Feature extraction and matching combined handwritten signature intelligent identification method and system

Abstract

The invention relates to an intelligent handwritten signature recognition method and system combining feature extraction and matching, and relates to the technical field of handwritten signature recognition, and the method comprises the steps: collecting image data of a multi-source handwritten signature, constructing a standardized signature data set, and generating a triple sample set; constructing a lightweight convolutional neural network as a feature extraction backbone network; designing a two-stage training strategy, and performing staged optimization and embedded space structural adjustment on the feature extraction backbone network; using the optimized feature extraction backbone network to collect depth feature embedding vectors of handwritten signatures in real time, and calculating similarity scores between signature pairs to form a total similarity matrix; and completing identity comparison and batch clustering of the handwritten signatures by adopting a self-adaptive threshold and combining a hierarchical clustering algorithm. The problems that traditional handwritten signature recognition depends on a fixed judgment standard, is difficult to adapt to various changes of signatures and cannot flexibly cope with complex scenes, so that recognition misjudgment is frequent, and accuracy is insufficient are solved.

Description

Technical Field

[0001] This application relates to the field of handwritten signature recognition, and in particular to an intelligent handwritten signature recognition method and system that combines feature extraction and matching. Background Technology

[0002] With the acceleration of digitalization in finance, commerce and other fields, handwritten signatures, as a core means of identity verification, have become a key technical requirement for supporting business security due to their accuracy of intelligent recognition and adaptability to various scenarios.

[0003] Currently, traditional handwritten signature recognition methods rely on fixed judgment criteria, which fail to fully adapt to the diverse changes in signature writing styles and imaging conditions. This not only easily leads to frequent misjudgments but also reduces the reliability of the results and increases the security risks of conducting business based on signature verification. Summary of the Invention

[0004] This application provides a handwritten signature intelligent recognition method and system that combines feature extraction and matching, which improves the problems of traditional handwritten signature recognition relying on fixed standards, poor adaptability, and frequent misjudgments in complex scenarios, thereby improving the accuracy and scenario adaptability of handwritten signature recognition.

[0005] The embodiments of this application disclose the following technical solutions: In a first aspect, embodiments of this application provide a method for intelligent recognition of handwritten signatures that combines feature extraction and matching, the method comprising: Collect image data of handwritten signatures from multiple sources, perform structured processing, construct a standardized signature dataset, and generate a triplet sample set based on signature identity labels; Based on the standardized signature dataset and the triplet sample set, a lightweight convolutional neural network is constructed as the feature extraction backbone network, wherein the lightweight convolutional neural network integrates a data augmentation layer and a style normalization preprocessing layer. A two-stage training strategy is designed to optimize the feature extraction backbone network in stages and adjust the embedding space structure. The optimized feature extraction backbone network is used to collect deep feature embedding vectors of several handwritten signatures in real time, and the similarity scores between signature pairs are calculated to form a full similarity matrix. Based on the full similarity matrix, an adaptive threshold based on the validation set distribution is used, combined with a hierarchical clustering algorithm, to complete the identity comparison and batch clustering of the several handwritten signatures.

[0006] Secondly, embodiments of this application provide a handwritten signature intelligent recognition system that combines feature extraction and matching, the system comprising: The signature data processing module is used to collect image data of handwritten signatures from multiple sources, perform structured processing, construct a standardized signature dataset, and generate a triplet sample set based on the signature identity label. The feature extraction network module is used to construct a lightweight convolutional neural network based on the standardized signature dataset and the triplet sample set, which serves as the backbone network for feature extraction. The lightweight convolutional neural network integrates a data augmentation layer and a style normalization preprocessing layer. The network training optimization module is used to design a two-stage training strategy to perform phased optimization and embedding space structure adjustment on the feature extraction backbone network. The similarity matrix construction module is used to collect deep feature embedding vectors of several handwritten signatures in real time using the optimized feature extraction backbone network, and calculate the similarity score between signature pairs to form a full similarity matrix. The identity comparison and clustering module is used to complete the identity comparison and batch clustering of the several handwritten signatures based on the full similarity matrix, using an adaptive threshold based on the distribution of the verification set, combined with a hierarchical clustering algorithm.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application proposes an intelligent handwritten signature recognition method and system that combines feature extraction and matching. By completing signature data processing, feature extraction network construction, network training and optimization, similarity matrix construction, and identity comparison clustering in steps, it achieves batch recognition and identity verification of handwritten signatures. First, multi-source handwritten signature image data is collected and structured through grayscale conversion, orientation correction, and pixel-level alignment to construct a standardized dataset. A triplet sample set is then generated based on identity labels. Next, a lightweight convolutional neural network integrating a data augmentation layer and a style normalization preprocessing layer is constructed as the backbone network for feature extraction. Subsequently, a two-stage training strategy combining ArcFace classification loss and triplet loss is adopted. Dynamic hard sample mining and dynamic interval adjustment are used to optimize the network and improve feature discrimination ability. Then, the optimized network is used to extract deep feature embedding vectors of the signatures to be recognized, calculate cosine similarity, and construct a full similarity matrix. Finally, an adaptive threshold is determined based on the validation set distribution, and a hierarchical clustering algorithm is used to complete batch clustering of signatures. Identity determination is achieved by matching cluster center features with an identity reference library, and a comparison list containing complete identity results is output.

[0008] The technical solution of this application solves the problems of traditional handwritten signature recognition relying on fixed standards, poor adaptability, and frequent misjudgments. It avoids recognition deviations caused by changes in handwriting and environmental interference, improves the accuracy and automation level of handwritten signature recognition, and can be applied to identity verification needs in multiple scenarios such as finance and commerce. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 A flowchart illustrating a handwritten signature intelligent recognition method combining feature extraction and matching, provided for an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a handwritten signature intelligent recognition system that combines feature extraction and matching, provided in an embodiment of this application.

[0011] The components represented by each number in the attached diagram are explained below: Signature data processing module 01, feature extraction network module 02, network training and optimization module 03, similarity matrix construction module 04, identity comparison and clustering module 05. Detailed Implementation

[0012] This application provides a handwritten signature intelligent recognition method and system that combines feature extraction and matching, which solves the technical problems in the prior art where traditional handwritten signature recognition relies on fixed judgment standards, is difficult to adapt to the diverse changes in signatures, cannot flexibly cope with complex scenarios, and leads to frequent misjudgments and insufficient accuracy.

[0013] 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 only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0015] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0016] Example 1, as shown in the appendix Figure 1 As shown, this application provides a handwritten signature intelligent recognition method that combines feature extraction and matching. The method includes the following steps: S110: Collect image data of handwritten signatures from multiple sources, perform structured processing, construct a standardized signature dataset, and generate a triplet sample set based on signature identity labels; In this embodiment of the application, in order to eliminate the interference caused by different collection conditions, a high-quality dataset needs to be constructed through multi-dimensional collection and standardized processing, and then targeted samples are generated to improve the model's discrimination ability, since handwritten signature intelligent recognition needs to be adapted to various application scenarios.

[0017] Specifically, we first extensively collect original images of handwritten signatures to ensure coverage of different writing tools, background environments, and imaging resolutions, thereby obtaining signature forms that may appear in practical applications and providing realistic basic data for subsequent model training.

[0018] Furthermore, the original images of the collected handwritten signatures are subjected to structured processing to ensure that the signatures collected under different conditions maintain the same shape and specifications, thus forming a structured signature image set.

[0019] Furthermore, the structured signature image set is divided into a training set, a validation set, and a test set according to a preset ratio. The three sets work together to form a standardized signature dataset, providing standardized data support for the phased training and performance verification of the model.

[0020] Finally, relying on the signature identity labels in the standardized signature dataset, a triplet sample set is generated by using a sampling strategy of the furthest positive sample and the nearest negative sample within the batch. Each triplet contains an anchor sample, a positive sample, and a negative sample. The above sampling method strengthens the discriminative features between samples, laying the foundation for subsequent model optimization of the embedding space structure and improvement of similarity discrimination ability.

[0021] Step S110 in the method provided in this application embodiment includes: Collect original images of handwritten signatures, including various writing tools, backgrounds, and imaging resolutions; The original handwritten signature image is sequentially subjected to grayscale processing, orientation correction processing based on the minimum bounding rectangle, and pixel-level alignment processing to be uniformly scaled to a predetermined fixed size to obtain a structured signature image set; The structured signature image set is divided into a training set, a verification set, and a test set according to a preset ratio to form the standardized signature dataset; Based on the signature identity labels in the standardized signature dataset, the triplet sample set is generated by using the sampling strategy of the furthest positive sample and the nearest negative sample within the batch. Each triplet includes an anchor sample, a positive sample, and a negative sample.

[0022] In this embodiment of the application, in order to avoid insufficient generalization ability of the model due to data differences, it is necessary to construct a high-quality dataset through multi-dimensional collection, standardized processing and targeted sample generation to improve the accuracy and scenario adaptability of intelligent handwritten signature recognition.

[0023] Specifically, the first step is to acquire original images of handwritten signatures, including various writing tools, backgrounds, and imaging resolutions. The choice of writing tools must cover common types used in practical applications, such as fountain pens, ballpoint pens, pencils, and brushes, to ensure that the model can adapt to the differences in signature features caused by different pen strokes and ink depths.

[0024] In addition, the background types should include white paper, gridded document paper, colored background paper, and official document paper with printed text, simulating the imaging environment of signatures in different usage scenarios. The imaging resolution should cover a common range from low-resolution mobile phone shooting to high-resolution scanner scanning to ensure that the data can reflect the details differences brought about by different acquisition devices.

[0025] At the same time, during the collection process, it is necessary to ensure that there are enough signature samples for each identity, including samples of regular writing, as well as samples of different states such as fast writing, slow writing, and slightly tilted writing, in order to comprehensively capture the changing patterns of individual signatures.

[0026] Furthermore, after the original image is acquired, it needs to be processed sequentially as grayscale, directional correction based on the minimum bounding rectangle, and pixel-level alignment by uniform scaling to a predetermined fixed size, in order to enhance the consistency of key structural features of the signature.

[0027] Among these methods, grayscale processing can strip away the color information of an image, simplifying the data dimensions, while preserving the stroke outlines and structural features of the signature, thus avoiding interference from color information in feature extraction. The orientation correction processing based on the minimum bounding rectangle addresses the potential tilting issue during signature writing. By locating the minimum bounding rectangle of the signature, calculating the tilt angle, and performing rotation correction, the writing direction of all signatures remains consistent.

[0028] In addition, the process of uniformly scaling to a predetermined fixed size requires determining a reasonable size specification, such as 128×256 pixels. This size can not only fully preserve the detailed features of the signature, but also control the amount of data to improve the model training efficiency. During the scaling process, a proportional scaling combined with boundary padding is used to avoid the signature structure from being deformed due to stretching or compression, and finally a structured signature image set with uniform specifications is formed.

[0029] Furthermore, the structured signature image set is divided into a training set, a validation set, and a test set according to a preset ratio to form a standardized signature dataset. The preset ratio must balance training effectiveness and evaluation accuracy; for example, an 8:1:1 ratio can be used. The training set is used for model parameter learning, the validation set is used for hyperparameter tuning and model performance monitoring during training, and the test set is used to independently evaluate the model's final recognition performance.

[0030] Simultaneously, during the partitioning process, it is crucial to ensure that signature samples of the same identity do not appear simultaneously in both the training and test sets to avoid data leakage leading to distorted test results. Furthermore, it is essential to maintain a balanced distribution of samples from different identities across each set, preventing excessive concentration of samples from any one identity in a particular set, and ensuring that the model can fairly learn the features of each identity.

[0031] Finally, based on the signature identity labels in the standardized signature dataset, a triplet sample set is generated using the sampling strategy of the furthest positive sample and the nearest negative sample within the batch. Each triplet consists of an anchor sample, a positive sample, and a negative sample, where the anchor sample and the positive sample belong to the same identity, and the anchor sample and the negative sample belong to different identities.

[0032] During the sampling process, within each training batch, for each anchor sample, the sample with the same identity that is furthest from the feature of that anchor sample is selected as a positive sample, and the sample with the different identity that is closest to the feature of that anchor sample is selected as a negative sample. The above sampling method strengthens the discriminativeness between samples, enabling the model to learn more accurately the common features of the same identity signature and the difference features of different identity signatures during the training process.

[0033] For example, in a certain batch of training, a signature of identity A is used as the anchor sample. The sample with the largest difference in stroke shape but still belonging to the same identity is selected from other samples of identity A as the positive sample. The sample with the most similar stroke structure to the anchor sample is selected from other identities such as B, C, and D as the negative sample, thus forming an effective triplet.

[0034] Ultimately, through training with a large number of such triplet samples, the model can better optimize the feature embedding space, thereby improving its ability to distinguish similar signatures and providing a foundation for subsequent similarity calculation and identity comparison.

[0035] S120: Based on the standardized signature dataset and the triplet sample set, a lightweight convolutional neural network is constructed as the feature extraction backbone network, wherein the lightweight convolutional neural network integrates a data augmentation layer and a style normalization preprocessing layer. In this embodiment of the application, in order to effectively weaken the differences in handwriting style and strengthen the structural features of signature characters, and at the same time adapt to the reasoning needs of different application scenarios, it is necessary to construct a lightweight convolutional neural network with an integrated dedicated preprocessing layer to form a feature extraction backbone network with balanced performance and efficiency.

[0036] Specifically, the training set from the standardized signature dataset is first used as the input sample set, and the corresponding signature identity labels are used as supervision signals to provide clear learning objectives and evaluation criteria for network training. The input sample set must cover structured signature images with different writing tools, backgrounds, and imaging resolutions to ensure the network is exposed to diverse signature features. The signature identity labels must correspond one-to-one with the samples, correctly labeling the identity to which each signature belongs to ensure the effectiveness of supervised training.

[0037] Furthermore, a lightweight convolutional neural network architecture is constructed based on machine learning techniques. Considering the need for handwritten signature feature extraction, a network structure with depthwise separable convolutions as its core is chosen as the foundation. This reduces the total number of network parameters and computational complexity while ensuring the ability to capture detailed signature features.

[0038] During the architecture design process, parameters such as the number of network layers, convolutional kernel size, stride, pooling method, and number of channels should be set reasonably according to the size of the standardized signature dataset and the complexity of the signature features.

[0039] For example, if the standardized signature dataset contains 10,000 to 30,000 samples and the signature features are mainly of low to medium complexity, an 8-10 layer network architecture can be constructed. The input layer receives a single-channel grayscale image of a predetermined fixed size. The encoding end is equipped with 5-6 depthwise separable convolutional layers with 3×3 kernels (stride 1, same padding). Every 1-2 convolutional layers are followed by a 2×2 max pooling layer (stride 2, no padding) to compress the feature map size and improve the feature abstraction. Meanwhile, the decoding end is equipped with 2-3 convolutional layers and batch normalization layers. Finally, a global average pooling layer outputs a fixed-dimensional feature vector.

[0040] Furthermore, if the dataset exceeds 50,000 images and has rich signature details, 1-2 convolutional layers can be added at the encoding stage, and the number of channels in some convolutional layers can be appropriately increased to enhance feature representation capabilities. If it is necessary to adapt to resource-constrained scenarios such as mobile devices, 1-2 convolutional layers can be reduced, and some convolutional layers can use 1×1 convolutional kernels for channel compression to further reduce network complexity.

[0041] Furthermore, a data augmentation layer and a style normalization preprocessing layer are sequentially integrated after the input layer to form a complete feature extraction backbone network. The data augmentation layer needs to design appropriate augmentation operations for the features of handwritten signatures, such as random rotation ±5°, slight scaling (0.9-1.1x), horizontal translation ±10 pixels, and adding low-variance Gaussian noise (variance ≤0.01), etc., to improve the network's robustness to slight signature deformations and positional shifts by enriching the sample morphology.

[0042] In addition, the style normalization preprocessing layer adopts an adaptive instance normalization method. By aligning the style feature distribution of different signatures, it weakens the style differences caused by writing strokes, pressure, speed, etc., and focuses on the key structural features of the signature to ensure that the features learned by the network are more universal.

[0043] Furthermore, the constructed feature extraction backbone network is trained under supervision using the input sample set and corresponding signature identity labels. During training, the structured signature images from the input sample set are input into the network, processed by a data augmentation layer and a style normalization preprocessing layer, and then deep features are extracted through operations such as convolution and pooling. Finally, the predicted results corresponding to the identity category are output.

[0044] Furthermore, the prediction results are compared with the signature identity labels, and the prediction error is quantified by calculating cross-entropy loss. The network parameters are continuously adjusted based on the gradient descent algorithm to optimize feature extraction capabilities. Simultaneously, during training, the model performance needs to be monitored in real-time using a validation set of a standardized signature dataset to avoid overfitting.

[0045] For example, after every 20 rounds of training, the network's identity classification accuracy is tested using validation set data. If the accuracy does not significantly improve for five consecutive rounds (the improvement is ≤0.5%), the network is considered converged. If a downward trend is observed in the validation set accuracy, the training parameters should be adjusted promptly, such as reducing the learning rate from 0.001 to 0.0005, or increasing the batch size from 64 images / batch to 128 images / batch. If the recognition accuracy for a specific style of signature is consistently low, samples of that style can be added to the input sample set for retraining to ensure a balanced recognition capability for different signature styles.

[0046] Finally, when the network's identity classification accuracy on the validation set converges and stabilizes above a preset threshold, the construction of the feature extraction backbone network is complete. This feature extraction backbone network possesses both the efficient inference capabilities of a lightweight architecture and the ability to weaken interfering features and strengthen core features through a dedicated preprocessing layer, providing a feature extraction foundation for the implementation of the subsequent two-stage training strategy and signature recognition.

[0047] S130: Design a two-stage training strategy to optimize the feature extraction backbone network in stages and adjust the embedding space structure. In this embodiment of the application, in order to enhance the intra-class clustering and inter-class separation of the feature embedding space and improve the model's ability to distinguish complex signatures, it is necessary to optimize the network through phased training and structured tuning to obtain a stable feature extraction backbone network.

[0048] Specifically, a fully connected classification layer is first connected to the end of the feature extraction backbone network to construct the first network structure for the first stage of training. The output dimension of the fully connected classification layer matches the number of signature identity categories, ensuring that the network can accurately output the identity prediction result corresponding to each signature.

[0049] Furthermore, the ArcFace classification loss function is used to supervise the training of the first network structure with a standardized signature dataset and corresponding signature identity labels. This ArcFace classification loss function helps the network quickly learn highly discriminative basic features by enhancing the differences between different identity features and compressing the dispersion within the same identity feature.

[0050] During training, the network parameters are continuously updated iteratively, and the changes in identity classification accuracy are monitored using the validation set. When the accuracy remains stable for several consecutive rounds, the network is considered to have converged, and the feature extraction backbone network that has been preliminarily optimized is obtained.

[0051] Furthermore, the fully connected classification layer in the first network structure is removed to construct a second network structure for the second stage of training. Based on the triplet sample set, the triplet loss function is used to structurally tune the second network structure. The triplet loss function optimizes the feature distance relationship between different samples, promoting feature aggregation of samples with the same identity and feature separation of samples with different identities, thus further optimizing the feature embedding space structure.

[0052] During the structured tuning process, a dynamic hard sample mining strategy is used to select effective triples from the triple sample set for training. Effective triples are samples in the current training batch that meet specific distance conditions, which can provide more effective gradient information for network training.

[0053] Meanwhile, during training, the change in the loss value of the triplet loss function on the validation set is continuously monitored. When the loss value stabilizes for several consecutive rounds, the second stage of training is stopped. At this point, the parameter update of the feature extraction backbone network and the structured tuning of the feature embedding space have been completed. The network has both accurate identity classification capabilities and can output discriminative deep feature embedding vectors, providing support for subsequent signature similarity calculation and identity comparison.

[0054] Step S130 in the method provided in this application embodiment includes: A fully connected classification layer is connected to the end of the feature extraction backbone network to construct the first network structure for the first stage of training; Using the ArcFace classification loss function, the first network structure is trained under supervision with the training set of the standardized signature dataset and the corresponding signature identity labels until the network's identity classification accuracy on the verification set converges, thus obtaining a preliminarily optimized feature extraction backbone network. Remove the fully connected classification layer and construct a second network structure for the second stage of training; Based on the triplet sample set, the second network structure is structurally tuned using the triplet loss function; The structured tuning uses a dynamic hard sample mining strategy to select effective triples from the triple sample set for training until the loss value of the triple loss function on the validation set converges. The effective triplet is a triplet sample in the current training batch that satisfies the feature distance difference between the anchor sample and the positive sample being less than the preset dynamic interval. The parameters of the feature extraction backbone network are updated and the structured adjustment of the feature embedding space is completed to obtain the optimized feature extraction backbone network.

[0055] In this embodiment of the application, in order to enable the feature extraction backbone network to have both identity classification capabilities and output highly discriminative deep feature embedding vectors, and to optimize the intra-class compactness and inter-class distinguishability of the feature embedding space, the network needs to be optimized through a combination of phased training and dynamic tuning to meet the accuracy requirements of intelligent handwritten signature recognition in real-world scenarios.

[0056] Specifically, a fully connected classification layer is first connected to the end of the feature extraction backbone network to construct the first network structure for the first stage of training. The output dimension of the fully connected classification layer must be consistent with the total number of identity categories in the standardized signature dataset to ensure that the network can directly output the identity category prediction result for each input signature. For example, if the dataset contains signature samples of 1000 different identities, the output dimension of the fully connected classification layer is set to 1000, corresponding to the predicted probability distribution for each identity.

[0057] Furthermore, the ArcFace classification loss function is used to supervise the training of the first network structure with a standardized signature dataset and corresponding signature identity labels. The ArcFace classification loss function, by introducing angular intervals into the feature space, effectively amplifies the differences between features of different identities while compressing the dispersion of features within the same identity, helping the network quickly learn highly discriminative basic features.

[0058] Specifically, during the training process, the structured signature images in the training set are input into the network in batches. The loss value between the prediction result and the real identity label is calculated, and the parameters of each layer of the network are iteratively updated based on the gradient descent algorithm.

[0059] Meanwhile, the model performance is monitored in real time using the validation set of the standardized signature dataset. After a certain number of training rounds, the network's identity classification accuracy is tested using the validation set data. When the accuracy remains stable for multiple rounds (e.g., the accuracy fluctuation does not exceed 0.3% for 10 consecutive rounds), the network is considered to have converged. At this point, a preliminary optimized feature extraction backbone network has been obtained, which has a stable basic identity classification capability.

[0060] Furthermore, the fully connected classification layer in the first network structure is removed to construct a second network structure for the second stage of training. After removing the fully connected classification layer, the network no longer directly outputs the identity category prediction result, but instead directly outputs the deep feature embedding vector of the signature. This allows the network to focus on the structured optimization of the feature embedding space, thereby improving the accuracy of similarity discrimination between samples.

[0061] Simultaneously, based on the triplet sample set, the second network structure is structurally tuned using the triplet loss function. The core of the triplet loss function is to optimize the feature distance relationship between anchor samples, positive samples, and negative samples, thereby promoting the clustering of features of samples with the same identity in the embedding space and distancing features of samples with different identities from each other, thus optimizing the overall structure of the feature embedding space.

[0062] Specifically, during the structured tuning process, effective triples are selected from the triple sample set for training through a dynamic hard sample mining strategy. Effective triples must meet the condition that the feature distance difference between the anchor sample and the positive sample in the current training batch is less than the preset dynamic interval, so as to ensure that the samples can provide more valuable gradient information for network training, thereby rapidly improving the network's similarity discrimination ability.

[0063] In the method provided in this application embodiment, the step of determining the preset dynamic interval includes: For each candidate triplet in the training set of the standardized signature dataset, the feature distance between the anchor sample and the positive sample, and the feature distance between the anchor sample and the negative sample are calculated respectively. Based on the feature distance set between the anchor point samples and positive samples of all candidate triplets in the current batch, calculate the corresponding mean and standard deviation; Based on the feature distance set between the anchor point samples and negative samples of all candidate triplets in the current batch, calculate the corresponding mean and standard deviation; The initial dynamic interval is obtained by calculating the difference between the first weighted sum of the mean and standard deviation of the feature distance between the anchor sample and the positive sample, and the second weighted sum of the mean and standard deviation of the feature distance between the anchor sample and the negative sample. During training, the initial dynamic interval is dynamically adjusted based on the change in the accuracy of identity classification on the validation set. The dynamic adjustment is to decrease the dynamic interval when the accuracy increases and increase the dynamic interval when the accuracy decreases. The adjusted dynamic interval is applied to the next round of dynamic hard sample mining to select the effective triplet.

[0064] Specifically, for each candidate triple in the training set of the standardized signature dataset, the feature distance between the anchor sample and the positive sample, and the feature distance between the anchor sample and the negative sample are calculated respectively.

[0065] The feature distance is calculated using cosine distance, maintaining consistency with the subsequent signature similarity determination logic to ensure the consistency and effectiveness of the distance metric. Through these two steps, the distribution of feature differences among samples in the current batch of candidate triplets can be determined, thereby clarifying the feature distance boundaries between samples with the same identity and those with different identities.

[0066] Furthermore, based on the feature distance set between the anchor samples and positive samples of all candidate triples in the current batch, the corresponding mean and standard deviation are calculated. Similarly, the mean and standard deviation corresponding to the feature distance set between the anchor samples and negative samples are calculated.

[0067] The mean reflects the overall level of the feature distance between samples of the same or different identities in the current batch, while the standard deviation reflects the dispersion of the distance distribution. Combining the two allows for the analysis of the statistical characteristics of the sample feature distances. For example, if the mean of the feature distance between the anchor sample and the positive sample is 0.3 and the standard deviation is 0.06, it indicates that the feature distances of samples of the same identity in the current batch are mostly concentrated between 0.24 (0.3-0.06=0.24) and 0.36 (0.3+0.06=0.36). If the mean of the feature distance between the anchor sample and the negative sample is 0.7 and the standard deviation is 0.12, then the feature distances of samples of different identities are mainly distributed between 0.58 (0.7-0.12=0.58) and 0.82 (0.7+0.12=0.82).

[0068] Furthermore, the initial dynamic interval is obtained by calculating the difference between the first weighted sum of the mean and standard deviation of the feature distances between the anchor sample and the positive sample, and the second weighted sum of the mean and standard deviation of the feature distances between the anchor sample and the negative sample. The introduction of the first and second weighted sums aims to balance the influence of the mean and standard deviation on the interval. The weighting coefficients need to be preset with reasonable values ​​based on the characteristics of the sample distribution. For example, setting the first weighting coefficient to 0.7 and the second weighting coefficient to 0.8 highlights the dominant role of the mean while also taking into account the distribution dispersion reflected by the standard deviation.

[0069] For example, assuming the mean distance between the anchor point and the positive sample is 0.3 and the standard deviation is 0.06, the first weighted sum is 0.3×0.7+0.06×(1-0.7)=0.228; the mean distance between the anchor point and the negative sample is 0.7 and the standard deviation is 0.12, the second weighted sum is 0.7×0.8+0.12×(1-0.8)=0.584, then the initial dynamic interval is 0.584-0.228=0.356, which can initially distinguish the difficult samples with training value.

[0070] Meanwhile, during training, the initial dynamic interval needs to be dynamically adjusted based on the changes in the accuracy of identity classification on the validation set. The adjustment formula can be expressed as: when the accuracy increases, the new dynamic interval = the old dynamic interval × (1 - the baseline adjustment coefficient × |accuracy change|); when the accuracy decreases, the new dynamic interval = the old dynamic interval × (1 + the baseline adjustment coefficient × |accuracy change|).

[0071] The old dynamic interval is the dynamic interval value used in the previous training round. The accuracy change is equal to the current round's validation set accuracy minus the previous round's validation set accuracy. The baseline adjustment coefficient is a preset adjustment range coefficient, with a value between 0 and 1. For example, setting it to 0.15 ensures both the flexibility of adjustment and avoids excessive interval changes that could lead to training instability.

[0072] For example, if the previous dynamic interval was 0.356, and the accuracy in the current round improves by 3% compared to the previous round (accuracy change of 0.03), then the new dynamic interval = 0.356 × (1 - 0.15 × 0.03) = 0.3544; if the accuracy decreases by 2% (accuracy change of -0.02), then the new dynamic interval = 0.356 × (1 + 0.15 × 0.02) = 0.3561. This dynamic adjustment mechanism can optimize the interval in real time based on the model's training performance. When model performance improves, the interval is appropriately reduced to select more difficult samples, further exploring the model's potential; when model performance declines, the interval is appropriately increased to reduce training difficulty, helping the model return to a stable learning state.

[0073] Finally, the adjusted dynamic interval is applied to the next round of dynamic hard sample mining to select effective triples. The preset dynamic interval, determined through statistical calculation, dynamic feedback, and iterative optimization, can adapt to the training data distribution and model state in each round, ensuring that the selected effective triples have sufficient training difficulty without causing the model to fail to learn due to excessive difficulty. This continuously optimizes the intra-class compactness and inter-class discriminative power of the feature embedding space.

[0074] Furthermore, after determining the adjusted dynamic interval, and combining the dynamic hard sample mining strategy, effective triples are selected from the triple sample set of the current training batch, that is, triple samples in the current training batch that satisfy the feature distance difference between the anchor sample and the positive sample is less than the preset dynamic interval.

[0075] Meanwhile, by continuously selecting such effective triples to train the second network structure, the model can focus on feature learning of hard-to-distinguish samples, continuously narrowing the discrete range of samples with the same identity in the feature embedding space, while widening the feature distance between samples with the same identity and samples with different identities.

[0076] During training, the change of the triplet loss function on the validation set is monitored in real time. When the loss value remains stable for several consecutive rounds (e.g., the loss value decreases by no more than 0.001 for 8 consecutive rounds) and no longer fluctuates significantly, it indicates that the model has fully learned the discriminative features of the samples and the structure of the feature embedding space tends to be optimal.

[0077] At this point, the second stage of structured tuning is stopped, and the parameter updates and structured tuning of the feature extraction backbone network and feature embedding space are completed to obtain the optimized feature extraction backbone network. This network retains the strong identity classification ability formed in the first stage of training, and through the fine tuning in the second stage, it has the ability to accurately distinguish similarity, which can provide technical support for subsequent deep feature extraction, similarity calculation and identity comparison of handwritten signatures.

[0078] S140: Using the optimized feature extraction backbone network, the deep feature embedding vectors of several handwritten signatures are collected in real time, and the similarity scores between signature pairs are calculated to form a full similarity matrix. In this embodiment of the application, in order to provide a quantitative basis for identity determination and batch clustering, it is necessary to construct a similarity matrix that can fully reflect the relationship between signatures through standardized preprocessing, deep feature extraction and full similarity calculation.

[0079] Specifically, the same grayscale processing as the structured processing is first performed on several handwritten signatures, followed by orientation correction processing based on the minimum bounding rectangle and pixel-level alignment processing to uniformly scale them to a predetermined fixed size, in order to ensure that the feature distribution remains consistent with that of the training samples.

[0080] Furthermore, the preprocessed set of signature images to be identified is sequentially input into the optimized feature extraction backbone network. After two stages of training and optimization, this network has improved feature extraction capabilities, enabling it to analyze surface style differences in signatures, capture key structures and intrinsic features, and then output the corresponding set of deep feature embedding vectors.

[0081] Based on the obtained set of deep feature embedding vectors, the cosine similarity between any two deep feature embedding vectors is calculated to obtain the corresponding similarity score. Cosine similarity effectively measures the angle between two feature vectors, thus reflecting the degree of similarity between signatures. A higher score indicates a higher feature matching degree and a greater likelihood that the signatures belong to the same person.

[0082] Finally, a full similarity matrix is ​​constructed based on the similarity scores between all signature pairs. The element in the i-th row and j-th column of the matrix corresponds to the similarity score between the i-th and j-th signatures. This matrix-based presentation intuitively displays the similarity relationships between all signatures to be identified, providing data support for the subsequent application of adaptive thresholds.

[0083] Step S140 in the method provided in this application embodiment includes: The same grayscale processing, orientation correction processing based on the minimum bounding rectangle, and pixel-level alignment processing based on uniform scaling to a predetermined fixed size are sequentially performed on the aforementioned handwritten signatures to obtain a set of signature images to be recognized. The set of signature images to be identified is sequentially input into the optimized feature extraction backbone network to extract the corresponding set of deep feature embedding vectors. Based on the set of deep feature embedding vectors, the cosine similarity between any two deep feature embedding vectors is calculated to obtain the corresponding similarity score; A full similarity matrix is ​​constructed based on the similarity scores between all signature pairs, where the element in the i-th row and j-th column of the full similarity matrix is ​​the similarity score between the i-th signature and the j-th signature.

[0084] In this embodiment of the application, in order to avoid recognition deviations caused by inconsistent feature expressions or incomplete similarity assessments, it is necessary to construct a quantitative matrix that can fully reflect the relationship between signatures through standardized preprocessing, deep feature extraction and full similarity calculation, so as to improve the accuracy of handwritten signature recognition.

[0085] First, several handwritten signatures undergo the same grayscale processing as the structured processing, followed by orientation correction based on the minimum bounding rectangle, and pixel-level alignment by uniformly scaling to a predetermined fixed size. Maintaining complete consistency between the preprocessing and training phases is crucial to ensuring that the signatures to be recognized share the same feature distribution as the training samples.

[0086] Specifically, grayscale processing can strip away redundant color information from the signature image, retaining only core structural features such as stroke outlines and thickness. Simultaneously, orientation correction processing based on the minimum bounding rectangle can accurately locate the signature's tilt angle and perform rotation correction to eliminate feature misalignment caused by directional deviations.

[0087] Meanwhile, the process of uniformly scaling to a predetermined fixed size requires determining reasonable specifications, such as 128×256 pixels. This size can not only fully preserve the stroke details of the signature, but also control the amount of data to improve the efficiency of subsequent feature extraction. During the scaling process, a proportional scaling combined with boundary padding is used to avoid structural deformation of the signature due to stretching or compression, and finally form a set of signature images to be recognized with uniform specifications.

[0088] Furthermore, the preprocessed set of images of the signatures to be identified is sequentially input into the optimized feature extraction backbone network. After two stages of training and embedding space structuring, this network possesses superior feature extraction and discrimination capabilities. The input images are processed layer by layer through the network's convolutional layers, pooling layers, style normalization layers, etc., gradually transforming two-dimensional image information into one-dimensional deep feature embedding vectors. Each vector is a high-dimensional numerical combination, capable of comprehensively and compactly representing the core features of the corresponding signature.

[0089] For example, the feature embedding vectors output by the network will show a high degree of similarity for signatures in different writing states of the same identity; while the feature embedding vectors of signatures of different identities will have significant differences, providing a feature basis for subsequent similarity calculation.

[0090] Furthermore, based on the obtained set of deep feature embedding vectors, the cosine similarity between any two deep feature embedding vectors is calculated to obtain the corresponding similarity score. Cosine similarity measures the degree of similarity by the angle between two vectors in high-dimensional space; the smaller the angle, the higher the similarity and the closer the score is to 1; the larger the angle, the lower the similarity and the closer the score is to 0. This measurement method effectively avoids the influence of feature vector scale differences, thus reflecting the intrinsic matching degree of the signature features.

[0091] During the calculation process, it is necessary to traverse each pair of vectors in the feature embedding vector set and perform similarity calculations one by one to ensure that no relationship between any two signatures is missed. For example, if there are 100 signatures to be identified, then 100×(100-1) / 2=4950 similarity calculations need to be performed to fully cover the similarity evaluation of all signature pairs.

[0092] Furthermore, a full similarity matrix is ​​constructed based on the similarity scores between all signature pairs. The number of rows and columns in the matrix corresponds to the number of signatures to be identified, and the element in the i-th row and j-th column directly corresponds to the similarity score between the i-th and j-th signatures, forming an intuitive and systematic similarity association matrix.

[0093] For example, the signature to be identified contains three samples A, B, and C. The element in the first row and second column of the matrix is ​​the similarity score between A and B, the element in the second row and third column is the similarity score between B and C, and so on.

[0094] Ultimately, the obtained full similarity matrix completely records the similarity relationships between all the signatures to be identified, providing specific quantitative data for the subsequent application of adaptive thresholds, and also providing a basis for the calculation of inter-cluster similarity in hierarchical clustering algorithms, ensuring that subsequent identity comparison and batch clustering can be carried out in an orderly manner based on comprehensive similarity information.

[0095] S150: Based on the full similarity matrix, an adaptive threshold based on the validation set distribution is used, combined with a hierarchical clustering algorithm, to complete the identity comparison and batch clustering of the several handwritten signatures.

[0096] In this embodiment of the application, in order to avoid clustering bias and misidentification caused by insufficient adaptability of fixed threshold, it is necessary to combine adaptive threshold and hierarchical clustering to achieve reasonable grouping of signatures and accurate identity determination.

[0097] Specifically, each of the several handwritten signatures is first used as an initial independent cluster to initialize the cluster set, laying the foundation for the subsequent cluster merging process and ensuring that each signature can participate in the cluster analysis.

[0098] Furthermore, based on the full similarity matrix, the inter-cluster similarity between any two clusters in the cluster set is calculated. This quantitative calculation of inter-cluster similarity clarifies the degree of association between different clusters, providing data support for cluster merging decisions.

[0099] Subsequently, the two clusters with the highest inter-cluster similarity are selected, and it is determined whether the inter-cluster similarity is greater than or equal to an adaptive threshold. The adaptive threshold is dynamically determined based on the distribution of the validation set, adapting to the similarity distribution characteristics of the current batch of signatures and effectively balancing the accuracy and flexibility of clustering.

[0100] If the inter-cluster similarity is greater than or equal to the adaptive threshold, it indicates that the signatures in the two clusters are highly likely to belong to the same identity. In this case, the two clusters are merged into a new cluster, the cluster set is updated, and the step of calculating the inter-cluster similarity is returned to continue the cluster merging analysis.

[0101] If the inter-cluster similarity is less than the adaptive threshold, or if the inter-cluster similarity has been determined, it means that there are no clusters that meet the merging conditions. At this point, the merging process is terminated, and the final cluster set is output as the batch clustering result.

[0102] Finally, identity comparison is completed based on the final cluster set. By associating the clusters with known identity features, the identity attribution of the signatures within each cluster is clarified, thereby achieving efficient classification and identity verification of large-scale handwritten signatures.

[0103] Step S150 in the method provided in this application embodiment includes: Initialize the cluster set by taking each of the aforementioned handwritten signatures as an initial independent cluster; Based on the full similarity matrix, calculate the inter-cluster similarity between any two clusters in the cluster set; Select the two clusters with the highest inter-cluster similarity and determine whether the inter-cluster similarity is greater than or equal to the adaptive threshold; If the inter-cluster similarity is greater than or equal to the adaptive threshold, the two clusters are merged into a new cluster, the cluster set is updated, and the inter-cluster similarity is recalculated. If the inter-cluster similarity is less than the adaptive threshold, or if the inter-cluster similarity has been determined, the merging process is terminated, the final cluster set is output as the batch clustering result, and identity verification is performed based on the cluster set.

[0104] In this embodiment of the application, in order to avoid misjudgment of identity due to insufficient adaptability of fixed threshold, it is necessary to combine adaptive threshold based on the distribution of the verification set with hierarchical clustering algorithm, so as to ensure the rationality of clustering results and the accuracy of identity determination through ordered cluster merging and identity association process.

[0105] Specifically, the cluster set is first initialized by using each of the handwritten signatures as an initial independent cluster. This initialization ensures that each signature to be identified can participate in subsequent cluster analysis as an independent entity, avoiding the omission of any signature's attribution. For example, if there are 50 handwritten signatures to be identified, 50 independent clusters are generated, with each cluster containing only one signature sample.

[0106] Furthermore, based on the full similarity matrix, the inter-cluster similarity between any two clusters in the cluster set is calculated. The inter-cluster similarity is calculated using the average similarity score between all samples within one cluster and all samples within another cluster. This comprehensively reflects the overall correlation between the two clusters, thus avoiding bias in cluster merging decisions due to abnormal similarity of a single sample.

[0107] For example, when calculating the inter-cluster similarity between cluster A and cluster B, it is necessary to extract the similarity score between each sample in cluster A and each sample in cluster B from the full similarity matrix, and then calculate the average of these scores as the inter-cluster similarity between the two clusters, so as to ensure that the result can objectively reflect the feature matching level of the two clusters.

[0108] Furthermore, the two clusters with the highest inter-cluster similarity scores are selected from all clusters. It is then determined whether the inter-cluster similarity is greater than or equal to an adaptive threshold, which in turn determines whether to perform a cluster merging operation. The adaptive threshold is dynamically adjusted based on the similarity distribution of the validation set, adapting to the similarity distribution characteristics of the current signature to be identified.

[0109] In the method provided in this application embodiment, the step of determining the adaptive threshold based on the validation set distribution includes: Using the optimized feature extraction backbone network, the deep feature embedding vectors of all signature samples in the validation set of the standardized signature dataset are extracted. Calculate the similarity score between all signature pairs in the verification set, and analyze the distribution of the verification set similarity scores. Based on the similarity score distribution of the validation set, the false recognition rate and rejection rate under different candidate thresholds are calculated, and the similarity score corresponding to the same error rate is used as the initial global threshold. Calculate the mean and variance of all scores in the full similarity matrix, and use them as the score statistics for the current task batch; Compare the current task batch score statistics with the similarity score distribution statistics of the validation set, and calculate the mean offset and variance change ratio. Based on the mean offset and variance change ratio, the initial global threshold is dynamically adjusted with weights to generate the final adaptive threshold applied to the current batch clustering.

[0110] Specifically, the optimized feature extraction backbone network is first used to extract deep feature embedding vectors from all signature samples in the validation set of the standardized signature dataset. The validation set serves as independent evaluation data during model training; its signature samples cover features from different identities, writing styles, and collection conditions, comprehensively reflecting the feature distribution range covered by the model training.

[0111] Furthermore, the similarity scores between all signature pairs in the validation set are calculated, and the distribution of similarity scores in the validation set is analyzed. The similarity scores are calculated using cosine similarity, ranging from 0 to 1; a higher score indicates a higher degree of feature matching between the signature pairs. By traversing all signature combinations in the validation set, the similarity score for each pair is calculated, forming a complete score set. Statistical analysis is then used to obtain the distribution patterns of the scores, including the concentration intervals, peak positions, and dispersion, providing data for subsequent threshold selection.

[0112] For example, the similarity scores of similar signature pairs in the verification set are mostly concentrated between 0.75 and 0.95, while the scores of dissimilar signature pairs are mostly concentrated between 0.1 and 0.4, forming a clear bimodal distribution.

[0113] Furthermore, based on the similarity score distribution of the validation set, the false recognition rate and false rejection rate under different candidate thresholds are calculated, and the similarity score corresponding to the same error rate is used as the initial global threshold, thereby providing a benchmark for the dynamic adjustment of the subsequent threshold.

[0114] The method provided in this application embodiment, based on the similarity score distribution of the validation set, calculates the false recognition rate and false rejection rate under different candidate thresholds, and uses the similarity score corresponding to the same error rate as the initial global threshold, including: From the similarity score distribution of the validation set, a set of candidate thresholds is extracted according to a preset step size; For each candidate threshold in the candidate threshold set, the proportion of similar signature pairs with scores lower than the candidate threshold in the similarity score distribution of the verification set is calculated as the rejection rate, and the proportion of dissimilar signature pairs with scores higher than the candidate threshold is calculated as the false recognition rate. Iterate through all candidate thresholds, find the candidate threshold that makes the rejection rate and the false recognition rate closest, and calculate the average of the rejection rate and the false recognition rate corresponding to the current candidate threshold as the equal error rate; The candidate threshold corresponding to the equal error rate is selected as the initial global threshold.

[0115] Specifically, a candidate threshold set is first extracted from the similarity score distribution of the validation set according to a preset step size. The preset step size needs to balance computational efficiency and threshold accuracy; for example, it can be set to 0.01 to cover the main distribution range of the similarity scores in the validation set, forming a set containing 85 candidate values.

[0116] In addition, the coverage of the candidate threshold should match the actual score distribution. If the lowest similarity score of similar signature pairs in the validation set is 0.5 and the highest similarity score of dissimilar signature pairs is 0.8, then the candidate threshold can focus on the range of 0.5 to 0.8 to further improve the screening efficiency.

[0117] Furthermore, for each candidate threshold in the candidate threshold set, the proportion of similar signature pairs in the similarity score distribution of the verification set with scores lower than the candidate threshold is calculated as the rejection rate, and the proportion of dissimilar signature pairs with scores higher than the candidate threshold is calculated as the false acceptance rate. The rejection rate reflects the probability of misidentifying the same identity signature as different identities, while the false acceptance rate reflects the probability of misidentifying different identity signatures as the same identity; both are core indicators for measuring the reasonableness of the thresholds.

[0118] For example, when the candidate threshold is 0.65, the proportion of all similar signature pairs in the verification set with scores below 0.65 is counted. If 3% of similar signature pairs are misjudged as mismatches, the rejection rate is 3%. The proportion of all dissimilar signature pairs with scores above 0.65 is counted. If 4% of dissimilar signature pairs are misjudged as matches, the false recognition rate is 4%.

[0119] Furthermore, iterate through all candidate thresholds to find the one that makes the false acceptance rate closest to the false acceptance rate. During this process, compare the difference between the two error rates for each candidate threshold; the smaller the difference, the better the threshold balances the two types of errors.

[0120] For example, a candidate threshold of 0.68 corresponds to a rejection rate of 3.2% and a false acceptance rate of 3.5%, with a difference of 0.3%; a candidate threshold of 0.69 corresponds to a rejection rate of 4% and a false acceptance rate of 3%, with a difference of 1%, so 0.68 is more in line with the balance requirement.

[0121] Furthermore, after finding the target candidate threshold, the average of the rejection rate and the false recognition rate corresponding to the current candidate threshold is calculated as the equal error rate, which reflects the overall error level under the current candidate threshold.

[0122] Finally, a candidate threshold corresponding to the equal error rate is selected as the initial global threshold. For example, the equal error rate corresponding to the candidate threshold of 0.68 is (3.2%+3.5%) / 2=3.35%. This threshold can avoid a large number of similar signatures being rejected due to an excessively high threshold, and can also reduce the misidentification of dissimilar signatures due to an excessively low threshold, thereby achieving a balanced control of the two types of error risks.

[0123] Furthermore, after determining the initial global threshold, the mean and variance of all scores in the full similarity matrix are calculated as the score statistics for the current task batch. The mean reflects the overall level of similarity between signature pairs in the current batch, while the variance reflects the dispersion of the score distribution. Combining these two metrics allows for analysis of the feature matching distribution characteristics of the current task. For example, if the mean of all scores in the full similarity matrix is ​​0.63 and the variance is 0.07, it indicates that the overall similarity of the current batch of signatures is moderate, and the score distribution is relatively concentrated.

[0124] Furthermore, the score statistics of the current task batch are compared with the statistics of the similarity score distribution of the validation set, and the mean shift and variance change ratio are calculated. Specifically, the mean shift is the mean score of the current batch minus the mean score of the validation set. If the mean score of the validation set is 0.59, then the mean shift is 0.63 - 0.59 = 0.04, indicating that the overall similarity of the current batch is higher than that of the validation set.

[0125] Furthermore, the variance change ratio is (current batch variance - validation set variance) / validation set variance. If the validation set variance is 0.05, then the variance change ratio is (0.07 - 0.05) / 0.05 = 40%, reflecting that the current batch score dispersion is 40% higher than that of the validation set. These two indicators can quantify the distribution difference between the current batch and the validation set, providing a clear basis for adjusting the initial global threshold.

[0126] Furthermore, based on the ratio of mean offset to variance change, the initial global threshold is dynamically adjusted with weights to generate the final adaptive threshold applied to the current batch of clustering.

[0127] Specifically, the adjustment formula is "adaptive threshold = initial global threshold + β1 × mean offset - β2 × variance change ratio", where β1 and β2 are preset adjustment weight coefficients, with values ​​ranging from 0 to 1, used to reasonably allocate the influence of the two indicators on the threshold.

[0128] For example, if β1=0.7, β2=0.3, and the initial global threshold is 0.68, and the mean offset of the current batch score is 0.04 and the variance change ratio is 0.4, substituting into the adjustment formula, we can get the adaptive threshold = 0.68 + 0.7 × 0.04 - 0.3 × 0.4 = 0.588.

[0129] Ultimately, the adjusted adaptive threshold not only maintains the balanced control of the initial global threshold on the rejection rate and the false recognition rate, but also adapts to the distribution differences of the signature similarity in the current batch through quantitative correction, thus accurately defining the similarity boundary between signatures of the same type and those of different types.

[0130] Furthermore, after determining the adaptive threshold, if the calculated inter-cluster similarity is greater than or equal to the adaptive threshold, it indicates that the signatures in the two clusters are highly matched in terms of features and have a very high probability of belonging to the same identity. In this case, the two clusters are merged into a new cluster, and the cluster set is updated.

[0131] Simultaneously, after merging, the step of calculating inter-cluster similarity needs to be returned to. The updated cluster set undergoes a full inter-cluster similarity calculation again, continuously searching for clusters that meet the merging criteria for iterative merging. For example, the initial 50 independent clusters become 48 clusters after the first round of merging. The inter-cluster similarity of these 48 clusters is then recalculated, and clusters that meet the threshold requirements are merged again until no combination meets the criteria.

[0132] Conversely, if the inter-cluster similarity is less than the adaptive threshold, or if the similarity between all clusters has been determined and no combination meets the merging criteria, the merging process terminates, and the final set of clusters is output as the batch clustering result. At this point, the signatures within each cluster have high feature consistency, and the signature features between different clusters differ significantly, providing a clear grouping basis for subsequent identity comparison.

[0133] Finally, identity comparison is completed based on the final cluster set. Using a known identity feature reference library as a basis, the identity of each cluster is determined by matching the similarity between the cluster center feature vector and the reference feature vector.

[0134] The method provided in this application embodiment, which performs identity comparison based on the cluster set, includes: Extract all signature samples belonging to known identities from the standardized signature dataset, use the optimized feature extraction backbone network to obtain deep feature embedding vectors, and construct an identity feature reference library, wherein each reference entry in the identity feature reference library is associated with a known identity label; For each cluster in the cluster set, the mean of the deep feature embedding vectors of all signature samples within the cluster is calculated to obtain the corresponding cluster center feature vector. The cosine similarity of each cluster center feature vector with all reference feature vectors in the identity feature reference library is calculated to obtain multiple reference similarities; Select the maximum value among the multiple reference similarities and compare it with a preset matching confidence threshold; If the maximum value is greater than or equal to the matching confidence threshold, then the identity of all signature samples in the current cluster is determined as the corresponding known identity label; If the maximum value is less than the matching confidence threshold, then the identities of all signature samples in the current cluster are marked as unknown. Iterate through all clusters in the cluster set and output a comparison list containing each signature sample and the final identity determination result.

[0135] Specifically, firstly, all signature samples belonging to known identities are extracted from the standardized signature dataset. Then, an optimized feature extraction backbone network is used to obtain their deep feature embedding vectors, constructing an identity feature reference library. The known identity samples cover the signature features of different individuals, and the feature vectors extracted by the optimized network can represent the core signature structure of each identity. Each reference entry in the identity feature reference library is associated with a unique known identity label, providing a clear identity reference standard for subsequent comparisons.

[0136] Furthermore, for each cluster in the cluster set, the mean of the deep feature embedding vectors of all signature samples within the cluster is calculated to obtain the corresponding cluster center feature vector. The cluster center feature vector is a concentrated representation of the features of all samples within the cluster, which can mitigate feature biases caused by writing fluctuations, imaging interference, etc., and more stably reflect the overall features of the cluster signature.

[0137] For example, a cluster contains 8 signature samples with the same identity. By calculating the mean of the feature vectors of the 8 samples, the cluster center feature vector can accurately analyze the core structure of the identity signature, while avoiding the influence of abnormal features of a single sample on identity determination.

[0138] Furthermore, cosine similarity is calculated between each cluster center feature vector and all reference feature vectors in the identity feature reference database to obtain multiple reference similarities. Cosine similarity can effectively measure the degree of matching between two feature vectors. The higher the score, the higher the fit between the cluster center feature and the reference feature, and the greater the probability that the signature within the cluster belongs to that reference identity.

[0139] During the calculation, all reference feature vectors in the reference library must be traversed to ensure that no possible identity matching options are missed. For example, the similarity of a cluster center feature vector with each of the 100 reference feature vectors in the reference library needs to be calculated to obtain 100 corresponding reference similarity values.

[0140] Simultaneously, the maximum value among multiple reference similarities is selected and compared with a preset matching confidence threshold. The matching confidence threshold is set to the low quantile of the score distribution of positive sample pairs within the same category, for example, 5%. This setting effectively reduces the risk of misjudgment, ensuring the rigor of identity determination.

[0141] For example, if the similarity scores of most positive sample pairs of the same type are concentrated between 0.7 and 0.9, the threshold corresponding to the 5% low score is set to 0.65. Only when the maximum reference similarity reaches or exceeds the matching confidence threshold is it determined to be an identity match.

[0142] Specifically, if the maximum value is greater than or equal to the matching confidence threshold, it means that the cluster center feature is highly matched with the feature of the corresponding reference identity, and the probability that all signature samples in the cluster belong to the identity is extremely high. In this case, the identity of all signature samples in the current cluster is determined as the corresponding known identity label.

[0143] Conversely, if the maximum value is less than the matching confidence threshold, it means that there is no known identity with sufficient matching degree in the identity feature reference library for this cluster feature, or the signature within the cluster may belong to a new identity that has not been recorded in the reference library. In this case, the identity of all signature samples in the current cluster will be marked as an unknown identity.

[0144] Finally, the algorithm iterates through all clusters in the cluster set, performs identity verification for each cluster, and outputs a comparison list containing each signature sample and the final identity verification result. The list clearly indicates the sample identifier, the cluster it belongs to, and the corresponding identity result for each signature. This facilitates subsequent classification and management of signatures with known identities and further verification of signatures with unknown identities, comprehensively meeting the needs of handwritten signature identity verification and management in practical applications.

[0145] The embodiments of this application, through the specific implementation methods described above, achieve the following technical effects: This application proposes an intelligent handwritten signature recognition method combining feature extraction and matching. First, it collects original images of handwritten signatures from multiple sources. After structured processing including grayscale conversion, orientation correction, and pixel-level alignment, these images are divided into training, validation, and test sets according to a preset ratio to construct a standardized signature dataset. Then, based on identity labels, a triplet sample set is generated using a sampling strategy of the furthest positive sample and the nearest negative sample within a batch. Next, a lightweight convolutional neural network integrating a data augmentation layer and a style normalization preprocessing layer is constructed using this dataset as the backbone network for feature extraction. Subsequently, the network is optimized through a two-stage training strategy. The first stage involves connecting a fully connected classification layer and training it using the ArcFace classification loss function until... Once the validation set accuracy converges, the fully connected classification layer is removed in the second stage. The feature embedding space is then tuned based on the triplet loss function and a dynamic hard sample mining strategy. The same structured processing is then applied to several handwritten signatures to be identified. The optimized network is then used to extract deep feature embedding vectors, and the cosine similarity between any two vectors is calculated to construct a full similarity matrix. Finally, an initial global threshold is determined based on the validation set similarity distribution. An adaptive threshold is dynamically adjusted using the current batch score statistics. Batch clustering is then performed using a hierarchical clustering algorithm. An identity feature reference library is then constructed, and identity comparison is completed based on the similarity matching between cluster center feature vectors and reference feature vectors. A comparison list containing all signatures and identity results is output.

[0146] The method provided in this application solves the problems of traditional handwritten signature recognition relying on fixed judgment standards, difficulty in adapting to diverse signature changes, and frequent misjudgments in complex scenarios through the technical solution of "multi-source data acquisition and standardized processing - lightweight feature extraction network construction - two-stage training and embedding space adjustment - similarity matrix construction - adaptive threshold clustering and identity comparison". It improves the accuracy and robustness of handwritten signature recognition.

[0147] Example 2, as shown in the appendix Figure 2 As shown, based on the inventive concept of a handwritten signature intelligent recognition method combining feature extraction and matching provided in Embodiment 1, this application also provides a handwritten signature intelligent recognition system combining feature extraction and matching, specifically including: The signature data processing module 01 is used to collect image data of handwritten signatures from multiple sources, perform structured processing, construct a standardized signature dataset, and generate a triplet sample set based on the signature identity label. Feature extraction network module 02 is used to construct a lightweight convolutional neural network based on the standardized signature dataset and the triplet sample set, as the backbone network for feature extraction, wherein the lightweight convolutional neural network integrates a data augmentation layer and a style normalization preprocessing layer; The network training optimization module 03 is used to design a two-stage training strategy to perform phased optimization and embedding space structure adjustment on the feature extraction backbone network. The similarity matrix construction module 04 is used to collect deep feature embedding vectors of several handwritten signatures in real time using the optimized feature extraction backbone network, and calculate the similarity score between signature pairs to form a full similarity matrix. The identity comparison and clustering module 05 is used to complete the identity comparison and batch clustering of the several handwritten signatures based on the full similarity matrix, using an adaptive threshold based on the distribution of the verification set, combined with a hierarchical clustering algorithm.

[0148] In one embodiment, the signature data processing module 01 is further configured to: Collect original handwritten signature images containing various writing tools, backgrounds, and imaging resolutions; sequentially perform grayscale processing, orientation correction processing based on the minimum bounding rectangle, and pixel-level alignment processing to uniformly scale to a predetermined fixed size on the original handwritten signature images to obtain a structured signature image set; divide the structured signature image set into a training set, a validation set, and a test set according to a preset ratio to form the standardized signature dataset; based on the signature identity labels in the standardized signature dataset, adopt the sampling strategy of the farthest positive sample and the nearest negative sample within the batch to generate the triplet sample set, wherein each triplet includes an anchor sample, a positive sample, and a negative sample.

[0149] In one embodiment, the network training optimization module 03 is further configured to: A fully connected classification layer is connected to the end of the feature extraction backbone network to construct a first network structure for the first stage of training. Using the ArcFace classification loss function, the first network structure is trained under supervision with the training set of the standardized signature dataset and the corresponding signature identity labels until the network's identity classification accuracy on the validation set converges, resulting in a pre-optimized feature extraction backbone network. The fully connected classification layer is removed to construct a second network structure for the second stage of training. Based on the triplet sample set, the second network structure is structurally tuned using a triplet loss function. This structural tuning employs a dynamic hard sample mining strategy to select effective triplets from the triplet sample set for training until the loss value of the triplet loss function converges on the validation set. The effective triplet is a triplet sample in the current training batch that satisfies the feature distance difference between the anchor sample and the positive sample being less than a preset dynamic interval. The parameter updates and structural tuning of the feature extraction backbone network and the feature embedding space are then completed to obtain the optimized feature extraction backbone network.

[0150] Furthermore, the network training optimization module 03 also includes: For each candidate triplet in the training set of the standardized signature dataset, the feature distance between the anchor sample and the positive sample, and the feature distance between the anchor sample and the negative sample are calculated respectively. Based on the feature distance set between the anchor sample and the positive sample of all candidate triplets in the current batch, the corresponding mean and standard deviation are calculated. Based on the feature distance set between the anchor sample and the negative sample of all candidate triplets in the current batch, the corresponding mean and standard deviation are calculated. The difference between the first weighted sum of the feature distance mean and standard deviation between the anchor sample and the positive sample, and the second weighted sum of the feature distance mean and standard deviation between the anchor sample and the negative sample is calculated to obtain the initial dynamic interval. During training, the initial dynamic interval is dynamically adjusted according to the change in the accuracy of identity classification on the validation set. The dynamic adjustment is to decrease the dynamic interval when the accuracy increases and increase the dynamic interval when the accuracy decreases. The adjusted dynamic interval is applied to the next round of dynamic hard sample mining to select the effective triplet.

[0151] In one embodiment, the similarity matrix construction module 04 is also used for: The handwritten signatures are sequentially subjected to the same grayscale processing as the structured processing, orientation correction processing based on the minimum bounding rectangle, and pixel-level alignment processing to be uniformly scaled to a predetermined fixed size to obtain a set of signature images to be identified. The set of signature images to be identified is sequentially input into the optimized feature extraction backbone network to extract the corresponding set of deep feature embedding vectors. Based on the set of deep feature embedding vectors, the cosine similarity between any two deep feature embedding vectors is calculated to obtain the corresponding similarity score. According to the similarity scores between all signature pairs, a full similarity matrix is ​​constructed, where the element in the i-th row and j-th column of the full similarity matrix is ​​the similarity score between the i-th signature and the j-th signature.

[0152] In one embodiment, the identity comparison and clustering module 05 is also used for: Each of the handwritten signatures is used as an initial independent cluster to initialize a cluster set. Based on the full similarity matrix, the inter-cluster similarity between any two clusters in the cluster set is calculated. The two clusters with the highest inter-cluster similarity are selected, and it is determined whether the inter-cluster similarity is greater than or equal to the adaptive threshold. If the inter-cluster similarity is greater than or equal to the adaptive threshold, the two clusters are merged into a new cluster, the cluster set is updated, and the inter-cluster similarity is recalculated. If the inter-cluster similarity is less than the adaptive threshold, or all inter-cluster similarities have been determined, the merging process is terminated, the final cluster set is output as the batch clustering result, and identity verification is performed based on the cluster set.

[0153] Furthermore, the identity comparison and clustering module 05 also includes: Using the optimized feature extraction backbone network, deep feature embedding vectors are extracted from all signature samples in the validation set of the standardized signature dataset. Similarity scores are calculated between all signature pairs in the validation set, and the validation set similarity score distribution is analyzed. Based on the validation set similarity score distribution, the false acceptance rate and false rejection rate are calculated under different candidate thresholds, and the similarity score corresponding to the equal error rate is used as the initial global threshold. The mean and variance of all scores in the full similarity matrix are calculated as the current task batch score statistics. The current task batch score statistics are compared with the statistics of the validation set similarity score distribution, and the mean shift and variance change ratio are calculated. Based on the mean shift and variance change ratio, the initial global threshold is dynamically adjusted with weights to generate the final adaptive threshold applied to the current batch clustering.

[0154] Furthermore, the identity comparison and clustering module 05 also includes: From the similarity score distribution of the verification set, a candidate threshold set is extracted according to a preset step size. For each candidate threshold in the candidate threshold set, the proportion of similar signature pairs with scores lower than the candidate threshold in the similarity score distribution of the verification set is calculated as the rejection rate, and the proportion of dissimilar signature pairs with scores higher than the candidate threshold is calculated as the false acceptance rate. All candidate thresholds are traversed to find the candidate threshold that makes the rejection rate and false acceptance rate closest, and the average of the rejection rate and false acceptance rate corresponding to the current candidate threshold is calculated as the equal error rate. The candidate threshold corresponding to the equal error rate is selected as the initial global threshold.

[0155] Furthermore, the identity comparison and clustering module 05 also includes: Extract all signature samples belonging to known identities from the standardized signature dataset. Use the optimized feature extraction backbone network to obtain deep feature embedding vectors and construct an identity feature reference library, where each reference entry in the identity feature reference library is associated with a known identity label. For each cluster in the cluster set, calculate the mean of the deep feature embedding vectors of all signature samples within the cluster to obtain the corresponding cluster center feature vector. Calculate the cosine similarity between each cluster center feature vector and all reference feature vectors in the identity feature reference library to obtain multiple reference similarities. Select the maximum value among the multiple reference similarities and compare it with a preset matching confidence threshold. If the maximum value is greater than or equal to the matching confidence threshold, determine the identity of all signature samples in the current cluster as the corresponding known identity label. If the maximum value is less than the matching confidence threshold, mark the identity of all signature samples in the current cluster as an unknown identity. Traverse all clusters in the cluster set and output a comparison list containing each signature sample and the final identity determination result.

[0156] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0157] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0158] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A method for intelligent recognition of handwritten signatures combining feature extraction and matching, characterized in that, The method includes: Collect image data of handwritten signatures from multiple sources, perform structured processing, construct a standardized signature dataset, and generate a triplet sample set based on signature identity labels; Based on the standardized signature dataset and the triplet sample set, a lightweight convolutional neural network is constructed as the feature extraction backbone network, wherein the lightweight convolutional neural network integrates a data augmentation layer and a style normalization preprocessing layer. A two-stage training strategy is designed to optimize the feature extraction backbone network in stages and adjust the embedding space structure. The optimized feature extraction backbone network is used to collect deep feature embedding vectors of several handwritten signatures in real time, and the similarity scores between signature pairs are calculated to form a full similarity matrix. Based on the full similarity matrix, an adaptive threshold based on the validation set distribution is used, combined with a hierarchical clustering algorithm, to complete the identity comparison and batch clustering of the several handwritten signatures.

2. The intelligent handwritten signature recognition method combining feature extraction and matching according to claim 1, characterized in that, Image data of handwritten signatures from multiple sources were collected and structured to construct a standardized signature dataset. A triplet sample set was then generated based on the signature identity labels, including: Collect original images of handwritten signatures, including various writing tools, backgrounds, and imaging resolutions; The original handwritten signature image is sequentially subjected to grayscale processing, orientation correction processing based on the minimum bounding rectangle, and pixel-level alignment processing to be uniformly scaled to a predetermined fixed size to obtain a structured signature image set; The structured signature image set is divided into a training set, a verification set, and a test set according to a preset ratio to form the standardized signature dataset; Based on the signature identity labels in the standardized signature dataset, the triplet sample set is generated by using the sampling strategy of the furthest positive sample and the nearest negative sample within the batch. Each triplet includes an anchor sample, a positive sample, and a negative sample.

3. The intelligent recognition method for handwritten signatures combining feature extraction and matching according to claim 1, characterized in that, A two-stage training strategy is designed to optimize and structurally tune the embedding space of the feature extraction backbone network in stages, including: A fully connected classification layer is connected to the end of the feature extraction backbone network to construct the first network structure for the first stage of training; Using the ArcFace classification loss function, the first network structure is trained under supervision with the training set of the standardized signature dataset and the corresponding signature identity labels until the network's identity classification accuracy on the verification set converges, thus obtaining a preliminarily optimized feature extraction backbone network. Remove the fully connected classification layer and construct a second network structure for the second stage of training; Based on the triplet sample set, the second network structure is structurally tuned using the triplet loss function; The structured tuning uses a dynamic hard sample mining strategy to select effective triples from the triple sample set for training until the loss value of the triple loss function on the validation set converges. The effective triplet is a triplet sample in the current training batch that satisfies the feature distance difference between the anchor sample and the positive sample being less than the preset dynamic interval. The parameters of the feature extraction backbone network are updated and the structured adjustment of the feature embedding space is completed to obtain the optimized feature extraction backbone network.

4. The intelligent recognition method for handwritten signatures combining feature extraction and matching according to claim 3, characterized in that, The steps for determining the preset dynamic interval include: For each candidate triplet in the training set of the standardized signature dataset, the feature distance between the anchor sample and the positive sample, and the feature distance between the anchor sample and the negative sample are calculated respectively. Based on the feature distance set between the anchor point samples and positive samples of all candidate triplets in the current batch, calculate the corresponding mean and standard deviation; Based on the feature distance set between the anchor point samples and negative samples of all candidate triplets in the current batch, calculate the corresponding mean and standard deviation; The initial dynamic interval is obtained by calculating the difference between the first weighted sum of the mean and standard deviation of the feature distance between the anchor sample and the positive sample, and the second weighted sum of the mean and standard deviation of the feature distance between the anchor sample and the negative sample. During training, the initial dynamic interval is dynamically adjusted based on the change in the accuracy of identity classification on the validation set. The dynamic adjustment is to decrease the dynamic interval when the accuracy increases and increase the dynamic interval when the accuracy decreases. The adjusted dynamic interval is applied to the next round of dynamic hard sample mining to select the effective triplet.

5. The intelligent handwritten signature recognition method combining feature extraction and matching according to claim 1, characterized in that, Using an optimized feature extraction backbone network, deep feature embedding vectors of several handwritten signatures are collected in real time, and similarity scores between signature pairs are calculated to form a full similarity matrix, including: The same grayscale processing, orientation correction processing based on the minimum bounding rectangle, and pixel-level alignment processing based on uniform scaling to a predetermined fixed size are sequentially performed on the aforementioned handwritten signatures to obtain a set of signature images to be recognized. The set of signature images to be identified is sequentially input into the optimized feature extraction backbone network to extract the corresponding set of deep feature embedding vectors. Based on the set of deep feature embedding vectors, the cosine similarity between any two deep feature embedding vectors is calculated to obtain the corresponding similarity score; A full similarity matrix is ​​constructed based on the similarity scores between all signature pairs, where the element in the i-th row and j-th column of the full similarity matrix is ​​the similarity score between the i-th signature and the j-th signature.

6. The intelligent recognition method for handwritten signatures combining feature extraction and matching according to claim 1, characterized in that, The step of determining the adaptive threshold based on the validation set distribution includes: Using the optimized feature extraction backbone network, the deep feature embedding vectors of all signature samples in the validation set of the standardized signature dataset are extracted. Calculate the similarity score between all signature pairs in the verification set, and analyze the distribution of the verification set similarity scores. Based on the similarity score distribution of the validation set, the false recognition rate and rejection rate under different candidate thresholds are calculated, and the similarity score corresponding to the same error rate is used as the initial global threshold. Calculate the mean and variance of all scores in the full similarity matrix, and use them as the score statistics for the current task batch; Compare the current task batch score statistics with the similarity score distribution statistics of the validation set, and calculate the mean offset and variance change ratio. Based on the mean offset and variance change ratio, the initial global threshold is dynamically adjusted with weights to generate the final adaptive threshold applied to the current batch clustering.

7. The intelligent recognition method for handwritten signatures combining feature extraction and matching according to claim 6, characterized in that, Based on the similarity score distribution of the validation set, the false recognition rate and false rejection rate under different candidate thresholds are calculated, and the similarity score corresponding to the same error rate is used as the initial global threshold, including: From the similarity score distribution of the validation set, a set of candidate thresholds is extracted according to a preset step size; For each candidate threshold in the candidate threshold set, the proportion of similar signature pairs with scores lower than the candidate threshold in the similarity score distribution of the verification set is calculated as the rejection rate, and the proportion of dissimilar signature pairs with scores higher than the candidate threshold is calculated as the false recognition rate. Iterate through all candidate thresholds, find the candidate threshold that makes the rejection rate and the false recognition rate closest, and calculate the average of the rejection rate and the false recognition rate corresponding to the current candidate threshold as the equal error rate; The candidate threshold corresponding to the equal error rate is selected as the initial global threshold.

8. The intelligent recognition method for handwritten signatures combining feature extraction and matching according to claim 1, characterized in that, Based on the full similarity matrix, an adaptive threshold based on the validation set distribution is used, combined with a hierarchical clustering algorithm, to complete the identity comparison and batch clustering of the several handwritten signatures, including: Initialize the cluster set by taking each of the aforementioned handwritten signatures as an initial independent cluster; Based on the full similarity matrix, calculate the inter-cluster similarity between any two clusters in the cluster set; Select the two clusters with the highest inter-cluster similarity and determine whether the inter-cluster similarity is greater than or equal to the adaptive threshold; If the inter-cluster similarity is greater than or equal to the adaptive threshold, the two clusters are merged into a new cluster, the cluster set is updated, and the inter-cluster similarity is recalculated. If the inter-cluster similarity is less than the adaptive threshold, or if the inter-cluster similarity has been determined, the merging process is terminated, the final cluster set is output as the batch clustering result, and identity verification is performed based on the cluster set.

9. The intelligent recognition method for handwritten signatures combining feature extraction and matching according to claim 8, characterized in that, Identity verification is performed based on the cluster set, including: Extract all signature samples belonging to known identities from the standardized signature dataset, use the optimized feature extraction backbone network to obtain deep feature embedding vectors, and construct an identity feature reference library, wherein each reference entry in the identity feature reference library is associated with a known identity label; For each cluster in the cluster set, the mean of the deep feature embedding vectors of all signature samples within the cluster is calculated to obtain the corresponding cluster center feature vector. The cosine similarity of each cluster center feature vector with all reference feature vectors in the identity feature reference library is calculated to obtain multiple reference similarities; Select the maximum value among the multiple reference similarities and compare it with a preset matching confidence threshold; If the maximum value is greater than or equal to the matching confidence threshold, then the identity of all signature samples in the current cluster is determined as the corresponding known identity label; If the maximum value is less than the matching confidence threshold, then the identities of all signature samples in the current cluster are marked as unknown. Iterate through all clusters in the cluster set and output a comparison list containing each signature sample and the final identity determination result.

10. A handwritten signature intelligent recognition system combining feature extraction and matching, characterized in that, The system is used to execute the intelligent handwritten signature recognition method combining feature extraction and matching as described in any one of claims 1-9, the system comprising: The signature data processing module is used to collect image data of handwritten signatures from multiple sources, perform structured processing, construct a standardized signature dataset, and generate a triplet sample set based on the signature identity label. The feature extraction network module is used to construct a lightweight convolutional neural network based on the standardized signature dataset and the triplet sample set, which serves as the backbone network for feature extraction. The lightweight convolutional neural network integrates a data augmentation layer and a style normalization preprocessing layer. The network training optimization module is used to design a two-stage training strategy to perform phased optimization and embedding space structure adjustment on the feature extraction backbone network. The similarity matrix construction module is used to collect deep feature embedding vectors of several handwritten signatures in real time using the optimized feature extraction backbone network, and calculate the similarity score between signature pairs to form a full similarity matrix. The identity comparison and clustering module is used to complete the identity comparison and batch clustering of the several handwritten signatures based on the full similarity matrix, using an adaptive threshold based on the distribution of the verification set, combined with a hierarchical clustering algorithm.

Images