Offline signature discrimination system and method based on sigpconvnext model

Abstract

This application provides an offline signature authentication system and method based on the SigPConvNeXt model. The system includes a feature extraction network PConvNeXt1, a twin feature extraction network PConvNeXt2, and a classifier. PConvNeXt1 and PConvNeXt2 are used to extract feature vectors from two offline signature images to be compared. The classifier calculates the distance between the two signature images in the same feature space and then determines whether the two signatures are from the same person based on a predetermined threshold. The feature extraction networks PConvNeXt1 and PConvNeXt2 include an encoder and a decoder. The encoder includes four ConvNeXt stages and a CBAM module connected to each stage. The CBAM module includes a channel attention module and a spatial attention module. The decoder includes four decoding stages and a feature pyramid fusion module connected to each decoding stage. This application enhances the feature extraction capability by combining the ConvNeXt network with the CBAM module and introduces a feature pyramid fusion module to achieve effective fusion of features from different dimensions.

Description

Technical Field

[0001] This application relates to the field of signature authentication technology, and more specifically, to an offline signature authentication system and method based on the SigPConvNeXt model. Background Technology

[0002] Definitions of abbreviations and key terms

[0003] SigPConvNeXt:

[0004] Definition: SigPConvNeXt (Signature Pyramid ConvNeXt) is an offline signature authentication model based on an improved ConvNeXt network and Siamese network structure. It combines deep learning techniques, particularly attention mechanisms and feature pyramid fusion, to improve the accuracy and efficiency of handwritten signature authentication.

[0005] ConvNeXt:

[0006] Definition: ConvNeXt is a deep convolutional network model proposed by Facebook AI and UC Berkeley. It borrows the structure and training strategy of the Swin-Transformer, improving computational efficiency and accuracy through depthwise separable convolutions and layer normalization.

[0007] CBAM (Convolutional Block Attention Module):

[0008] Definition: CBAM is an attention mechanism module comprising a Channel Attention Module (CAM) and a Spatial Attention Module (SAM) to enhance the expressive power of feature maps in convolutional neural networks. By focusing on more important channels and spatial locations, CBAM improves the model's sensitivity to key features, thereby enhancing the performance of classification and recognition tasks.

[0009] FPN (Feature Pyramid Network):

[0010] Definition: Feature Fusion Network (FPN) is a feature fusion network structure that integrates features at different scales through top-down lateral connections to enhance the model's ability to detect objects at multiple scales. FPN combines high-resolution information from low layers with strong semantic information from high layers, improving the accuracy of tasks such as object detection.

[0011] SVM (Support Vector Machine):

[0012] Definition: Support Vector Machine (SVM) is a supervised learning model used for classification and regression analysis. It distinguishes data points from different classes by finding the optimal hyperplane. SVM is known for its effectiveness in high-dimensional spaces and with small sample sizes, especially excelling in few-shot learning problems.

[0013] Numpy:

[0014] Definition: NumPy is an open-source Python scientific computing library that provides multidimensional array objects, derived objects (such as masked arrays and matrices), and a variety of routines for fast array manipulation. NumPy is a fundamental tool for data processing and scientific computing, playing a particularly important role in array manipulation, mathematical computation, and data analysis.

[0015] Biometric authentication is an identity verification technology based on an individual's physiological or behavioral characteristics. It utilizes unique biometric features for secure access, such as device login, access control systems, and payment verification. Biometric authentication can be divided into two main categories: physiological features and behavioral features. Physiological features are related to an individual's physical attributes, such as fingerprint, face mask, iris, palm print, and voiceprint recognition, while behavioral features are related to an individual's behavioral patterns, such as typing rhythm, gait recognition, and signature authentication. Each person's biometric features are generally unique and difficult to forge, making them more secure than traditional passwords. However, collecting and storing biometric data can raise privacy and security concerns. Furthermore, some biometric authentication methods may be affected by environmental factors, posing risks of false positives and false negatives.

[0016] Signature authentication is a common biometric authentication technology that uses individual signature characteristics to verify identity. It falls under behavioral authentication and is widely used in finance, law, and contract signing. Depending on the signature acquisition method and the characteristics obtained, signature authentication can be divided into online and offline signatures. Online signatures are typically written in real-time on a device using a touchscreen or digital pen, focusing on analyzing dynamic characteristics such as signature speed, pressure, and stroke order. This is commonly used in electronic payments and mobile device unlocking. Offline signatures, on the other hand, are static images, completed on paper. They can be digitized through scanning, but with the development of information technology, they can also be directly acquired using devices such as signature pads. This method focuses on analyzing the shape, size, angle, and other static characteristics of the signature, and is commonly used in contract signing and document verification.

[0017] Offline signature authentication can be divided into two modes: writer-dependent and writer-independent. Writer-dependent systems require retraining for each new writer added, leading to significant update and storage costs for consumer-based systems where new consumers join daily. In contrast, the writer-independent mode establishes a general system to simulate the differences between genuine and forged signatures. During training, available signatures are divided into training and test sets. For a single signer, their signatures are categorized into two classes: similar (genuine signature, genuine signature) or dissimilar (genuine signature, forged signature). This process applies to all signers in both the training and test sets, building training and test instances for the classifier. Therefore, the writer-independent mode is preferable and more general than the writer-dependent mode, and this paper will focus on the writer-independent signature authentication method.

[0018] Research on offline signature authentication methods has progressed through two phases: manual feature extraction and deep learning. Early research focused on manually processing static images to extract various features (global feature extraction such as block codes, wavelets, and Fourier sequences; local geometric and topological features such as position, tangent direction, blob structure, connected components, and curvature). Methods based on projection and contours (oriented contours, wraparound features, mesh-based methods, geometric moment-based methods, and texture-based methods) and structural methods considering relationships between local features (such as image matching and the recently proposed compact correlation features) have also emerged. However, with the rapid development of neural networks and deep learning in recent years, more and more researchers are using deep learning to solve offline signature authentication problems. On the other hand, Siamese networks, consisting of two networks, accept two types of images for feature extraction. At the top of each network, a function calculates the distance between the two images in the feature space. The two networks share parameters, ensuring that similar images cannot be mapped to significantly different locations in the feature space. These characteristics make Siamese networks popular for various verification tasks such as signature verification and face verification.

[0019] A patent document entitled "Multi-scale Offline Handwritten Signature Authentication Method Based on Attention Mechanism" discloses a multi-scale offline handwritten signature authentication method based on attention mechanism, which includes the following steps:

[0020] S1: Obtain the original offline signature image: Read the original image data and convert it to NumPy format;

[0021] S2: Preprocessing based on the above signature image using Gaussian filtering and OTSU algorithm: Gaussian filtering is applied to the NumPy data to remove image noise, and then the OTSU algorithm is used to perform the maximum inter-class variance method to divide the image into background and foreground;

[0022] S3: Based on the preprocessed signature image data described above, the proposed network structure is trained: The entire neural network consists of four blocks, each using two different convolutional blocks to convolve the same image or feature map to obtain multi-scale features. A polarization self-attention mechanism is used to find important features, and then feature fusion is performed using ADD or Concat.

[0023] S4: Optimize the model based on the above pre-trained model results: By analyzing the initial and subsequent training logs and test results, adjust parameters such as learning rate, decay strategy, and weight decay to accelerate model convergence and improve accuracy.

[0024] S5: Based on the above-tuned pre-trained model, perform prediction and signature authentication: Input test signature image data, obtain predicted labels through the pre-trained model, extract classification features from the fully connected layer of the model, train the SVM model, use the pre-trained SVM to perform batch offline signature authentication, and obtain the final authentication result.

[0025] This patented solution has achieved good results in offline signature authentication, but it still has the following shortcomings:

[0026] 1. Insufficient precision in feature extraction: Existing technologies fail to fully utilize key information in images, resulting in deficiencies in feature extraction. The feature extraction is not precise enough, affecting the final identification effect.

[0027] 2. Limited generalization ability: Existing technologies perform poorly in generalization across languages ​​or different types of signatures, especially in non-training languages.

[0028] 3. Room for improvement in accuracy: There is still considerable room for improvement in the accuracy of existing technologies, especially in the identification of complex or forged signatures, where the error rate may be relatively high.

[0029] 4. Dependence on training samples: Existing technologies are highly dependent on the quantity and quality of training samples, and the problem of few-shot learning has not been effectively solved.

[0030] 5. Model complexity and overfitting issues: Existing technologies have problems in controlling model complexity and overfitting, resulting in insufficient stability and reliability of the model in practical applications. Summary of the Invention

[0031] In view of this, this application provides an offline signature authentication system and method based on the SigPConvNeXt model to overcome the above-mentioned problems existing in the prior art.

[0032] To achieve the above objectives, the technical solution adopted in this application is as follows:

[0033] The offline signature authentication system based on the SigPConvNeXt model includes a feature extraction network PConvNeXt1, a twin feature extraction network PConvNeXt2, and a classifier. PConvNeXt1 and PConvNeXt2 are used to extract feature vectors from two offline signature images to be compared. The classifier calculates the distance between the two offline signature images in the same feature space and then determines whether the two signatures are from the same person based on a predetermined threshold. PConvNeXt1 and PConvNeXt2 have the same network architecture and share the same network parameters. The feature extraction networks PConvNeXt1 and PConvNeXt2 include an encoder and a decoder. The encoder includes four ConvNeXt stages and a CBAM module connected to each stage. The CBAM module includes a channel attention module and a spatial attention module. Each stage contains multiple convolutional layers for extracting features at different levels. The decoder includes four decoding stages and a feature pyramid fusion module connected to each decoding stage.

[0034] Furthermore, each decoding stage consists of a 1x1 convolutional layer, a 3x3 convolutional layer, and an upsampling module. The 1x1 convolutional layer is used to adjust the number of channels in the feature map, the 3x3 convolutional layer is used to extract features, and the upsampling is used to restore the spatial resolution of the feature map.

[0035] Furthermore, the feature pyramid fusion module includes a feature fusion module and a downsampling module. The feature fusion module fuses feature maps of different resolutions through a feature fusion operation. The downsampling module performs a downsampling operation after feature fusion to reduce the size of the feature maps.

[0036] This application also provides an offline signature authentication method based on the SigPConvNeXt model, the method being based on the offline signature authentication system and including:

[0037] Step 1: Input Processing: Input the two offline signature images to be compared into the feature extraction networks PConvNeXt1 and PConvNeXt2 of the offline signature authentication system, respectively;

[0038] Step 2: Feature extraction: Deep features of the offline signature image are extracted step by step through the four stages of the encoder, and the discriminative ability of the features is enhanced by the CBAM module after each stage.

[0039] Step 3: Feature Fusion: In the decoder, features at different scales are integrated through the feature pyramid fusion module to enhance the model's ability to capture features of different sizes and orientations;

[0040] Step 4: Output Generation: After feature fusion, a downsampling operation is performed to reduce the size of the feature map and generate a feature vector for signature authentication.

[0041] Step 5: Identification Task: Using the generated feature vectors, calculate the distance between the two offline signature images in the same feature space through a classifier, and then determine whether the two signatures are from the same person based on a predetermined threshold.

[0042] Furthermore, in step 5, calculating the distance between two offline signature images in the same feature space using a classifier specifically involves using Euclidean distance to calculate the distance between the two offline signature images in the same feature space. The specific formula is as follows: ;

[0043] in, and These are the feature vectors of the two signature images, respectively. The dimension of the feature space. Let be the index, representing the th element in the feature space. One characteristic.

[0044] Furthermore, the method for determining the threshold is an automatic threshold determination method, specifically including:

[0045] 1) After each round of model training, segment values ​​are taken from the mean distance between true-false signature pairs and true-true signature pairs to establish... One test value: ;

[0046] in The mean distance between true-true signature pairs. This represents the mean distance between true and false signature pairs;

[0047] 2) Use each test value Perform signature verification and calculate the current model's verification accuracy for each test value. Take the test value corresponding to the highest accuracy value as the model threshold.

[0048] Compared with the prior art, the beneficial effects of this application are:

[0049] 1. Improved Refinement of Feature Extraction: By combining the ConvNeXt network and the CBAM module to form the PConvNeXt network, this application can enhance the representativeness of features from both channel and spatial dimensions. The channel attention (CAM) and spatial attention (SAM) of the CBAM module can adaptively emphasize important features and key regions, thereby improving the refinement of feature extraction. This refined feature extraction helps to more accurately capture subtle differences in handwritten signatures and improve the accuracy of authenticity verification.

[0050] 2. Enhanced Model Generalization Ability: This application employs a Siamese network structure and trains and cross-validates on multiple datasets. This helps the model learn more general feature representations, thereby enhancing its generalization ability across different datasets and language environments. This improved generalization ability means that the identification method in this application is not only applicable to specific datasets but can also adapt to a wider range of application scenarios.

[0051] 3. Improved processing speed and efficiency: The improved ConvNeXt network structure and Feature Pyramid Fusion (PFF) module can efficiently process and fuse features at different scales, reducing computational resource consumption and improving processing speed and efficiency. This enables this application to respond quickly in practical applications and meet real-time or near-real-time identification requirements.

[0052] 4. Improved accuracy and reduced error rate: By fusing multi-scale features and an attention mechanism, this application can more comprehensively capture the features of handwritten signatures, thereby improving the accuracy of identification. Simultaneously, the automatic threshold determination method can dynamically adjust the optimal threshold based on the model training results, further reducing the error rate and improving the reliability of identification.

[0053] 5. Reduced Model Complexity and Overfitting: The PConvNeXt network in this application effectively balances model complexity and performance by combining CBAM and PFF modules, reducing the risk of overfitting. This structure not only improves model stability but also ensures good performance on unknown data.

[0054] 6. Promoting the development of information security in the field of biometric identification: This application improves the accuracy and efficiency of offline signature authentication, reduces the subjectivity of manual authentication, and enhances the fairness and objectivity of offline signature authentication. This has significant practical application value for the field of information security, especially in areas with high security requirements such as finance, law, and contract signing. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 This is a network structure diagram of the feature extraction network PConvNeXt in this application;

[0057] Figure 2 This is a structural diagram of the offline signature authentication system based on the SigPConvNeXt model in this application;

[0058] Figure 3 This is a flowchart of the offline signature authentication method based on the SigPConvNeXt model in this application. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of this application, but not all embodiments.

[0060] ConvNeXt is a deep convolutional network model proposed by Facebook AI and UC Berkeley. In recent years, deep learning models based on transformers have gradually become mainstream. The Swin-Transformer, which appeared in 2021, even won the best paper at ICCV that year. ConvNeXt is an improved model based on ResNet, which borrows the structure and training strategy of Swin-Transformer. ConvNeXt references the scaling strategy of Swin-Transformer and adjusts the four blocks of ResNet-50, changing the stacking number in ResNet-50 from the original (3, 4, 6, 3) to (3, 3, 9, 3). ConvNeXt borrows the advantages of depthwise separable convolution, replacing the 3×3 convolution in the bottleneck layer with depthwise separable convolution, while increasing the width of the network, thereby improving computational efficiency and accuracy.

[0061] ConvNeXt is mainly composed of four stacked stages. Each stage contains a downsampling module and multiple ConvNeXt Blocks. The initial layer uses a 4×4 convolutional layer with a stride of 4 for partitioning, and layer normalization is used to improve the training stability and convergence speed of the neural network. Subsequently, each stage extracts features by downsampling and stacking ConvNeXt Blocks at different numbers of times. The downsampling part consists of layer normalization and convolutional kernels with a size of 2×2 and a stride of 2. Finally, global average pooling, layer normalization, and fully connected layers are used to obtain the result. There are four versions of ConvNeXt (T, S, B, L). The specific implementation of this application uses the ConvNeXt-B version, and its network structure is shown in Table 1.

[0062] Table 1 ConvNeXt-B Network Structure

[0063]

[0064] The Convolutional Block Attention Module (CBAM) comprises two independent sub-modules: the Channel Attention Module (CAM) and the Spartial Attention Module (SAM), employing attention mechanisms in the channel and spatial dimensions, respectively. This not only saves parameters and computational power but also ensures its plug-and-play integration into existing network architectures.

[0065] Feature Pyramid Networks (FPNs) are implemented through top-down lateral connections. Lower-level features contain rich localization information, but this information may be lost when passed to higher levels, making it difficult to obtain accurate localization information. This application employs a Bidirectional Feature Fusion Module (BiFFM). This module adds a bottom-up path to the end of the FPN, enhancing the accurate localization information from the lower levels into the entire feature hierarchy through bottom-up path enhancement.

[0066] like Figure 1 and 2As shown, this application provides an offline signature authentication system based on the SigPConvNeXt model, including a feature extraction network PConvNeXt1, a twin feature extraction network PConvNeXt2 of PConvNeXt1, and a classifier. PConvNeXt1 and PConvNeXt2 are used to extract feature vectors from two offline signature images to be compared. The classifier calculates the distance between the two offline signature images in the same feature space and then determines whether the two signatures are from the same person based on a predetermined threshold. PConvNeXt1 and PConvNeXt2 have the same network architecture and share the same network parameters. The feature extraction networks PConvNeXt1 and PConvNeXt2 include an encoder and a decoder. The encoder includes four stages and a CBAM module connected to each stage. The CBAM module includes a channel attention module and a spatial attention module. Each stage contains multiple convolutional layers for extracting features at different levels. The decoder includes four decoding stages and a feature pyramid fusion module connected to each decoding stage.

[0067] The feature extraction network PConvNeXt (Pyramid ConvNeXt) in this application adds a CBAM module after each stage of the ConvNeXt model, and uses a pyramid bidirectional feature fusion module to fuse the feature maps processed by each CBAM module to obtain the final feature map. The specific structure is as follows: Figure 1 As shown, a CBAM module is applied after each stage to enhance the expressive power of features. Through channel attention and spatial attention mechanisms, the model is able to focus on more important regions in the image.

[0068] The entire SigPConvNeXt model, such as Figure 2 As shown, it consists of two feature extraction networks PConvNeXt (PConvNeXt1 and PConvNeXt2). The two offline signature images to be compared are input into the trained feature extraction network to obtain two feature vectors, which are then passed to the classifier to calculate the distance between the two images in the same feature space. Finally, based on a predetermined threshold, the classifier judges whether the two signatures are from the same person.

[0069] By combining the ConvNeXt network with the CBAM (Convolutional Block Attention Module), the feature extraction capability is enhanced. The Feature Pyramid Fusion (PFF) module is introduced to achieve effective fusion of features from different dimensions.

[0070] As a further implementation, each decoding stage consists of a 1x1 convolutional layer, a 3x3 convolutional layer, and an upsampling module. The 1x1 convolutional layer is used to adjust the number of channels in the feature map, the 3x3 convolutional layer is used to extract features, and the upsampling is used to restore the spatial resolution of the feature map.

[0071] As a further implementation, the feature pyramid fusion module includes a feature fusion module and a downsampling module. The feature fusion module fuses feature maps of different resolutions through a feature fusion operation. The downsampling module is used to perform a downsampling operation after feature fusion to reduce the size of the feature maps.

[0072] After each decoding stage, feature maps of different resolutions are fused through feature fusion (P1, P2, P3, P4) operations. After feature fusion, downsampling is performed to reduce the size of the feature maps while retaining important features.

[0073] like Figure 3 As shown, this application also provides an offline signature authentication method based on the SigPConvNeXt model. The method is based on the offline signature authentication system and includes:

[0074] Step 1: Input Processing: Input the two offline signature images to be compared into the feature extraction networks PConvNeXt1 and PConvNeXt2 of the offline signature authentication system, respectively;

[0075] Step 2: Feature extraction: Deep features of the offline signature image are extracted step by step through the four stages of the encoder, and the discriminative ability of the features is enhanced by the CBAM module after each stage.

[0076] Step 3: Feature Fusion: In the decoder, features at different scales are integrated through the feature pyramid fusion module to enhance the model's ability to capture features of different sizes and orientations;

[0077] Step 4: Output Generation: After feature fusion, a downsampling operation is performed to reduce the size of the feature map and generate a feature vector for signature authentication.

[0078] Step 5: Identification Task: Using the generated feature vectors, calculate the distance between the two offline signature images in the same feature space through a classifier, and then determine whether the two signatures are from the same person based on a predetermined threshold.

[0079] As a further implementation, in step 5, calculating the distance between the two offline signature images in the same feature space using a classifier specifically involves using Euclidean distance to calculate the distance between the two offline signature images in the same feature space. The specific formula is as follows: ;

[0080] in, and These are the feature vectors of the two signature images, respectively. The dimension of the feature space. Let be the index, representing the th element in the feature space. One characteristic.

[0081] As a further implementation, the method for determining the threshold is an automatic threshold determination method, specifically including:

[0082] 1) After each round of model training, segment values ​​are taken from the mean distance between true-false signature pairs and true-true signature pairs to establish... One test value: ;

[0083] in The mean distance between true-true signature pairs. This represents the mean distance between true and false signature pairs;

[0084] 6) Use each test value Perform signature verification and calculate the current modulus for each verification value.

[0085] The model threshold is determined by the accuracy of the type identification and the test value corresponding to the highest accuracy value.

[0086] During model training, Generative Adversarial Networks (GANs) can be used to generate forged signature samples, increasing the number of samples and enhancing the model's generalization ability. Simultaneously, the samples generated by GANs can be used to train and test the model, improving its ability to identify forged signatures.

[0087] By automatically determining the threshold, manual intervention is reduced, and the automation level of the identification process is improved. This automated threshold determination method allows the application to adapt to different identification environments and conditions, improving operational convenience and result consistency.

[0088] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An offline signature authentication system based on the SigPConvNeXt model, characterized in that, The system includes a feature extraction network PConvNeXt1, a twin feature extraction network PConvNeXt2, and a classifier. PConvNeXt1 and PConvNeXt2 are used to extract feature vectors from two offline signature images to be compared. The classifier calculates the distance between the two offline signature images in the same feature space and then determines whether the two signatures are from the same person based on a predetermined threshold. PConvNeXt1 and PConvNeXt2 have the same network architecture and share the same network parameters. The feature extraction networks PConvNeXt1 and PConvNeXt2 include an encoder and a decoder. The encoder includes four ConvNeXt stages and a CBAM module connected to each stage. The CBAM module includes a channel attention module and a spatial attention module. Each stage contains multiple convolutional layers for extracting features at different levels. The decoder includes four decoding stages and a feature pyramid fusion module connected to each decoding stage.

2. The offline signature authentication system based on the SigPConvNeXt model as described in claim 1, characterized in that, Each decoding stage consists of a 1x1 convolutional layer, a 3x3 convolutional layer, and an upsampling module. The 1x1 convolutional layer is used to adjust the number of channels in the feature map, the 3x3 convolutional layer is used to extract features, and the upsampling is used to restore the spatial resolution of the feature map.

3. The offline signature authentication system based on the SigPConvNeXt model as described in claim 2, characterized in that, The feature pyramid fusion module includes a feature fusion module and a downsampling module. The feature fusion module fuses feature maps of different resolutions through feature fusion operations. The downsampling module performs downsampling operations after feature fusion to reduce the size of the feature maps.

4. An offline signature authentication method based on the SigPConvNeXt model, characterized in that, The method is based on the offline signature authentication system according to any one of claims 1-3, and the method includes: Step 1: Input Processing: Input the two offline signature images to be compared into the feature extraction networks PConvNeXt1 and PConvNeXt2 of the offline signature authentication system, respectively; Step 2: Feature extraction: Deep features of the offline signature image are extracted step by step through the four stages of the encoder, and the discriminative ability of the features is enhanced by the CBAM module after each stage. Step 3: Feature Fusion: In the decoder, features at different scales are integrated through the feature pyramid fusion module to enhance the model's ability to capture features of different sizes and orientations; Step 4: Output Generation: After feature fusion, a downsampling operation is performed to reduce the size of the feature map and generate a feature vector for signature authentication. Step 5: Identification Task: Using the generated feature vectors, calculate the distance between the two offline signature images in the same feature space through a classifier, and then determine whether the two signatures are from the same person based on a predetermined threshold.

5. The offline signature authentication method based on the SigPConvNeXt model as described in claim 4, characterized in that, In step 5, the distance between the two offline signature images in the same feature space is calculated using a classifier. Specifically, Euclidean distance is used to calculate the distance between the two offline signature images in the same feature space. The specific formula is as follows: ； in, and These are the feature vectors of the two signature images, respectively.

6. The offline signature authentication method based on the SigPConvNeXt model as described in claim 5, characterized in that, The method for determining the threshold is an automatic threshold determination method, specifically including: 1) After each round of model training, segment values ​​are taken from the mean distance between true-false signature pairs and true-true signature pairs to establish... One test value: ; in The mean distance between true-true signature pairs. This represents the mean distance between true and false signature pairs; 2) Use each test value Perform signature verification and calculate the current model's verification accuracy for each test value. Take the test value corresponding to the highest accuracy value as the model threshold.

Images