrimination rules learned in the previous stage, the model performs fusion analysis and authenticity determination on various features, ultimately generating the authentication result corresponding to the target voucher, such as authenticity probability, counterfeit type, and uncertainty probability. When the uncertainty probability exceeds a threshold, a manual review process is triggered.

[0111] In the embodiments of this application, multi-band image acquisition and preprocessing provide high-quality basic data for subsequent spectral image analysis. A spectral texture attention fusion model is used to finely fuse multispectral information, deeply mine cross-band complementary features, and extract spectral texture fusion features, balancing image fidelity and detail. A physical information neural network combines spectral physical characteristics to perform inference, extracting physical attribute features and accurately capturing the physical attributes of the voucher. A counterfeit prototype is constructed based on a meta-learning task model, adapting to various counterfeiting methods. A multi-modal counterfeit detection model integrates multiple feature categories for comprehensive discrimination, significantly improving the accuracy, stability, and generalization ability of voucher counterfeiting detection, effectively preventing various counterfeiting risks, and ensuring the security of financial transactions.

[0112] According to an embodiment of this application, the credential authentication method further includes: calling a real-time authentication interface to obtain the authentication result of the target credential, and generating an abnormal area heat map and physical attribute report based on the authentication result; and calculating the transaction risk level based on the authentication result and transaction information, and performing a risk warning operation if the transaction risk level is higher than a preset risk threshold.

[0113] This system integrates models such as a spectral texture attention fusion model, a physical information neural network model, a meta-learning task model, and a multimodal authentication model to provide real-time authentication services. Upon calling the real-time authentication interface and acquiring the multispectral image of the target document, the interface will return the authentication result within a short time (e.g., 500 milliseconds). Based on this result, a heatmap of abnormal areas, such as areas with missing ultraviolet fluorescence and abnormal infrared absorption, is generated. Simultaneously, deviations of the physical properties of the paper and ink from standard values ​​are output, forming a physical property report for manual verification. The system comprehensively calculates the transaction risk level by combining the document authentication result with transaction information such as transaction amount and historical transaction records. If the risk level exceeds a preset risk threshold, a risk warning is issued. If an immediate risk warning is triggered, high-risk documents are automatically intercepted and notifications are sent to risk control personnel.

[0114] In the embodiments of this application, the authentication results are quickly returned through a real-time authentication interface, generating an anomaly heatmap and physical attribute report to intuitively display the counterfeit areas and abnormal attributes. The authentication results are combined with transaction data to determine the risk level and trigger an alert, realizing the integration of authentication, analysis, and risk control. This not only completes the identification of the authenticity of the vouchers but also assists staff in locating problems, assessing transaction risks, proactively preventing financial risks, and forming an intelligent risk control system.

[0115] According to embodiments of this application, the credential authentication method further includes: in response to a user modifying the authentication result of the target credential, obtaining the corrected authentication result and storing it in a false alarm database; and based on the false alarm database and a timed training strategy, using continuous learning technology to incrementally train the credential authentication model; wherein the credential authentication model includes one or more of the following: a spectral texture attention fusion model, a physical information neural network model, a meta-learning task model, and a multimodal authentication model.

[0116] It provides a false alarm correction interface, supporting manual online review and correction of identification results. When a user manually modifies the identification result of the target credential, the corrected identification result is collected and the corrected data is completely entered into the false alarm database for retention.

[0117] Based on the preset timed training strategy, such as reading updated corrected samples from the false alarm database during predetermined time periods (e.g., weekly off-peak business hours, daily non-business hours), incremental training of the credential authentication model is carried out using continuous learning techniques such as elastic weight consolidation and memory-aware synapses. While learning from newly added corrected data and optimizing the model's recognition capabilities, this effectively avoids the catastrophic forgetting problem in the model and achieves iterative upgrades of the model.

[0118] In the embodiments of this application, incremental model training is achieved based on a false positive database and continuous learning technology. This supports updating the sample library based on manual correction results and periodically iteratively optimizing each core model. Without large-scale retraining, model misjudgment issues can be corrected, continuously adapting to ever-changing counterfeiting methods. This forms a closed-loop mechanism of manual correction, data accumulation, and model iteration, continuously reducing the false positive rate of counterfeit detection, maintaining long-term counterfeit detection performance and generalization ability, and extending the model's lifespan.

[0119] For example, based on the above-mentioned credential authentication method, a multispectral image fusion intelligent authentication system is constructed. Figure 3 A schematic diagram illustrating the structure of a multispectral image fusion intelligent anti-counterfeiting system according to an embodiment of this application is shown. Figure 3As shown, the multispectral image fusion intelligent anti-counterfeiting system mainly includes a multispectral image acquisition and preprocessing module, a spectral-texture attention fusion module, a neurophysical attribute reasoning module, a meta-learning few-sample adaptive module, a multimodal anti-counterfeiting large model, a real-time anti-counterfeiting and risk warning module, and an incremental update module. The multispectral image acquisition and preprocessing module acquires visible light, infrared, and ultraviolet multi-band images of documents (such as tickets, certificates, and cash). After preprocessing, these images are input into the spectral-texture attention fusion module to generate enhanced spectral-texture fusion features that fuse multispectral anti-counterfeiting information. The spectral-texture attention fusion model achieves deep interaction of multi-band information through spectral spatial demixing, significantly improving the recognition ability of complex anti-counterfeiting features (such as combined ultraviolet fluorescence and infrared absorption anti-counterfeiting). Compared with existing simple fusion methods, the anti-counterfeiting accuracy is improved. The neurophysical attribute reasoning module processes spectral images based on physical priors to simulate the optical response of paper and ink, outputting physical attribute features. The meta-learning few-shot adaptive module enables the model to quickly adapt to new forgery methods from a small number of new samples. It constructs a forgery prototype of the spectral image, extracts the prototype's features, embeds prior physical knowledge, and combines it with meta-learning training. This effectively improves the model's robustness to domain changes, maintaining stable performance under different ticket types, imaging devices, and lighting conditions, and enhancing cross-scene generalization ability. Spectral texture fusion features, physical attribute features, and forgery prototype features are input into the multimodal authentication module for authenticity determination to obtain the identification result. The real-time authentication and risk warning module outputs the identification result and triggers an alert. The incremental update module continuously optimizes the model using new data.

[0120] Figure 4 A schematic diagram illustrating the spectral texture attention fusion flowchart of the credential authentication method according to an embodiment of this application is shown. Figure 4 As shown, according to an embodiment of this application, preprocessing a multi-band image may include steps S410 to S440.

[0121] In step S410, image registration is performed on the multi-band images.

[0122] Image registration can employ feature-point-based registration algorithms, such as accelerated robust feature algorithms combined with random sample consensus algorithms, or scale-invariant feature transformation algorithms combined with robust principal component analysis algorithms, to ensure pixel-level alignment of images across different bands and achieve pixel-level registration. Alternatively, a closed-loop control and state diagnosis mechanism can be used to monitor and correct the spatial pose (e.g., position, orientation) of the data in real time, dynamically correcting deviations.

[0123] In step S420, based on the band information of the multi-band image, the registered multi-band image is subjected to illumination unevenness correction to obtain a multi-band corrected image.

[0124] During imaging, the layout of the light source and the ambient light can affect the image, which may result in brightness gradients and local shadows. The illumination and shadow correction can be performed on each registered band image individually. By fitting the background illumination distribution model of each band, the brightness deviation caused by uneven light source can be offset one by one, the influence of uneven light source distribution can be eliminated, the gray level of the whole image can be balanced, and the imaging interference caused by light and shadow distortion can be removed.

[0125] In step S430, the multi-band corrected image is input into the pre-trained image reconstruction model of the corresponding band for super-resolution reconstruction to obtain the multi-band reconstructed image.

[0126] Multi-band corrected images are input into corresponding lightweight image reconstruction models, one band at a time. The original band images have low resolution due to hardware limitations, resulting in blurred subtle anti-spoofing textures. Based on pre-trained reconstruction models for each band, missing pixel information is filled in. This restores fine textures while preserving the spectral characteristics and spatial correspondence of the bands, improving image clarity and resolution, and generating a multi-band reconstructed image with complete details.

[0127] When training the image reconstruction model, it is first pre-trained using publicly available high- and low-resolution image data to give the model basic image reconstruction capabilities. Then, a small number of high-resolution voucher images captured by high-precision equipment are used as ground truth, along with corresponding low-resolution images collected by a large-scale deployment of low-precision terminals, to construct a precisely matched, dedicated dataset. Based on these real industry samples, the pre-trained model is fine-tuned by band and category parameters, continuously optimizing the image reconstruction model's detail restoration capabilities, ultimately resulting in a dedicated super-resolution image reconstruction model adapted to the imaging characteristics of the terminals.

[0128] In super-resolution reconstruction, the input multi-band images correspond to multiple image reconstruction models, and their training data are different, resulting in different network parameters. However, the network structure and training method are the same. The loss function is a weighted sum of norm loss and perceptual loss, with high-precision images as the training target.

[0129] According to embodiments of this application, the image reconstruction model includes a shallow feature extraction layer, a deep feature extraction layer, and a reconstruction layer. A multi-band corrected image is input into a pre-trained image reconstruction model for the corresponding band for super-resolution reconstruction. This includes: using the shallow feature extraction layer of the current band to extract shallow features from the corrected image of the current band, obtaining target shallow features; using the deep feature extraction layer of the current band to perform differential modulation and residual fusion on the target shallow features, generating target deep features; wherein the deep feature extraction layer includes a local feature enhancement network, a differential feature modulation network, a multi-scale feature aggregation network, a channel attention enhancement network, a residual feature extraction network, and a residual fusion network; and using the reconstruction layer of the current band to perform convolutional reconstruction of the target deep features, obtaining a reconstructed image of the current band.

[0130] Figure 5A The diagram schematically illustrates the structure of an image reconstruction model for a credential authentication method according to an embodiment of this application. For example... Figure 5A As shown, in the shallow feature extraction layer, the corrected image of the current band is input and subjected to convolution for shallow feature extraction to obtain the target's shallow features. These shallow features are then input into the deep feature extraction layer to output the target's deep features. The high-resolution reconstruction layer processes the target's deep features through three 3×3 convolutions to reconstruct a high-resolution image. The following processing is performed sequentially in the deep feature extraction layer:

[0131] The local feature enhancement network performs deep convolution operations on the shallow features of the target and sets residual connections to enhance the local feature representation.

[0132] The Differential Feature Modulation (DIM) network first extracts local features through convolution, then obtains the global context vector through global average pooling. It calculates the difference between the local features and the global context vector, multiplies this difference by learnable weight parameters to scale the features, thus reweighting them. The weighted features are then added to and fused with the local features, and finally activated by an activation function to suppress interactions with low information content, allowing the network to focus on learning a balanced global feature representation.

[0133] The multi-scale feature aggregation network divides the activated features into multiple groups along the channel dimension and samples them through max pooling operations with multiple strides (e.g., 1, 2, 4, 8). For sampling results with a stride greater than 1, convolution processing is performed first, followed by upsampling to restore the original resolution. The multi-scale processed features are then concatenated along the channel dimension, combining local details and semantic information, and finally fused through convolution.

[0134] The channel attention enhancement network processes the activated features (features before channel segmentation) through convolution and activation functions, multiplies them pixel-by-pixel with the activated features after multi-scale aggregation, generates a channel attention map to highlight important features, and then outputs it through convolution to integrate multi-scale context, local details and long-term dependencies.

[0135] The residual feature extraction network divides the channel attention-enhanced features into two parallel branches, each undergoing convolutional processing. One branch performs only convolution to obtain convolutionally processed features, while the other branch performs self-attention to obtain self-attention processed features, enhancing its ability to focus on autocorrelated features. The output features of the two branches (the convolutionally processed features and the self-attention processed features) are concatenated channel by channel and then further processed by residual blocks for feature extraction. Finally, the self-attention processed features are concatenated again with the features output from the residual blocks to reduce feature forgetting.

[0136] The residual fusion network sequentially passes the spliced ​​features through a combination of two convolutional, normalization, and activation functions to adjust the information and obtain the initial deep features. Then, it performs a skip connection with the target shallow features to superimpose them and obtain the target deep features.

[0137] For example, Figure 5B The diagram illustrates the network connection of an image reconstruction model for a credential authentication method according to an embodiment of this application. For example... Figure 5B As shown, the image reconstruction model first extracts shallow features from the input image through a 3×3 convolution (Conv), then performs deep feature extraction to refine features and expand information. Subsequently, it connects to three layers of three 3×3 convolutions for reconstruction to complete image reconstruction and output a high-resolution image. The backbone is set with global residual connections.

[0138] In deep feature extraction, depthwise separable convolution (DWConv) is first performed with residual connections to shallow features to enhance local feature representation capabilities with fewer parameters. Subsequently, feature modulation is implemented, using a differential feature extraction strategy to improve the continuity and diversity of global texture representation. First, 1×1 convolution is used to extract local features, then global average pooling (GAP) is used to obtain the global context vector. The difference between the local features and the global context vector is calculated, and this difference is multiplied by learnable weight parameters to achieve scaling weighting. The weighted features are then added to and fused with the local features before being activated by a Gaussian error linear activation function (GeLU). This dynamically eliminates low-information feature interactions, allowing the network to focus more on learning a balanced global feature representation. Next, a multi-scale feature aggregation structure is introduced, adaptively aggregating features at different scales from local to global. The activated features are divided into four groups along the channel dimension and processed separately. Each group is sampled by max pooling with strides of 1, 2, 4, and 8 in sequence. Branches with strides greater than 1 perform downsampling to obtain a larger receptive field. Each group of features is then processed by 3×3 convolution. Branches with strides of 1 use nearest neighbor interpolation to upsample and restore the original resolution. The features after the four scale processing are concatenated in the original direction, combining fine local details and coarse semantic information. After the scale features are concatenated, they are fused by 1×1 convolution and GeLU activation is performed to obtain scale-activated features. At the same time, the original features before channel segmentation are separately processed by 1×1 convolution and GeLU activation to obtain channel-activated features. The channel-activated features and scale-activated features are multiplied pixel by pixel to generate a channel attention map to highlight important features. Finally, the features are output by 1×1 convolution to obtain the channel attention-enhanced features, realizing the overall fusion of multi-scale context, local details, and long dependencies. Subsequently, a new round of internal feature extraction is carried out. The obtained channel attention-enhanced features are fed into two 3×3 convolutional branches in parallel. One branch introduces a self-attention mechanism to enhance the self-associative feature capture capability. The output features of the two branches are concatenated along the channel dimension (Cat), and then three residual blocks (Resblock) are connected to further extract features. The features that have undergone self-attention operation are concatenated again to enhance features and alleviate feature forgetting. Then, the feature information is corrected by a combination layer of two sets of convolution (Conv) + batch normalization (BN) + linear rectified activation function (ReLU). Finally, the feature is fused with the shallow features by skip connections to obtain the final deep features.

[0139] In the embodiments of this application, super-resolution processing is achieved through shallow and deep feature extraction and image reconstruction. The deep network integrates a local feature enhancement network, a differential feature modulation network, a multi-scale feature aggregation network, a channel attention enhancement network, a residual feature extraction network, and a residual fusion network. Combining differential modulation and residual fusion to optimize feature representation, it can specifically enhance the detailed information of images in different spectral bands. The image reconstruction models for different spectral bands can adapt to the imaging characteristics of different bands such as visible light, infrared, and ultraviolet, accurately repairing blurred areas, restoring microscopic anti-counterfeiting details, and improving the resolution and quality of multispectral images.

[0140] In step S440, image enhancement is performed on the multi-band reconstructed image to obtain a multispectral image.

[0141] Image enhancement is performed on the reconstructed image of the current band to obtain the spectral image corresponding to the current band. Image enhancement methods can include adaptive histogram equalization to improve local contrast and highlight microscopic physical features; linear / nonlinear grayscale mapping to adjust pixel brightness and contrast and improve overall visual and feature representation; and sharpening filtering to enhance edge and texture details.

[0142] In the embodiments of this application, multi-band images are preprocessed sequentially, including registration, illumination correction, super-resolution reconstruction, and image enhancement, to obtain enhanced spectral images for each band. Image registration eliminates multi-band positional deviations, illumination correction avoids interference from ambient light, super-resolution reconstruction improves image clarity, and image enhancement highlights anti-counterfeiting details. The preprocessing workflow comprehensively optimizes the quality of the original multi-band images, eliminates imaging noise and environmental interference, and provides a high-quality data source for subsequent feature processing, ensuring the accuracy of counterfeit detection from the source.

[0143] Figure 6 This schematically illustrates a flowchart of the spectral texture attention fusion method for credential authentication according to an embodiment of this application. Figure 7A This diagram schematically illustrates the structure of a spectral texture attention fusion model for a credential authentication method according to an embodiment of this application. Figure 7B The diagram illustrates the network connection of the spectral texture attention fusion model of the credential authentication method according to an embodiment of this application. Figure 6 , 7A As shown in Figure 7B, according to an embodiment of this application, the spectral texture attention fusion model includes a spectral unmixing network, a spatial texture feature extraction network, a spectral feature extraction network, and an attention fusion network; processing multispectral images based on the pre-trained spectral texture attention fusion model to obtain spectral texture fusion features may include steps S710 to S740.

[0144] In step S710, the spectral and spatial information of the multispectral image are extracted by a spectral unmixing network, the spectral information of the multispectral image is spliced ​​together to obtain spectral features, and the spatial information of the multispectral image is spliced ​​together to obtain spatial features.

[0145] The spectral unmixing network decouples spectral features, separating the spectral and spatial dimensions of multispectral images (such as visible light images, infrared images, ultraviolet fluorescence images, lateral light images, and transmission light images), and extracting the spectral and spatial information of each image separately. The extracted spectral information from all images is then concatenated and merged to form a complete spectral feature. Similarly, the spatial information corresponding to all images is concatenated and merged separately to obtain a complete spatial feature.

[0146] The spectral unmixing network uses two cascaded 1×1 convolutional layers with activation functions to achieve pixel-level spectral transformation and unmixing. After unmixing, the spectral and spatial information of each image are concatenated and input into the spectral feature extraction network and the spatial texture feature extraction network, respectively.

[0147] The spectral unmixing network consists of two 1×1 convolutional layers (Conv) with consecutively configured activation functions. This structure performs spectral feature transformation independently for each single pixel. The two-layer design improves feature representation while ensuring computational efficiency, resulting in superior extraction performance compared to conventional single-layer unmixing structures. The input is feature data arranged in a spatial dimension of H×W×C, where H, W, and C represent the height, width, and number of spectral bands, respectively. The learnable weight space of the first convolutional layer is 1×1×C×F, where F is the learnable weight of the output feature dimension. After nonlinear mapping using the Rectified Linear Activation Function (ReLU), the data is fed into the second convolutional layer, which has a learnable weight space of 1×1×F×F. Spectral unmixing is then performed again. After unmixing, the spectral dimension information of each spectral image is concatenated, and the spatial dimension information is concatenated separately. Independent feature extraction is then performed on both the concatenated spectral and spatial information.

[0148] In step S720, high and low frequency features are extracted from spatial features based on the spatial texture feature extraction network to obtain enhanced texture features.

[0149] The spatial texture feature network first performs wavelet transform on the spatial features to separate low-frequency and high-frequency components. The low-frequency components are enhanced by combining Fourier transform, convolution, and double pooling. The high-frequency components are concatenated and convolved before being connected to a pyramid network with skip connections to extract features. Finally, the enhanced high and low frequency components are subjected to inverse wavelet transform to output the enhanced texture features.

[0150] According to embodiments of this application, high- and low-frequency features are extracted from spatial features using a spatial texture feature extraction network to obtain enhanced texture features. This includes: performing wavelet transform on the spatial features to obtain low-frequency and high-frequency components; performing Fourier transform on the low-frequency components, convolving and double-pooling the transformed low-frequency components to obtain low-frequency fusion components, and performing inverse Fourier transform on the low-frequency fusion components to obtain enhanced low-frequency features; concatenating high-frequency components, performing convolution adjustment on the concatenated high-frequency components, and adding the adjusted high-frequency components to the spatial features to obtain spatial high-frequency features; performing jump fusion on the spatial high-frequency features using a U-shaped network to obtain high-frequency fusion features, and performing wavelet transform on the high-frequency fusion features to obtain enhanced high-frequency features; and performing inverse wavelet transform on the enhanced high-frequency features and enhanced low-frequency features to obtain enhanced texture features.

[0151] To highlight key structures such as printing edges, dot boundaries, and fiber textures, wavelet transform is used to decompose spatial features into low-frequency components (such as the lowest band) and high-frequency components (such as bands two, three, and four). The low-spectrum data is an approximate low-frequency component with a relatively flat overall value. Therefore, this low-frequency component is separated and subjected to Fourier transform to refine the frequency structure of the low-frequency component, achieving higher frequency resolution in the low-frequency range and facilitating the identification of weak and adjacent low-frequency components. Subsequently, a convolution (Conv)-rectified linear activation function (ReLU)-convolution (Conv) operation process is used to extract feature information. Then, average pooling (GAP) and max pooling (GMP) operations are performed in parallel to complete feature processing along the channel dimension. Channel weight information under multiple strategies is fused, and the two pooling results are added together and then convolution is performed again. Finally, the enhanced low-frequency features are obtained through inverse Fourier transform, thus enhancing the low-frequency components.

[0152] For high-frequency components across multiple bands, a stitching (Cat) operation is first performed, using 1×1 convolution to adapt the feature dimensions. Then, the features are added element-wise to the original spatial features before wavelet transform, and fed into a U-Net (U-shaped symmetric encoder-decoder network) with a skip connection mechanism. The U-Net possesses pyramidal multi-scale hierarchical characteristics. At the encoding end, each layer is connected to a cubic residual module (Resblock) to extract features and downsample them progressively. At the decoding end, after upsampling at each layer, the corresponding shallow features are fused using U-Net skip connections. This is then further refined and reconstructed using a cubic residual module, resulting in optimized high-frequency fused features. A wavelet transform is then applied to the three optimized sets of high-frequency fused features to obtain enhanced high-frequency features. These enhanced high-frequency features are then fed into an inverse wavelet transform operation along with enhanced low-frequency features to obtain enhanced texture features. After multiple rounds of feature extraction and optimization, the high-frequency components have richer high-frequency information details, ultimately improving image clarity.

[0153] In the embodiments of this application, wavelet transform, Fourier transform, and U-shaped network are used to process spatial features, and texture information is optimized step by step according to high and low frequency components. The low frequency components are transformed and convolved to improve the overall image quality, while the high frequency components are enhanced to extract details and retain key features such as edges and textures through skip fusion. Finally, high-quality enhanced texture features are reconstructed to enhance spatial texture details such as voucher paper and printing marks, solve the problem of detail loss in traditional methods, and facilitate accurate identification of texture-based counterfeit features.

[0154] In step S730, the spectral features are fused at multiple scales based on the spectral feature extraction network to obtain enhanced spectral features.

[0155] The spectral feature extraction network achieves feature reweighting and activation through convolution, global pooling, and learnable weights. Then, the features are grouped by channel, and multi-scale feature fusion is completed by asynchronous long pooling, residual modules, and upsampling. Finally, combined with the channel attention mechanism, the network outputs enhanced spectral features.

[0156] According to embodiments of this application, multi-scale feature fusion of spectral features based on a spectral feature extraction network to obtain enhanced spectral features includes: convolving the spectral features to extract local spectral features; performing global average pooling on the local spectral features to obtain a global context vector; performing difference weight activation on the global context vector and local spectral features based on weight parameters to obtain activated features; segmenting the activated features based on channel dimensions to obtain multiple sets of spectral features; performing multi-scale feature mapping on the multiple sets of spectral features according to preset multiple strides to obtain multi-scale features; concatenating and fusing the multi-scale features to obtain multi-scale fused features; performing convolution and activation on the multi-scale fused features to obtain a first activated feature; performing convolution and activation on the activated features to obtain a second activated feature; multiplying the first activated feature and the second activated feature pixel by pixel to obtain channel attention features, and performing convolution on the channel attention features to obtain enhanced spectral features.

[0157] To calculate the features of different scale band feature maps in the channel dimension and learn the contribution of each band to the current scale, the structure of the spectral feature extraction network is similar to that of the image reconstruction model mentioned above. First, 1×1 convolution is used for initial feature extraction to obtain local spectral features. Then, global average pooling is used to obtain the global context vector. The difference between the local spectral features and the global context vector is calculated. This difference is multiplied by the learnable weight parameters to complete the scaling operation, thereby realizing feature reweighting. The weighted features are added to the original local spectral features to complete feature fusion. Subsequently, Gaussian error linear activation function (GeLU) is used for activation processing to obtain activated features. Low information content feature interaction is dynamically suppressed, prompting the network to learn a balanced global feature representation.

[0158] The activated features are then divided into four groups along the channel dimension and processed independently. Each group uses max pooling for feature sampling, and the preset strides for the multiple groups can be set to 1, 2, 4, and 8 respectively. Each group of features then undergoes scale feature mapping, which includes max pooling, convolution, and upsampling. Upsampling uses nearest neighbor interpolation to restore the original resolution. The scale features processed at the four scales are then concatenated along their original dimensions to complete multi-scale feature fusion, simultaneously taking into account both fine local details and coarse-grained semantic information. The concatenated multi-scale fused features are then integrated using a 1×1 convolution and activated by GeLU to obtain the first activated feature.

[0159] Simultaneously, for the activated features (features before the channel segmentation operation), a 1×1 convolution combined with GeLU activation is performed separately to obtain the second activation feature. The second activation feature is then multiplied pixel by pixel with the first activation feature obtained after fusion to generate channel attention features, i.e., channel attention maps, thereby strengthening the weights of key features. Finally, an enhanced spectral feature is output through a 1×1 convolution layer, enabling the enhanced spectral feature to integrate spectral information from multi-scale context, local detail spectral information, and long-term dependencies.

[0160] In the embodiments of this application, spectral information is optimized by combining global context, weighted activation, multi-scale downsampling, and channel attention mechanisms to achieve multi-scale fusion of spectral features. This allows for the simultaneous capture of local spectral details and global spectral patterns, adapting to differences in spectral features across different bands and enhancing the characterization capabilities of optical features such as ultraviolet fluorescence and infrared absorption. The multi-scale structure enables the model to adapt to anti-counterfeiting areas of different sizes and locations, significantly improving the ability to distinguish between genuine and counterfeit credentials across multiple spectral dimensions.

[0161] In step S740, the enhanced texture features and enhanced spectral features are fused based on the attention fusion network to generate spectral texture fusion features.

[0162] The attention fusion network has spectral self-attention, spatial self-attention and cross-attention branches. The outputs of each branch are spliced ​​and processed by a multilayer perceptron and a convolution module. Then, they are superimposed with enhanced spectral features and enhanced texture features to finally generate spectral texture fusion features.

[0163] In the embodiments of this application, a spectral texture attention fusion model is built based on a spectral unmixing network, a spatial texture feature extraction network, a spectral feature extraction network, and an attention fusion network. First, the spectral unmixing network separates spectral and spatial information. Then, the two types of features are enhanced by dedicated feature extraction networks. Finally, the attention fusion network completes the fusion, realizing the decoupling processing of spectral and spatial information and refined feature extraction. This breaks through the limitations of traditional simple splicing and fusion, fully explores the complementary information of multi-band images such as visible light, infrared, and ultraviolet, takes into account both image texture details and spectral feature integrity, and effectively improves the recognition accuracy of complex anti-counterfeiting features.

[0164] According to embodiments of this application, enhanced texture features and enhanced spectral features are fused based on an attention fusion network to generate spectral texture fusion features, including: performing self-attention extraction on enhanced texture features to obtain texture self-attention features; performing self-attention extraction on enhanced spectral features to obtain spectral self-attention features; performing cross-attention extraction on enhanced spectral features and enhanced texture features to obtain cross-attention features; concatenating the cross-attention features, texture self-attention features, and spectral self-attention features, and performing multilayer perception and convolution processing on the concatenated attention features to obtain enhanced attention features; and fusing the enhanced attention features, enhanced spectral features, and enhanced texture features to generate spectral texture fusion features.

[0165] When fusing enhanced texture features and enhanced spectral features using an attention fusion network, the process begins by first extracting texture self-attention features from the enhanced texture features using a spatial self-attention branch, and extracting spectral self-attention features from the enhanced spectral features using a spectral self-attention branch. A cross-attention branch then performs two sets of cross-attention processes in opposite order to perform interactive computation of spectral and spatial information, extracting the cross-attention features. The features output from these branches are then concatenated (Cat), and processed by a multilayer perceptron (Mlp) and two sets of convolutions (Conv)-batch normalization (BN)-rectified linear activation functions (ReLU) to obtain enhanced attention features. These enhanced attention features are then superimposed with the original enhanced spectral features and enhanced texture features. Finally, a three-layer convolution adjustment is performed to achieve multi-feature fusion, generating a spectral-texture fusion feature.

[0166] In the embodiments of this application, the intrinsic correlation between texture and spectral features is mined using a self-attention mechanism. After splicing and fusion, the features are further enhanced by multi-layer perception and convolution. The original spectral features are then superimposed to generate the final spectral-texture fusion feature. Through multi-feature fusion processing, key anti-counterfeiting areas in the image can be autonomously focused, effective features are strengthened, and invalid interference information is suppressed. This achieves deep complementary fusion of spectral and texture features, highlighting core anti-counterfeiting information while ensuring image fidelity, and improving the effectiveness and recognizability of multi-spectral fusion features.

[0167] According to embodiments of this application, the physical information neural network model includes convolutional layers, residual modules, and a multi-task dense prediction module. The pre-trained physical information neural network model performs physical attribute reasoning on multispectral images to obtain physical attribute features, including: extracting features from the multispectral images based on convolutional layers to obtain shallow features; processing the shallow features through the residual module to obtain deep features; and predicting physical attributes based on the deep features through the multi-task dense prediction module to obtain physical attribute features.

[0168] Figure 8A This diagram schematically illustrates the structure of a physical information neural network model for a credential authentication method according to an embodiment of this application. Figure 8B The diagram illustrates the network connection diagram of the physical information neural network model of the credential authentication method according to an embodiment of this application, such as... Figure 8A , 8B As shown, firstly, any spectral image is input into the convolutional layer, and shallow features are extracted through one convolution (Conv) layer to obtain shallow features. Then, the shallow features are input into the residual (Resblock) module for deep feature learning, and deep features are output. Features can be extracted using 16 residual modules. Finally, the deep features are processed by the multi-task dense prediction module, and the physical property features such as paper optical properties, ink optical properties, printing pressure features, and fluorescence attenuation characteristics are finally output.

[0169] It is worth noting that after the deep features are input into the multi-task dense prediction module, the common features are first extracted through multiple shared convolutional layers consisting of 3×3 Conv-BN-ReLU to enhance the overall feature representation capability. Then, the modules are split into multi-task branches, and each branch uses an independent 1×1 convolutional layer or multiple 3×3 convolutions to map the shared features to the specific features of the channel dimension required by the corresponding task. Finally, the output features of each branch are activated by applying sigmoid (S-type activation function) to obtain the regression prediction results for different tasks.

[0170] In the embodiments of this application, the physical information neural network model performs feature extraction and physical attribute inference in a hierarchical manner. The convolutional layer quickly extracts shallow basic features, the residual module solves the problem of stable output of deep features, and the multi-task dense prediction module accurately infers various optical physical attributes. Each network module has a clear division of labor and high feature extraction efficiency, which can accurately model physical anti-counterfeiting features such as paper, ink, and fluorescence, and improve the accuracy of physical attribute recognition.

[0171] According to embodiments of this application, the meta-learning task model includes a decoupling layer, a feature extraction layer, and a prototype computation layer. Based on the pre-trained meta-learning task model, a prototype is constructed from a multispectral image to obtain prototype features. This includes: performing convolution operations and gradient inversion on the multispectral image based on the decoupling layer to obtain decoupling features; performing partial convolution residual processing on the decoupling features based on the feature extraction layer to obtain microscopic physical features; adding the decoupling features and microscopic physical features based on a skip connection method to obtain microscopic fusion features; and calculating the prototype features corresponding to the microscopic fusion features using an embedding function based on the prototype computation layer.

[0172] Figure 9A This diagram schematically illustrates the structure of the meta-learning task model of the credential authentication method according to an embodiment of this application. Figure 9B The diagram schematically illustrates the network connection diagram of the meta-learning task model of the credential authentication method according to an embodiment of this application. For example... Figure 9A , 9B As shown, the meta-learning task model introduces a physical decoupling mechanism. In the adversarial decoupling layer, convolution (Conv) and gradient inversion are used to remove the textual semantic information in the image embedding vector, thereby decoupling the macroscopic semantic features and microscopic physical features of the credentials.

[0173] The decoupled features are input into the feature extraction layer, which consists of three residual blocks (Resblock) based on partial convolution (PConv) connected in series. The feature extraction layer performs partial convolution residual processing on the decoupled features and outputs microscopic physical features.

[0174] The decoupling features are directly added to the microscopic physical features output by the three residual blocks via global skip connections. Specifically, the decoupling features are treated as residual features, bypassing the three residual blocks, and are added element-wise to the microscopic physical features output by the residual blocks. The output of the feature extraction layer (microscopic physical features) is then added to the output of the decoupling layer (decoupling features) to obtain the microscopic physical fusion features. Partial convolution enhances the ability to extract microscopic physical features while maintaining computational efficiency, and skip connections form a global residual learning mechanism.

[0175] Based on the prototype computation layer, microscopic physical fusion features are mapped to embedding vectors through an embedding function, and the mean of the embedding vectors for each category is calculated to obtain the imitation prototype features for that category. The prototype computation network is input, containing an embedding function with learnable parameters. The embedding function calculates an M-dimensional prototype representation for each category, where each prototype is the mean vector of all embedding support point vectors in the corresponding category. The output of the prototype computation layer is further processed by a metric learning layer to obtain the classification probability of the image.

[0176] In the embodiments of this application, the meta-learning task model achieves effective decoupling of image features through convolution and gradient inversion in the decoupling layer. Based on the partial convolutional residual structure of the feature extraction layer, it deeply mines microscopic physical information. Global skip connections fuse decoupling features and microscopic physical features, preserving microscopic details. The embedding function of the prototype computation layer accurately generates counterfeit prototype features. The model has a clear hierarchy and efficiently transmits features, enhancing the meta-learning task model's ability to represent various types of counterfeiting and providing reliable feature support for identifying new counterfeit credentials.

[0177] According to embodiments of this application, the multimodal anti-counterfeiting model includes a sliding window transformation network, a physical network, a convolutional network, a cross-modal interaction layer, and a multi-task prediction layer. The pre-trained multimodal anti-counterfeiting model performs identification processing on spectral texture fusion features, physical attribute features, and counterfeit prototype features to obtain the identification result of the target credential. This includes: inputting the spectral texture fusion features into the sliding window transformation network to extract visual features; inputting the physical attribute features into the physical network to extract physical features; wherein the physical network includes a multilayer perceptron, a residual connection structure, and a one-dimensional convolutional unit; inputting the counterfeit prototype features into the convolutional network to extract counterfeit category features; fusing the visual features, physical features, and counterfeit category features through the cross-modal interaction layer based on a cross-attention mechanism to generate joint features; and using the multi-task prediction layer to predict the authenticity of the joint features to obtain the identification result of the target credential.

[0178] Figure 10A This schematic diagram illustrates the structure of a multimodal authentication model for a credential authentication method according to embodiments of this application. Figure 10B The diagram illustrates the network connections of a multimodal authentication model for a credential authentication method according to embodiments of this application. For example... Figure 10A , 10BAs shown, the multimodal feature encoding of the multimodal anti-counterfeiting model adopts a multi-branch encoder structure. The visual branch processes spectral and texture fusion features and uses the feature extraction part of the sliding window transformation network to capture global dependencies. The sliding window transformation network is the backbone network of the sliding window transformation encoder. Based on the visual feature extraction backbone architecture, attention calculation is achieved through a hierarchical sliding window mechanism, balancing computational efficiency and multi-scale feature extraction capability. The physical branch processes the physical attribute features output by the neurophysical engine. The physical attribute features are input into the physical network to extract physical features. The physical network uses a multilayer perceptron (Mlp) + residual connection + 1×1 convolution (Conv) to adjust the size. The meta-learning branch inputs the imitation prototype features of the task adaptive features, adjusts the size through 1×1 convolution, and uses them as the imitation category task embedding features, i.e., imitation category features.

[0179] This attention-based cross-modal interaction layer enhances visual, physical, and imitation category features, fusing multimodal features through a cross-attention mechanism to generate joint features. For example, if the physical branch detects an abnormal paper absorption coefficient, it guides the visual branch to focus on the texture details of that area. When a new imitation type emerges, it guides the visual branch, achieving query-key-value embedding cross-attention fusion.

[0180] The joint features are input into the multi-task prediction layer to complete the tasks of binary classification of authenticity, classification of counterfeit types, regression of physical attributes and estimation of uncertainty, and output the identification results of the target certificate, including at least the authenticity probability and the uncertainty probability.

[0181] In the embodiments of this application, the multimodal authentication model extracts three types of features: visual, physical, and counterfeit category. Through a cross-attention mechanism, it achieves cross-modal interactive fusion, ultimately completing the authentication process. It comprehensively integrates key information such as spectrum, texture, physical properties, and counterfeit prototypes, achieving multi-dimensional feature complementarity. The cross-attention mechanism strengthens the association learning between different features, improves the model's comprehensive discrimination ability, adapts to complex counterfeiting scenarios, and ensures accurate and stable authentication results.

[0182] Based on the above-described method for authenticating credentials, embodiments of this application also provide a device for authenticating credentials. The following will combine... Figure 11 The device is described in detail. Figure 11 The diagram illustrates the structure of a credential authentication device according to an embodiment of this application.

[0183] like Figure 11 As shown, the credential authentication device 1000 of this embodiment includes an image acquisition and preprocessing module 1010, an attention fusion module 1020, a physical attribute reasoning module 1030, a prototype construction module 1040, and a multimodal authentication module 1050.

[0184] The image acquisition and preprocessing module 1010 can be used to acquire multi-band images of the target document and preprocess the multi-band images to obtain a multispectral image. In one embodiment, the image acquisition and preprocessing module 1010 can be used to perform step S210 described above, which will not be repeated here.

[0185] The attention fusion module 1020 can be used to process the multispectral image based on a pre-trained spectral texture attention fusion model to obtain spectral texture fusion features. In one embodiment, the attention fusion module 1020 can be used to perform step S220 described above, which will not be repeated here.

[0186] The physical attribute reasoning module 1030 can be used to perform physical attribute reasoning on the multispectral image through a pre-trained physical information neural network model to obtain physical attribute features. In one embodiment, the physical attribute reasoning module 1030 can be used to execute step S230 described above, which will not be repeated here.

[0187] The prototype construction module 1040 can be used to construct a replica prototype from the multispectral image based on a pre-trained meta-learning task model, thereby obtaining replica prototype features. In one embodiment, the prototype construction module 1040 can be used to execute step S240 described above, which will not be repeated here.

[0188] The multimodal anti-counterfeiting module 1050 can be used to identify the spectral texture fusion features, the physical attribute features, and the counterfeit prototype features using a pre-trained multimodal anti-counterfeiting model, thereby obtaining the identification result of the target credential. In one embodiment, the multimodal anti-counterfeiting module 1050 can be used to execute step S250 described above, which will not be repeated here.

[0189] According to an embodiment of this application, the meta-learning task model includes a decoupling layer, a feature extraction layer, and a prototype calculation layer; the prototype construction module 1040 can also be used to perform convolution operations and gradient inversion on the multispectral image based on the decoupling layer to obtain decoupling features; based on the feature extraction layer, perform partial convolution residual processing on the decoupling features to obtain microscopic physical features; based on a skip connection method, add the decoupling features and the microscopic physical features to obtain microscopic fusion features; and based on the prototype calculation layer, use an embedding function to calculate the imitation prototype features corresponding to the microscopic fusion features.

[0190] According to an embodiment of this application, the credential authentication device 1000 further includes a meta-learning training module, which can be used to construct a task pool based on a learning task of the imitation type, wherein the learning task includes a support set and a query set; sample a target task based on the task pool, sample the support set of the target task, and calculate a loss function value based on the initialization parameters of the meta-learning task model and the sampled support set, and perform an inner loop gradient update based on the loss function value to obtain the inner loop update parameters of the target task; calculate a meta-loss based on the query set of the target task and the inner loop update parameters, and perform an outer loop backpropagation update based on the meta-loss to update the initialization parameters; dynamically adjust the weighting of the global parameters after the outer loop update according to the task characteristics to obtain task adaptive weighting parameters; and in response to updating the learning task of the imitation type, obtain update samples according to the updated learning task, and perform model parameter gradient update based on the task adaptive weighting parameters according to the update samples.

[0191] According to embodiments of this application, the spectral texture attention fusion model includes a spectral unmixing network, a spatial texture feature extraction network, a spectral feature extraction network, and an attention fusion network. The attention fusion module 1020 can also be used to extract spectral and spatial information of the multispectral image through the spectral unmixing network, concatenate the spectral information of the multispectral image to obtain spectral features, concatenate the spatial information of the multispectral image to obtain spatial features, extract high- and low-frequency features from the spatial features based on the spatial texture feature extraction network to obtain enhanced texture features, perform multi-scale feature fusion on the spectral features based on the spectral feature extraction network to obtain enhanced spectral features, and fuse the enhanced texture features and the enhanced spectral features based on the attention fusion network to generate the spectral texture fusion feature.

[0192] According to embodiments of this application, the attention fusion module 1020 can also be used to perform wavelet transform on the spatial features to obtain low-frequency components and high-frequency components; perform Fourier transform on the low-frequency components, perform convolution and double pooling on the transformed low-frequency components to obtain low-frequency fusion components, and perform inverse Fourier transform on the low-frequency fusion components to obtain enhanced low-frequency features; concatenate the high-frequency components, perform convolution adjustment on the concatenated high-frequency components, and add the adjusted high-frequency components to the spatial features to obtain spatial high-frequency features; perform jump fusion on the spatial high-frequency features through a U-shaped network to obtain high-frequency fusion features, and perform wavelet transform on the high-frequency fusion features to obtain enhanced high-frequency features; and perform inverse wavelet transform on the enhanced high-frequency features and the enhanced low-frequency features to obtain the enhanced texture features.

[0193] According to an embodiment of this application, the attention fusion module 1020 can also be used to perform convolution processing on the spectral features to extract local spectral features; perform global average pooling processing on the local spectral features to obtain a global context vector; perform difference weight activation on the global context vector and the local spectral features based on weight parameters to obtain activated features; segment the activated features based on channel dimensions to obtain multiple sets of spectral features; perform multi-scale feature mapping processing on the multiple sets of spectral features according to preset multiple strides to obtain multi-scale features; concatenate and fuse the multi-scale features to obtain multi-scale fused features; perform convolution and activation processing on the multi-scale fused features to obtain a first activated feature; perform convolution and activation processing on the activated features to obtain a second activated feature; multiply the first activated feature and the second activated feature pixel by pixel to obtain channel attention features, and perform convolution processing on the channel attention features to obtain the enhanced spectral features.

[0194] According to an embodiment of this application, the attention fusion module 1020 can also be used to perform self-attention extraction on the enhanced texture features to obtain texture self-attention features; perform self-attention extraction on the enhanced spectral features to obtain spectral self-attention features; perform cross-attention extraction on the enhanced spectral features and the enhanced texture features to obtain cross-attention features; concatenate the cross-attention features, the texture self-attention features and the spectral self-attention features, and perform multilayer perception and convolution processing on the concatenated attention features to obtain enhanced attention features; and fuse the enhanced attention features, the enhanced spectral features and the enhanced texture features to generate the spectral texture fusion features.

[0195] According to an embodiment of this application, the physical information neural network model includes a convolutional layer, a residual module, and a multi-task dense prediction module; the physical attribute inference module 1030 can also be used to extract features from the multispectral image based on the convolutional layer to obtain shallow features; process the shallow features through the residual module to obtain the deep features; and perform physical attribute prediction based on the deep features through the multi-task dense prediction module to obtain the physical attribute features.

[0196] According to an embodiment of this application, the loss function of the physical information neural network model includes data loss of physical attributes and physical consistency loss; wherein, the physical attributes include paper optical attributes, ink optical attributes, printing pressure characteristics, and fluorescence attenuation characteristics; the physical consistency loss includes energy conservation constraints, wavelength constraints, and optical transmission theory constraints.

[0197] According to embodiments of this application, the multimodal anti-counterfeiting model includes a sliding window transformation network, a physical network, a convolutional network, a cross-modal interaction layer, and a multi-task prediction layer. The multimodal anti-counterfeiting module 1050 can also be used to input the spectral texture fusion features into the sliding window transformation network to extract visual features; input the physical attribute features into the physical network to extract physical features; wherein the physical network includes a multilayer perceptron, a residual connection structure, and a one-dimensional convolutional unit; input the counterfeit prototype features into the convolutional network to extract counterfeit category features; through the cross-modal interaction layer, based on a cross-attention mechanism, fuse the visual features, the physical features, and the counterfeit category features to generate joint features; and through the multi-task prediction layer, perform authenticity prediction on the joint features to obtain the identification result of the target credential.

[0198] According to an embodiment of this application, the image acquisition and preprocessing module 1010 can also be used to perform image registration on the multi-band image; based on the band information of the multi-band image, perform illumination unevenness correction on the registered multi-band image to obtain a multi-band corrected image; input the multi-band corrected image into a pre-trained image reconstruction model of the corresponding band for super-resolution reconstruction to obtain a multi-band reconstructed image; and perform image enhancement on the multi-band reconstructed image to obtain the multispectral image.

[0199] According to an embodiment of this application, the image reconstruction model includes a shallow feature extraction layer, a deep feature extraction layer, and a reconstruction layer. The image acquisition and preprocessing module 1010 can also be used to extract shallow features from the corrected image of the current band based on the shallow feature extraction layer of the current band to obtain target shallow features; and to perform differential modulation and residual fusion on the target shallow features based on the deep feature extraction layer of the current band to generate target deep features. The deep feature extraction layer includes a local feature enhancement network, a differential feature modulation network, a multi-scale feature aggregation network, a channel attention enhancement network, a residual feature extraction network, and a residual fusion network. Finally, the target deep features are convolutionally reconstructed through the reconstruction layer of the current band to obtain the reconstructed image of the current band.

[0200] According to an embodiment of this application, the credential authentication device 1000 further includes: a risk warning module, which can be used to call a real-time authentication interface to obtain the authentication result of the target credential, and generate an abnormal area heat map and physical attribute report based on the authentication result; and calculate the transaction risk level based on the authentication result and transaction information, and perform a risk warning operation when the transaction risk level is higher than a preset risk threshold.

[0201] According to an embodiment of this application, the credential authentication device 1000 further includes: a model update module, which can be used to respond to a user modifying the authentication result of the target credential, obtain the corrected authentication result, and store it in a false alarm database; and to perform incremental training of the credential authentication model using continuous learning technology based on the false alarm database and a timed training strategy; wherein the credential authentication model includes one or more of the following: the spectral texture attention fusion model, the physical information neural network model, the meta-learning task model, and the multimodal authentication big model.

[0202] According to embodiments of this application, any and multiple modules among the image acquisition and preprocessing module 1010, attention fusion module 1020, physical attribute reasoning module 1030, prototype construction module 1040, multimodal fraud detection module 1050, meta-learning training module, risk warning module, and model update module can be merged into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this application, at least one of the image acquisition and preprocessing module 1010, attention fusion module 1020, physical attribute reasoning module 1030, prototype construction module 1040, multimodal fraud detection module 1050, meta-learning training module, risk warning module, and model update module can be at least partially implemented as hardware circuits, such as field-programmable gate arrays, programmable logic arrays, systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits, or any other reasonable means of integrating or packaging circuits, or implemented in software, hardware, or firmware, or in any appropriate combination of any of these three implementation methods. Alternatively, at least one of the image acquisition and preprocessing module 1010, attention fusion module 1020, physical attribute reasoning module 1030, prototype construction module 1040, multimodal fraud detection module 1050, meta-learning training module, risk warning module, and model update module can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0203] Figure 12 A block diagram schematically illustrates an electronic device suitable for implementing a credential authentication method according to an embodiment of this application.

[0204] like Figure 12As shown, an electronic device 1200 according to an embodiment of this application includes a processor 1201, which can perform various appropriate actions and processes according to a program stored in a read-only memory 1202 or a program loaded from a storage portion 1208 into a random access memory 1203. The processor 1201 may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a dedicated microprocessor. The processor 1201 may also include onboard memory for caching purposes. The processor 1201 may include a single processing unit or multiple processing units for executing different steps of the method flow according to an embodiment of this application.

[0205] Random access memory 1203 stores various programs and data required for the operation of electronic device 1200. Processor 1201, read-only memory 1202, and random access memory 1203 are interconnected via bus 1204. Processor 1201 executes various steps of the method flow according to embodiments of this application by executing programs in read-only memory 1202 and / or random access memory 1203. It should be noted that the programs may also be stored in one or more memories other than read-only memory 1202 and random access memory 1203. Processor 1201 may also execute various steps of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0206] According to embodiments of this application, the electronic device 1200 may further include an input / output interface 1205, which is also connected to the bus 1204. The electronic device 1200 may also include one or more of the following components connected to the input / output interface 1205: an input section 1206 including a keyboard, mouse, etc.; an output section 1207 including a cathode ray tube, liquid crystal display, etc., and a speaker, etc.; a storage section 1208 including a hard disk, etc.; and a communication section 1209 including a network interface card, such as a local area network card, modem, etc. The communication section 1209 performs communication processing via a network such as the Internet. A drive 1210 is also connected to the input / output interface 1205 as needed. A removable medium 1211, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 1210 as needed so that computer programs read from it can be installed into the storage section 1208 as needed.

[0207] Embodiments of this application also provide a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0208] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination thereof. In embodiments of this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include the read-only memory 1202, and / or random access memory 1203, and / or one or more memories other than read-only memory 1202 and random access memory 1203 described above.

[0209] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code is used to cause the computer system to implement the methods provided in the embodiments of this application.

[0210] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 1209, and / or installed from the removable medium 1211. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0211] In embodiments of this application, the computer program can be downloaded and installed from a network via communication section 1209, and / or installed from removable medium 1211. When the computer program is executed by processor 1201, it performs the functions defined in the system of this application embodiment. According to embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0212] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0213] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0214] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

Claims

1. A method for authenticating credentials, characterized in that, The method includes: Acquire multi-band images of the target document and preprocess the multi-band images to obtain a multispectral image; The multispectral image is processed based on a pre-trained spectral texture attention fusion model to obtain spectral texture fusion features; The physical property features are obtained by inferring the physical properties of the multispectral image through a pre-trained physical information neural network model. Based on a pre-trained meta-learning task model, a replica prototype is constructed from the multispectral image to obtain replica prototype features; and By using a pre-trained multimodal anti-counterfeiting model, the spectral texture fusion features, physical attribute features, and counterfeit prototype features are processed for identification to obtain the identification result of the target credential.

2. The method according to claim 1, characterized in that, The meta-learning task model includes a decoupling layer, a feature extraction layer, and a prototype computation layer; the pre-trained meta-learning task model constructs a prototype from the multispectral image to obtain prototype features, including: Based on the decoupling layer, convolution and gradient inversion are performed on the multispectral image to obtain decoupling features. Based on the feature extraction layer, the decoupled features are partially processed by convolutional residuals to obtain microscopic physical features; Based on the skip connection method, the decoupling feature is added to the microscopic physical feature to obtain the microscopic fusion feature; and Based on the prototype calculation layer, the imitation prototype features corresponding to the micro-fusion features are calculated using the embedding function.

3. The method according to claim 1, characterized in that, The training process of the meta-learning task model includes: Based on the learning tasks of imitation type, a task pool is constructed, wherein the learning tasks include a support set and a query set; Based on the target task sampled from the task pool, the support set of the target task is sampled, and based on the initialization parameters of the meta-learning task model, the loss function value is calculated according to the sampled support set. Based on the loss function value, the inner loop gradient update is performed to obtain the inner loop update parameters of the target task. Based on the query set of the target task and the inner loop update parameters, the meta-loss is calculated, and the outer loop backpropagation update is performed based on the meta-loss to update the initialization parameters. Based on task characteristics, the weighted values ​​of the globally updated parameters are dynamically adjusted in the outer loop to obtain adaptive weighted parameters for the task; and In response to updating the learning task of the imitation type, an updated sample is obtained according to the updated learning task, and the model parameters are updated based on the task adaptive weighted parameters according to the updated sample.

4. The method according to claim 1, characterized in that, The spectral texture attention fusion model includes a spectral unmixing network, a spatial texture feature extraction network, a spectral feature extraction network, and an attention fusion network. The pre-trained spectral texture attention fusion model processes the multispectral image to obtain spectral texture fusion features, including: The spectral and spatial information of the multispectral image are extracted by the spectral unmixing network, the spectral information of the multispectral image is stitched together to obtain spectral features, and the spatial information of the multispectral image is stitched together to obtain spatial features. Based on the spatial texture feature extraction network, high and low frequency features are extracted from the spatial features to obtain enhanced texture features; Based on the spectral feature extraction network, multi-scale feature fusion is performed on the spectral features to obtain enhanced spectral features; and The enhanced texture features and the enhanced spectral features are fused based on the attention fusion network to generate the spectral texture fusion features.

5. The method according to claim 4, characterized in that, The process of extracting high- and low-frequency features from the spatial features using the spatial texture feature extraction network to obtain enhanced texture features includes: The spatial features are subjected to wavelet transform to obtain low-frequency and high-frequency components. The low-frequency components are subjected to Fourier transform, and the transformed low-frequency components are subjected to convolution and double pooling to obtain low-frequency fusion components. The low-frequency fusion components are then subjected to inverse Fourier transform to obtain enhanced low-frequency features. The high-frequency components are spliced ​​together, the spliced ​​high-frequency components are convolved and adjusted, and the adjusted high-frequency components are added to the spatial features to obtain the spatial high-frequency features. The spatial high-frequency features are fused using a U-shaped network to obtain high-frequency fused features, and then enhanced high-frequency features are obtained by wavelet transform of the high-frequency fused features; and The enhanced high-frequency features and the enhanced low-frequency features are subjected to inverse wavelet transform to obtain the enhanced texture features.

6. The method according to claim 4, characterized in that, The step of performing multi-scale feature fusion on the spectral features based on the spectral feature extraction network to obtain enhanced spectral features includes: The spectral features are convolved to extract local spectral features; the local spectral features are then subjected to global average pooling to obtain a global context vector; based on weight parameters, the global context vector and the local spectral features are activated by difference weights to obtain the activated features. Based on the channel dimension, the activated features are segmented to obtain multiple sets of spectral features; according to preset multiple step sizes, multi-scale feature mapping processing is performed on the multiple sets of spectral features to obtain multi-scale features; the multi-scale features are then concatenated and fused to obtain multi-scale fused features; and The multi-scale fusion features are convolved and activated to obtain a first activation feature; the activated features are convolved and activated to obtain a second activation feature; the first activation feature and the second activation feature are multiplied pixel by pixel to obtain a channel attention feature, and the channel attention feature is convolved to obtain the enhanced spectral feature.

7. The method according to claim 4, characterized in that, The process of fusing the enhanced texture features and the enhanced spectral features based on the attention fusion network to generate the spectral texture fusion features includes: Self-attention extraction is performed on the enhanced texture features to obtain texture self-attention features; The enhanced spectral features are subjected to self-attention extraction to obtain spectral self-attention features; Cross-attention extraction is performed on the enhanced spectral features and the enhanced texture features to obtain cross-attention features; The cross-attention features, texture self-attention features, and spectral self-attention features are concatenated, and the concatenated attention features are then subjected to multi-layer perceptron and convolution processing to obtain enhanced attention features; and The enhanced attention feature, the enhanced spectral feature, and the enhanced texture feature are fused to generate the spectral texture fusion feature.

8. The method according to claim 1, characterized in that, The physical information neural network model includes convolutional layers, residual modules, and a multi-task dense prediction module; the physical attribute inference of the multispectral image using the pre-trained physical information neural network model to obtain physical attribute features includes: Based on the convolutional layer, feature extraction is performed on the multispectral image to obtain shallow features; The shallow features are processed by the residual module to obtain the deep features; and The physical attribute features are obtained by using the multi-task intensive prediction module to predict physical attributes based on the deep features.

9. The method according to claim 1, characterized in that, The loss function of the physical information neural network model includes data loss of physical attributes and physical consistency loss; The physical properties include paper optical properties, ink optical properties, printing pressure characteristics, and fluorescence attenuation characteristics; the physical consistency loss includes energy conservation constraints, wavelength constraints, and optical transmission theory constraints.

10. The method according to claim 1, characterized in that, The multimodal anti-counterfeiting model includes a sliding window transformation network, a physical network, a convolutional network, a cross-modal interaction layer, and a multi-task prediction layer. The pre-trained multimodal anti-counterfeiting model performs identification processing on the spectral texture fusion features, the physical attribute features, and the counterfeit prototype features to obtain the identification result of the target credential, including: The spectral texture fusion features are input into the sliding window transformation network to extract visual features; The physical attribute features are input into the physical network to extract physical features; wherein, the physical network includes a multilayer perceptron, a residual connection structure, and a one-dimensional convolutional unit; The features of the counterfeit prototype are input into the convolutional network to extract the counterfeit category features; Through the cross-modal interaction layer, based on the cross-attention mechanism, the visual features, the physical features, and the imitation category features are fused using multimodal features to generate joint features; and The multi-task prediction layer is used to predict the authenticity of the joint features, thereby obtaining the identification result of the target credential.

11. The method according to claim 1, characterized in that, The preprocessing of the multi-band image includes: Image registration is performed on the multi-band images; Based on the band information of the multi-band image, the registered multi-band image is corrected for uneven illumination to obtain a multi-band corrected image. The multi-band corrected image is input into a pre-trained image reconstruction model for the corresponding band for super-resolution reconstruction, resulting in a multi-band reconstructed image; and Image enhancement is performed on the multi-band reconstructed image to obtain the multispectral image.

12. The method according to claim 11, characterized in that, The image reconstruction model includes a shallow feature extraction layer, a deep feature extraction layer, and a reconstruction layer. The step of inputting the multi-band corrected image into the pre-trained image reconstruction model for the corresponding band for super-resolution reconstruction includes: Based on the shallow feature extraction layer of the current band, shallow feature extraction is performed on the corrected image of the current band to obtain the target shallow feature. Based on the deep feature extraction layer of the current band, differential modulation and residual fusion are performed on the shallow features of the target to generate deep features; wherein, the deep feature extraction layer includes a local feature enhancement network, a differential feature modulation network, a multi-scale feature aggregation network, a channel attention enhancement network, a residual feature extraction network, and a residual fusion network; and The deep features of the target are reconstructed by convolution through the reconstruction layer of the current band to obtain the reconstructed image of the current band.

13. The method according to any one of claims 1 to 12, characterized in that, The method further includes: The system invokes a real-time authentication interface to obtain the authentication result of the target credential, and generates an anomaly area heatmap and physical attribute report based on the authentication result; and Based on the identification results and transaction information, the transaction risk level is calculated, and a risk warning operation is performed if the transaction risk level is higher than a preset risk threshold.

14. The method according to any one of claims 1 to 12, characterized in that, The method further includes: In response to a user modifying the authentication result of the target credential, the system obtains the corrected authentication result and stores it in the false alarm database; and Based on the false alarm database and the timed training strategy, continuous learning technology is used to incrementally train the credential authentication model; wherein, the credential authentication model includes one or more of the following: the spectral texture attention fusion model, the physical information neural network model, the meta-learning task model, and the multimodal authentication model.

15. A certificate authentication device, characterized in that, The device includes: The image acquisition and preprocessing module is used to acquire multi-band images of the target document and preprocess the multi-band images to obtain a multispectral image. The attention fusion module is used to process the multispectral image based on a pre-trained spectral texture attention fusion model to obtain spectral texture fusion features; The physical attribute reasoning module is used to perform physical attribute reasoning on the multispectral image through a pre-trained physical information neural network model to obtain physical attribute features; A prototype construction module is used to construct a replica prototype from the multispectral image based on a pre-trained meta-learning task model, thereby obtaining replica prototype features; and The multimodal counterfeit detection module is used to identify the spectral texture fusion features, the physical attribute features, and the counterfeit prototype features through a pre-trained multimodal counterfeit detection model, and obtain the identification result of the target credential.

16. An electronic device comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 14.

17. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 14.

18. A computer program product comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 14.

Images