Delegation signature identification method and device, electronic equipment and storage medium

Abstract

The application relates to a method and device for identifying a proxy signature, electronic equipment and a storage medium, and relates to the technical field of information security and pattern recognition. The method comprises the following steps: acquiring a set of signature images without labels; for each original signature image, performing standardization preprocessing on the original signature image; inputting each obtained target signature image into a feature extraction model trained through migration learning to extract a deep feature vector; freezing the parameters of a pre-trained bottom model and fine-tuning the parameters of a top model during training; using a density-based unsupervised clustering algorithm to cluster the target signature images according to the deep feature vectors; performing character recognition on each target signature image in each cluster to obtain signature texts; and if each signature text corresponding to a cluster contains at least two different names, it is determined that a proxy signature exists. The application realizes efficient, automatic and high-credibility identification of the specific illegal mode of one person proxying for multiple people without relying on supervised learning.

Description

Technical Field

[0001] This invention relates to the fields of information security and pattern recognition technology, and in particular to a method, apparatus, electronic device and storage medium for identifying proxy signatures. Background Technology

[0002] With the deepening of digital transformation, the compliance of business processes in areas such as financial transactions, contract signing, logistics receipts, and government services increasingly relies on customers' handwritten signatures as key credentials. However, in actual operation, there is a real risk that salespersons may improperly sign on behalf of customers. Such proxy signing will lead to inaccurate business processes, harm customer rights, and render the company's internal control mechanisms ineffective. Therefore, signature authentication is necessary. Signature verification technologies are mainly divided into online and offline categories, with offline signature verification technology becoming the main research direction.

[0003] Currently, offline signature verification schemes mainly include the following three types: 1) Character-level open-set author recognition scheme based on contrastive mask autoencoders, which achieves feature extraction by fusing mask autoencoders and contrastive learning. 2) Text-independent handwritten author recognition scheme based on multi-stream convolutional neural networks, which uses multi-stream convolutional neural networks to extract local features and spatial correlation features of handwritten text. 3) Chinese character handwriting recognition scheme, which achieves character-level author recognition by combining path signature features with deep convolutional neural networks.

[0004] However, when the above three offline signature verification schemes are applied to the specific business scenario of batch signature recognition, there are the following significant shortcomings: 1) They heavily rely on supervised learning, and their effective application requires a large number of signature samples with known identities for training, which contradicts the conditions of no preset templates and no labeled data in the actual scenario; 2) The technical goals are misaligned with the business logic. They mostly serve single signature comparison or open set identity recognition, and do not cluster documents collected in batches by the same salesperson to optimize the core business pain point of identifying multiple signatures signed by the same person under different names, resulting in insufficient efficiency and accuracy of signature recognition. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method, device, electronic device and storage medium for identifying proxy signatures.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: Firstly, a method for identifying proxy signatures includes: Multiple raw signature images to be identified are obtained to form an unlabeled signature image set; For each original signature image in the signature image set, a standardization preprocessing is performed on the original signature image to obtain the target signature image; Each of the target signature images is input into a feature extraction model trained by transfer learning to extract the deep feature vector of each target signature image; wherein, the feature extraction model is constructed based on a visual transformer model; during model training, the model parameters of the pre-trained multi-layer bottom coding layer are frozen, and the model parameters of the multi-layer top coding layer are fine-tuned. Using a density-based unsupervised clustering algorithm, the target signature images are clustered according to their depth feature vectors to obtain at least one cluster. For each cluster, text recognition is performed on each target signature image within the cluster to obtain the signature text corresponding to each target signature image within the cluster. If the signature text corresponding to each target signature image within the cluster contains at least two different names, then it is determined that there is a proxy signature.

[0007] The beneficial effects of this invention are as follows: First, acquiring an unlabeled set of signature images and performing standardized preprocessing ensures the uniformity of model input, a prerequisite for automated batch processing. Second, by employing a feature extraction model based on a visual transformer and incorporating a hierarchical transfer learning strategy, the model parameters of the pre-trained multi-layered bottom-level encoding layers are frozen during training, while the parameters of the multi-layered top-level encoding layers are fine-tuned. This effectively reduces the risk of overfitting in small-sample scenarios and allows the model's training computation and gradient updates to focus on learning higher-order semantic features in the later multi-layered top-level encoding layers. Third, a density-based unsupervised clustering algorithm is used to cluster these feature vectors. Without pre-setting the number of categories, signatures with similar handwriting are automatically grouped into the same cluster. This accurately corresponds to the business reality that the same person may sign multiple documents, achieving a preliminary physical identification of the writer's identity. Fourth, text recognition is performed for each cluster to extract the signature text, completing the conversion from handwriting images to semantic information. Finally, by checking whether at least two different names appear within the same handwriting cluster, a rigid logical contradiction is established between the handwriting originating from the same person and the signature belonging to different people. This contradiction directly and objectively reveals the act of signing on behalf of others, transforming the complex problem of handwriting identification into a calculable and verifiable consistency verification problem. Thus, without the need for any pre-stored signature templates or authenticity marking data, it enables efficient, automatic, and highly reliable identification of the specific violation pattern of one person signing on behalf of multiple people.

[0008] Based on the above technical solution, the present invention can be further improved as follows.

[0009] Furthermore, the feature extraction model includes a basic feature extraction network and a dual-output head network; The basic feature extraction network is a visual transformer model with the original classification head removed; The dual-output head network includes a classification head network and a projection head network. The classification head network is used to predict the writer's identity category based on the output feature vector of the basic feature extraction network, and the projection head network is used to map the output feature vector of the basic feature extraction network to a contrastive feature space for unsupervised clustering.

[0010] The beneficial effects of adopting the above-mentioned further scheme are as follows: First, the basic feature extraction network uses a visual transformer model with the original classification head removed, allowing it to focus on capturing fine-grained visual features such as stroke direction, stroke spacing, and ligature characteristics of the signature image. In the dual-output head network, the classification head network forces the feature extraction model to predict the writer's identity based on fine-grained visual features, thereby learning highly discriminative identity-identifying features that can finely represent the writing styles of different people. At the same time, the projection head network maps the fine-grained visual features to a low-dimensional vector space (i.e., the contrastive feature space) optimized through contrastive learning, which makes the features of different signatures of the same person closer together in this space, while widening the distance between the features of signatures of different people. In this way, the feature extraction module can output an ideal representation containing strong identity discrimination information and superior clustering space characteristics, thus providing a fundamental guarantee for unsupervised clustering.

[0011] Furthermore, the classification head network includes a compression layer, an activation layer, and a random deactivation layer; wherein, the compression layer is used for progressive feature dimension compression, the activation layer uses Gaussian error linear units as activation functions, and the random deactivation layer is used to randomly discard some output features of the activation layer with a preset probability during training.

[0012] The beneficial effects of adopting the above-mentioned further scheme are as follows: The classification head network includes a compression layer, an activation layer, and a random deactivation layer. The compression layer performs progressive feature dimensionality compression, effectively filtering and refining high-dimensional features, gradually focusing on the most discriminative handwriting details, and avoiding the loss of effective features due to information redundancy or a sudden drop in dimensionality. The activation layer uses Gaussian error linear units, which, due to their smooth non-linear characteristics, can better adapt to and retain subtle visual features commonly found in signature handwriting, such as continuous grayscale changes and gradual changes in pressure, thereby improving the subtlety and accuracy of feature expression. The random deactivation layer randomly discards some activation outputs during training via forward propagation, serving as a powerful regularization method. This prevents the model from over-relying on accidental features of specific samples in the training set (such as accidental ink stains or specific creases in individual signatures), significantly enhancing the model's generalization ability to unseen signature samples with reasonable fluctuations. The synergistic effect of these three elements enables the classification head network to complete its core task more efficiently and robustly during the training phase: driving the entire feature extraction model to learn highly discriminative writer identity features.

[0013] Furthermore, the projection head network includes a linear transformation layer, a batch normalization layer, and an activation layer connected in sequence; wherein the activation layer uses a rectified linear unit as the activation function.

[0014] The beneficial effects of adopting the above-mentioned further scheme are as follows: The projector network consists of a linear transformation layer, a batch normalization layer, and an activation layer connected in sequence. The linear transformation layer projects high-dimensional input features into a lower-dimensional, more compact vector space, directly reducing the complexity of subsequent computations and facilitating the extraction of more essential feature representations. The batch normalization layer, by standardizing the feature distribution, effectively alleviates internal covariate shifts during training, accelerates model convergence, and improves the stability and consistency of feature representations. Using rectified linear units as the activation function, due to their sparse activation, computational efficiency, and ability to alleviate the gradient vanishing problem, is very suitable for introducing necessary nonlinearity into features in signature recognition tasks, thereby enhancing the projector network's ability to fit and express complex handwriting feature patterns. The sequential connection of these three layers ensures that features are normalized and nonlinearly enhanced during dimensionality reduction, ultimately enabling the projector network to more reliably map input features into a lower-dimensional contrastive feature space that is more discriminative and conducive to subsequent unsupervised clustering algorithms for similarity measurement.

[0015] Furthermore, the feature extraction model is trained through transfer learning by iterating through the following steps until the model converges; Obtain signature image samples from multiple different writers to form a training dataset; The training dataset is input into the feature extraction model for forward propagation to obtain the identity category prediction result and the contrast feature vector. The cross-entropy loss is calculated based on the identity category prediction results, and the contrast loss is calculated based on the contrast feature vector. The total loss is obtained by weighted summation of the cross-entropy loss and the contrastive loss, wherein the weight of the contrastive loss is lower than the weight of the cross-entropy loss. Backpropagation is performed based on the total loss; wherein, during backpropagation, the model parameters of the pre-trained multi-layer bottom coding layer are kept unchanged, and the model parameters of the multi-layer top coding layer, the classification head network, and the projection head network are updated based on the gradient of this iteration.

[0016] The beneficial effects of adopting the above-mentioned further scheme are as follows: Calculating cross-entropy loss based on identity category prediction results ensures that the model is directly driven by strong and explicit identity-discriminating supervision signals during training, thereby forcing the model to learn highly discriminative features that can accurately distinguish different writers. This is the foundation for the model to acquire core classification capabilities. Simultaneously, calculating contrastive loss based on contrastive feature vectors introduces a representation learning objective to the model. Its core is to drive the model to learn a feature space distribution where different signatures of the same person are close to each other, while signatures of different people are far apart. This directly optimizes the intra-class clustering and inter-class segregation of features, providing forward-looking preparation for subsequent unsupervised clustering tasks. The two losses are weighted and summed, with the weight of the contrastive loss lower than that of the cross-entropy loss, prioritizing the model's identity classification accuracy to ensure its discriminative power. At the same time, a relatively low weight is used to introduce the contrastive learning objective as an auxiliary regularization and guide, subtly optimizing the spatial structure of features without interfering with the core discrimination task, making it more cluster-friendly. By keeping the model parameters of the multi-layer bottom-level encoding layers unchanged during backpropagation, and only updating the model parameters of the multi-layer top-level encoding layers, the classification head network, and the projection head network, the hierarchical transfer learning strategy is precisely implemented. This not only efficiently reuses the general visual knowledge of the pre-trained multi-layer bottom-level encoding layers, but also focuses on fine-tuning the multi-layer top-level encoding layers, effectively preventing overfitting under small sample conditions and improving the model's generalization ability.

[0017] Furthermore, before performing text recognition on each target signature image within the cluster, the relevant hyperparameters for text detection are adjusted according to at least one of the following parameter adjustment strategies: Lower the text region detection threshold; Lower the detection frame threshold; Increase the reverse clipping ratio of the detection frame; Enable morphological expansion operation.

[0018] The beneficial effects of adopting the above-mentioned further solutions are as follows: Lowering the text region detection threshold and the detection box threshold can reduce the detection threshold of the system for weak signals, blurred or discontinuous stroke regions, allowing more potential, low-contrast signature areas to be initially identified as candidate text regions, effectively addressing the signal unevenness problem caused by writing pressure, tools, or paper background in handwriting. Increasing the detection box anti-cropping ratio allows the system to expand outward from the initially detected core text region when forming the final text detection box, thereby ensuring that the generated text bounding box can completely cover the flowing, elegant stroke edges commonly found in handwritten signatures, avoiding character cropping caused by an overly tight box. Enabling morphological dilation, by dilating the feature map, helps to connect internal character breaks caused by poor image quality, natural stroke discontinuities, or writing style, connecting discrete stroke pixels into a coherent text region, which is particularly important for cursive writing or lightly written signatures. By using at least one of the above parameter adjustment strategies to adjust the relevant hyperparameters of text detection, the text content in the signature image can be located and extracted as completely and accurately as possible.

[0019] Furthermore, the standardization preprocessing of the original signature image to obtain the target signature image includes: The original signature image is decoded and standardized to obtain a bitmap format image; The bitmap format image is geometrically corrected and sized to obtain a standard-sized image; The standard-sized image is subjected to denoising and feature enhancement processing to obtain the target signature image.

[0020] The beneficial effects of adopting the above-mentioned further scheme are as follows: First, decoding and format standardizing the original signature image yields a bitmap format image. This ensures that original signature images of different formats can be uniformly and error-free parsed into a bitmap format image that the program can process, eliminating heterogeneity at the data input level and laying a unified format foundation for subsequent processing. Then, geometric correction and size standardization are performed on the bitmap format image to obtain a standard-sized image. This effectively avoids character distortion and feature aberration caused by image stretching, distortion, or improper sizing, ensuring the consistency of the input image received by the model in the spatial dimension, which is crucial for models that rely on fine stroke features. Finally, denoising and feature enhancement processing are performed on the standard-sized image to obtain the target signature image. This significantly suppresses irrelevant information such as scanning noise, paper texture, and background interference, while sharpening and highlighting the main outline and details of the signature, thereby purifying the input image and allowing the model to focus more on the handwriting features of the writer, rather than noise introduced during image acquisition. In this way, regardless of the quality of the original image, the final output target signature image can meet the strict requirements of the feature extraction model for input image format uniformity, size standardization, clear features, and controllable noise, which greatly improves the adaptability to signature images from different sources and of different quality and the reliability of the processing results.

[0021] Secondly, the present invention provides a signature identification device, comprising: The acquisition module is used to acquire multiple raw signature images to be identified, forming a set of unlabeled signature images; The preprocessing module is used to perform standardization preprocessing on each original signature image in the signature image set to obtain a target signature image; The extraction module is used to input each of the target signature images into a feature extraction model trained by transfer learning, and extract the depth feature vector of each of the target signature images; wherein, the feature extraction model is constructed based on a visual transformer model; during model training, the model parameters of the pre-trained multi-layer bottom coding layer are frozen, and the model parameters of the multi-layer top coding layer are fine-tuned. The clustering module is used to cluster the target signature images according to the depth feature vector of each target signature image using a density-based unsupervised clustering algorithm to obtain at least one cluster. The recognition module is used to perform text recognition on each target signature image within each cluster for each cluster, so as to obtain the signature text corresponding to each target signature image within the cluster. The determination module is used to determine that there is a proxy signing behavior if the signature text corresponding to each of the target signature images in the cluster contains at least two different names.

[0022] Thirdly, the present invention provides an electronic device that adopts the following technical solution: An electronic device includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the signature identification method as described in any of the first aspects.

[0023] Fourthly, the present invention provides a computer-readable storage medium, which adopts the following technical solution: A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the signature identification method as described in any of the first aspects.

[0024] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the proxy signature identification method provided by the present invention; Figure 2 A schematic diagram of the standardized preprocessing flow provided by this invention; Figure 3 A schematic diagram of the training process of the feature extraction model provided by the present invention; Figure 4 This is a schematic diagram of the structure of the feature extraction model provided by the present invention; Figure 5 This is a schematic diagram of the structure of the signature identification device provided by the present invention; Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0026] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0027] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating the proxy signature identification method provided by the present invention. Figure 1 As shown, the method may include the following steps S101-S106.

[0028] S101. Obtain multiple original signature images to be identified, forming a set of unlabeled signature images.

[0029] For example, at bank account opening, logistics delivery signing, or government service windows, salespersons may process multiple documents requiring customer signatures within a single day or a specific time period. These documents, after being scanned or photographed, will generate electronic signature images. The original signature image refers to the electronic signature image directly obtained from the business system, scanning equipment, or image database. The original signature image may be in various formats, such as JPEG, PNG, etc., or may be stored as a Base64 encoded string. The massive amount of original signature images obtained constitutes an unlabeled signature image set, which does not possess any authentic labels or prior classification information regarding the signatory's identity.

[0030] S102. For each original signature image in the signature image set, perform standardization preprocessing on the original signature image to obtain the target signature image.

[0031] Specifically, converting massive amounts of raw signature images from diverse sources and of varying quality into target signature images with uniform format, controllable quality, and clear features can ensure the uniformity of model input and facilitate automated batch processing.

[0032] In one implementation, such as Figure 2 As shown, step S102 may include the following sub-steps S1021-S1023.

[0033] S1021. Decode and standardize the original signature image to obtain a bitmap format image.

[0034] Specifically, the original signature image may not be stored and transmitted as a direct image file in the business system. A common and efficient method is to encode it as a Base64 string. Therefore, the first step is to determine the data format of the original signature image. If the original signature image is in Base64 encoded string format, the appropriate decoding library (e.g., the Base64 standard library can be used in a Python environment) is called to decode it and restore it to the original binary image.

[0035] After decoding, the specific image format (such as PNG or JPEG) needs to be read. This is done by loading the binary image using an image processing library (such as Python's PIL / Pillow or OpenCV). Once loaded, the binary image is represented in memory as a data structure containing a pixel matrix. Specifically, some formats (such as PNG) may include an alpha transparency channel, stored as an RGBA four-channel image. However, subsequent feature extraction models typically require an RGB three-channel image as input. Therefore, the RGBA image is converted to an RGB image, and the transparency channels are discarded or merged.

[0036] After decoding and format standardization, a standard, three-channel bitmap image is obtained. This image exists in memory as a pixel matrix, with each pixel defined by three values: R, G, and B, laying a unified foundation for all subsequent pixel-based operations.

[0037] S1022. Perform geometric correction and size standardization on the bitmap format image to obtain a standard size image.

[0038] Specifically, directly scaling a bitmap image of arbitrary size alters its aspect ratio, causing the signature to be stretched or compressed, severely interfering with handwriting characteristics. A reverse fill operation is used to fill the bitmap image into a square image with equal length and width. The reverse fill operation is implemented as follows: calculate the height and width of the bitmap image, determining the longer side; then, uniformly fill the shorter side (or both sides) with a specific color (usually white or the average color of the image edges) until the image becomes a square (i.e., height equals width). The reverse fill operation changes the size of the image's bounding rectangle while completely preserving the original aspect ratio and shape of the internal signature area.

[0039] The square images obtained after the above filling process are then uniformly scaled to a preset fixed size using an image interpolation algorithm (such as bilinear interpolation). This size must strictly match the standard input size of the selected feature extraction model (e.g., 224 pixels × 224 pixels). At this point, all square images have been transformed into standard-sized images with completely consistent dimensions.

[0040] S1023. Perform denoising and feature enhancement processing on the standard-sized image to obtain the target signature image.

[0041] Specifically, a standard-sized RGB three-channel image is converted into a single-channel grayscale image. This is because for handwriting analysis tasks, color discrimination is usually not critical, while luminance (grayscale) information is sufficient to characterize the shape, density, and contrast of the strokes. This reduces the data volume by two-thirds while preserving core information, improving subsequent processing efficiency.

[0042] Next, Gaussian filtering is applied to the obtained grayscale image to remove noise. This effectively filters out high-frequency random noise caused by scanning sensor noise, paper particles, or slight stains, thus smoothing the image.

[0043] Finally, binarization (thresholding) is performed on the denoised grayscale image. By setting a global or adaptive threshold, the grayscale value of each pixel in the image is compared with the threshold: pixels with grayscale values ​​higher than the threshold are set to white (e.g., a value of 255), representing the background; pixels with grayscale values ​​lower than the threshold are set to black (e.g., a value of 0), representing the signature. This achieves thorough, high-contrast separation of the signature from the background, removing all grayscale information and simplifying the image to pure black and white, thereby maximizing the highlighting of key structural features such as the outline of the signature, the thickness of the strokes, and whether they are connected or disconnected, while completely eliminating complex background textures.

[0044] In this implementation, firstly, the original signature image is decoded and standardized to obtain a bitmap format image. This ensures that original signature images of different formats can be uniformly and error-free parsed into a bitmap format image that the program can process, eliminating heterogeneity at the data input level and laying a unified format foundation for subsequent processing. Then, the bitmap format image undergoes geometric correction and size standardization to obtain a standard-sized image. This effectively avoids character distortion and feature aberration caused by image stretching, distortion, or improper sizing, ensuring the consistency of the input image received by the model in the spatial dimension, which is crucial for models that rely on fine stroke features. Finally, the standard-sized image undergoes denoising and feature enhancement processing to obtain the target signature image. This significantly suppresses irrelevant information such as scanning noise, paper texture, and background interference, while sharpening and highlighting the main outline and details of the signature, thereby purifying the input image and allowing the model to focus more on the writer's handwriting features rather than noise introduced during image acquisition. In this way, regardless of the quality of the original image, the final output target signature image can meet the strict requirements of the feature extraction model for input image format uniformity, size standardization, clear features, and controllable noise, which greatly improves the adaptability to signature images from different sources and of different quality and the reliability of the processing results.

[0045] It should be noted that the above description provides a detailed explanation of the implementation process of step S102 of the present invention by listing a specific embodiment. However, those skilled in the art will understand that this embodiment is merely an example and is not intended to limit the scope of protection of the present invention.

[0046] S103. Input the signature images of each target into the feature extraction model trained by transfer learning, and extract the depth feature vector of each signature image.

[0047] In one implementation, such as Figure 3 As shown, the feature extraction model is trained by transfer learning through iterative steps S301-S305 until the model converges.

[0048] S301. Obtain signature image samples from multiple different writers to form a training dataset.

[0049] Specifically, a number of volunteers (i.e., different writers) are organized or invited to simulate real business scenarios by writing their own signatures on electronic devices or through scanning. These signatures are collected and saved as electronic images to simulate signature samples in real business documents. This ensures that the data is consistent with the form of real signature images.

[0050] For example, 287 valid training samples were collected, covering signature categories from 18 different writers, with an average of about 16 original signature image samples provided by each writer. This scale is consistent with the characteristics of small-sample machine learning scenarios, namely, a limited total number of samples and a small number of original signature image samples for each category. It can accurately simulate the situation in real-world signature recognition scenarios, where salespeople face a large number of unfamiliar customers, and each customer may only have a few signature records.

[0051] To adapt to subsequent model training, the acquired original signature image samples will undergo standardized preprocessing according to the aforementioned step S102 to obtain images with normalized format, size, and quality, thus forming a training dataset that can be directly used for model training. This preprocessing process ensures the consistency of the input data and lays the foundation for stable model training; its specific details will not be repeated here.

[0052] For the data processing of the original signature image sample, at least two modes are supported: one is to crop it to the target size that fits the model input; the other is to retain the Base64 encoding format (this format can ensure the high definition of the image data, which is convenient for subsequent direct parsing and processing).

[0053] S302. Input the training dataset into the feature extraction model for forward propagation to obtain the identity category prediction results and the comparison feature vector.

[0054] Specifically, the feature extraction model is built upon a Vision Transformer (ViT) model (e.g., the vit_base_patch16_224 model). Based on a self-attention mechanism, the ViT model can model the global contextual information of the input target signature image, thereby more effectively capturing fine-grained visual features in the signature image that require long-distance dependencies to understand, such as the direction of the strokes, the spacing between strokes, and the habit of connecting strokes. Compared to the local feature extraction characteristics of traditional convolutional neural networks (CNNs), the ViT model is better suited to the technical requirements of distinguishing subtle differences in the strokes of different writers in the task of signature recognition.

[0055] Optionally, such as Figure 4As shown, the feature extraction model includes a basic feature extraction network 1 and a dual-output head network 2.

[0056] The basic feature extraction network 1 is a visual transformer model with the original classification head removed. The original classification head in the ViT model can be disabled by setting `num_classes=0`, causing the network to output only a 768-dimensional global feature vector. This fully preserves the basic general visual feature extraction capabilities learned by the pre-trained multi-layer low-level encoding layers in basic feature extraction network 1 on massive datasets, while also reserving a standardized feature input interface for the subsequent custom dual-output head network 2.

[0057] During model training, the model parameters of the pre-trained multi-layer bottom coding layers are frozen, and the model parameters of the multi-layer top coding layers are fine-tuned.

[0058] For example, suppose the ViT model includes 12 Transformer layers, each with a clear hierarchical feature division of labor, and the features learned by each Transformer layer have significant differences in task relevance, rather than being homogeneous feature extraction.

[0059] 1) Transformer layers 1 to 6, the bottom-level encoding layers, primarily learn low- to mid-level basic general visual features. These include fundamental common features such as image edges, contours, textures, and grayscale gradient changes. These features are fundamental features of all visual images and are irrelevant to specific business tasks. Higher-level semantic features such as stroke edges and handwriting texture in signature images are essentially combinations of these basic general visual features. These basic general visual features have already been optimally learned and fitted by pre-trained models on massive general datasets like ImageNet. They do not need to be retrained for the signature recognition task. The model parameters of the first 6 Transformer layers can be directly reused after freezing the pre-trained model parameters, effectively saving training computational power and sample resources.

[0060] 2) Transformer layers 7 to 12, the top-level encoding layers, primarily learn high-order semantic features. These features are combinations and abstract representations of basic general visual features, possessing strong task relevance and enabling the mapping between image features and business semantics. Specifically, they represent high-order discrimination of the category of a signature image and the identity features of the writer. These features are highly dependent on the business scenario of signature recognition and require customized adaptation learning for the signature features of 18 types of writers. Therefore, only the model parameters of these 6 Transformer layers are fine-tuned.

[0061] In small-sample scenarios for signature recognition, if full parameter fine-tuning is performed on all 12 layers of the ViT model's Transformer encoder, its more than 80M trainable parameters will far exceed the fitting ability of the limited training samples. This can easily lead to the model overlearning noise in the training data (such as scanning impurities or slight pen stroke deviations), resulting in severe overfitting. Specifically, the model's accuracy is artificially high on the training set, while its generalization ability drops sharply on the validation set.

[0062] To address this issue, this embodiment employs a hierarchical transfer learning strategy: freezing the pre-trained model parameters of the first six Transformer encoding layers and performing supervised fine-tuning only on the parameters of the last six Transformer encoding layers. This strategy is based on the functional differentiation of each layer in the ViT model: its lower layers (such as the first six layers) mainly encode general visual features such as edges and textures, which have cross-task transferability; while its upper layers (such as the last six layers) are responsible for combining these lower-level features to form higher-order semantic features relevant to a specific task.

[0063] By employing a hierarchical transfer learning strategy, firstly, by freezing the bottom layer, the optimal low-level visual feature extraction capability already learned by the model on massive general images is directly and efficiently reused, avoiding repeated learning and computational consumption of this part of general knowledge. Secondly, the number of trainable parameters is reduced by about 50%, which fundamentally and significantly alleviates the inherent risk of overfitting when training a large model on small sample data. Finally, the limited training computing power and gradient updates are precisely focused on the top layer of the model, driving the last 6 layers of the model to learn exclusive high-order semantic features strongly related to handwriting feature identification, achieving precise adaptation from general visual knowledge to business-specific features, and ultimately achieving dual optimization of model training efficiency and generalization performance.

[0064] The dual-output head network 2 comprises a classification head network 21 and a projection head network 22. The classification head network 21 predicts the writer's identity category based on the output feature vector of the basic feature extraction network 1. By forcing the feature extraction model to predict the writer's identity based on fine-grained visual features, the classification head network 21 learns highly discriminative identity-identifying features that finely represent the writing styles of different people. The projection head network 22 maps the output feature vector of the basic feature extraction network 1 to a contrastive feature space used for unsupervised clustering. The projection head network 22 maps fine-grained visual features to a low-dimensional vector space (i.e., the contrastive feature space) optimized through contrastive learning. This allows the features of different signatures of the same person to be closer together, while the features of signatures of different people are further apart. This enables the feature extraction module to output an ideal representation containing strong identity-discriminating information and superior clustering space characteristics, thus providing a fundamental guarantee for unsupervised clustering.

[0065] In one implementation, the classification head network includes a compression layer, an activation layer, and a random deactivation layer.

[0066] The compression layer is used for progressive feature dimensionality compression. For example, the compression layer receives a 768-dimensional global feature vector output from the basic feature extraction network as input and, through a step-by-step, non-linear mapping process, compresses it to an output dimension that matches the total number of identity categories (e.g., 18 categories). This process abandons the approach of directly performing a large dimensionality reduction with a single linear layer, instead employing a multilayer perceptron structure containing two 512-dimensional hidden layers to achieve progressive transformation. This allows the feature vector to flow sequentially through two intermediate hidden layers with sufficient representational capacity, gradually completing feature fusion, filtering, and refinement. Through this progressive compression mechanism, key handwriting details effective for identity verification can be preserved to the greatest extent, effectively avoiding the risk of losing valuable discriminative features due to a sudden reduction in dimensionality, thus further enhancing the class discriminative ability of the feature vector while reducing dimensionality.

[0067] The activation layer employs the Gaussian Error Linear Unit (GELU) as the activation function. Compared to the traditional Rectified Linear Unit (ReLU) activation function, GELU possesses a smooth, continuous, and everywhere differentiable nonlinear characteristic. This mathematical property allows it to better model and convey the inherent visual attributes of signature images that manifest as continuous value changes, such as continuous transitions in gray levels, gradual changes in ink density, and subtle variations in writing pressure. Conversely, the hard truncation characteristic of the ReLU activation function at zero may lead to the loss of gradient information or insufficient feature representation when processing such subtle continuous features. Therefore, choosing GELU can more accurately adapt to the representation requirements of fine-grained signature features, thereby preserving and enhancing the micro-stroke information that is crucial for distinguishing different calligrapher styles in deep networks, contributing to improved overall discriminative power of feature extraction.

[0068] Random deactivation layers are used to randomly discard a portion of the output features of the activation layers during training with a preset probability (e.g., 0.3). By randomly masking (deactivating) a portion of the feature neurons in each training iteration, diverse and simplified network substructures are constructed. This effectively forces the network to avoid over-reliance on any few fixed neural pathways when making forward predictions, thus preventing the model from making excessive and erroneous dependencies on random, irrelevant noise features in the training data (e.g., random ink spots caused by scanning, specific creases in the paper, or accidental jitter during writing). This regularization strategy can significantly reduce the risk of overfitting the model on the training dataset, thereby enhancing the model's stability and discrimination accuracy when facing unknown signature samples, i.e., improving the model's generalization ability.

[0069] In this embodiment, the classification head network includes a compression layer, an activation layer, and a random deactivation layer. The compression layer performs progressive feature dimensionality compression, effectively filtering and refining high-dimensional features, gradually focusing on the most discriminative handwriting details, and avoiding the loss of effective features due to information redundancy or a sudden drop in dimensionality. The activation layer uses Gaussian error linear units, which, due to their smooth non-linear characteristics, better adapt to and preserve subtle visual features commonly found in signature handwriting, such as continuous grayscale changes and gradual changes in pressure, thereby improving the subtlety and accuracy of feature representation. The random deactivation layer randomly discards some activation outputs during training via forward propagation, serving as a powerful regularization method. This prevents the model from over-relying on accidental features of specific samples in the training set (such as accidental ink stains or specific creases in individual signatures), significantly enhancing the model's generalization ability to unseen signature samples with reasonable fluctuations. The synergistic effect of these three layers enables the classification head network to complete its core task more efficiently and robustly during the training phase: driving the entire feature extraction model to learn highly discriminative writer identity features.

[0070] In one embodiment, the projection head network includes a linear transformation layer, a batch normalization layer, and an activation layer connected in sequence.

[0071] The linear transformation layer uses a learnable weight matrix to map the input 768-dimensional global feature vector to a pre-defined, lower-dimensional (e.g., 256-dimensional) vector space. The 256-dimensional dimension balances the effectiveness of feature representation with model computational efficiency. It can perform dimensionality reduction and preliminary linear recombination of features, extract latent variables more critical to subsequent tasks, and set the basic output dimension for the entire projector network.

[0072] The batch normalization layer follows the linear transformation layer. During training, the batch normalization layer standardizes each dimension of the current feature vector, making its mean 0 and variance 1. This operation can alleviate the internal covariate bias problem during training, stabilize the input distribution of each layer, thereby effectively accelerating the convergence speed of the model and allowing the use of a higher learning rate. It also plays a role in regularization to some extent, enhancing the stability of training.

[0073] The activation layer uses rectified linear units as the activation function and is located after the batch normalization layer. Through its nonlinear transformation characteristics, the activation layer introduces nonlinear expressive power into the linearly and normally processed features. This enables the projector network to learn and fit more complex interactions between features, thereby enhancing the network's ability to combine and abstract input features, and generating more expressive and discriminative feature representations.

[0074] Furthermore, under the direct supervision of the contrastive loss function, the projector network is driven to make the output feature space exhibit ideal distribution characteristics of significant clustering of similar samples and separation of dissimilar samples.

[0075] In this embodiment, the projection head network comprises a linear transformation layer, a batch normalization layer, and an activation layer connected in sequence. The linear transformation layer projects high-dimensional input features into a lower-dimensional, more compact vector space, directly reducing the complexity of subsequent computations and facilitating the extraction of more fundamental feature representations. The batch normalization layer, by standardizing the feature distribution, effectively mitigates internal covariate shifts during training, accelerates model convergence, and improves the stability and consistency of feature representations. The use of rectified linear units as the activation function, due to their sparse activation, computational efficiency, and ability to alleviate the vanishing gradient problem, is well-suited for introducing necessary nonlinearity into features in signature recognition tasks, thereby enhancing the projection head network's ability to fit and express complex handwriting feature patterns. The sequential connection of these three layers ensures that features are normalized and nonlinearly enhanced during dimensionality reduction, ultimately enabling the projection head network to more reliably map input features into a lower-dimensional contrastive feature space that is more discriminative and conducive to subsequent unsupervised clustering algorithms for similarity measurement.

[0076] S303. Calculate the cross-entropy loss based on the identity category prediction results, and calculate the contrast loss based on the contrast feature vector.

[0077] S304. The cross-entropy loss and the contrastive loss are weighted and summed to obtain the total loss; the contrastive loss has a lower weight than the cross-entropy loss.

[0078] For example, the total loss = cross-entropy loss + alpha × contrastive loss, where alpha represents the weight of the contrastive loss and can be set to 0.1.

[0079] Setting the weight of the contrastive loss lower than that of the cross-entropy loss allows for collaborative optimization and a balance between primary and secondary tasks in multi-task learning. On one hand, by giving the cross-entropy loss a higher weight in the total loss, the core optimization objective of model training remains focused on improving the classification accuracy of identity categories. This clarifies that the primary task of model training is to learn feature representations that can accurately distinguish the identities of different writers, thus ensuring that the model meets its basic performance requirements as a classifier and effectively preventing the contrastive learning objective from interfering with or dominating the optimization direction due to excessive weight, which could lead to the model deviating from solving the core business problem of identity recognition. On the other hand, introducing the contrastive loss with a relatively low weight, while ensuring classification accuracy, serves as an auxiliary and regularized optimization constraint. This loss term is not intended to compete with the classification task, but rather to guide the model to consciously learn and optimize the intrinsic structure of its feature space while striving for correct classification. This can drive the model's output feature vectors to not only be correctly classified, but also to exhibit excellent geometric characteristics of compact intra-class and loose inter-class features within the feature space itself. This characteristic naturally gives the features powerful similarity measurement and clustering analysis capabilities.

[0080] Therefore, the aforementioned weight allocation strategy enables the deep feature vectors output by the trained feature extraction model to possess both high-precision identity discrimination and superior unsupervised clustering adaptability. This not only perfectly serves the current task of signature fraud detection but also allows the feature to be seamlessly and directly adapted to a wider range of business expansion scenarios, such as cross-database signature retrieval, historical handwriting comparison, and similarity matching, greatly enhancing the model's feature versatility and system scalability.

[0081] S305. Perform backpropagation based on the total loss; during backpropagation, keep the pre-trained multi-layer bottom-level encoding layer model parameters unchanged, and update the model parameters of the multi-layer top-level encoding layer, classification head network, and projection head network based on the gradient of this iteration.

[0082] In this step, cross-entropy loss is calculated based on the identity category prediction results. This ensures that the model is directly driven by strong and explicit identity-discriminating supervision signals during training, forcing the model to learn highly discriminative features that can accurately distinguish different writers. This is the foundation for the model to acquire core classification capabilities. Simultaneously, contrastive loss is calculated based on contrastive feature vectors, introducing a representation learning objective to the model. Its core is to drive the model to learn a feature space distribution where different signatures of the same person are close to each other, while signatures of different people are far apart. This directly optimizes the intra-class clustering and inter-class segregation of features, providing forward-looking preparation for subsequent unsupervised clustering tasks. The two losses are weighted and summed, with the contrastive loss having a lower weight than the cross-entropy loss. Ensuring the model's identity classification accuracy is the primary goal, guaranteeing its discriminative power. At the same time, the contrastive learning objective is introduced with a relatively low weight as an auxiliary regularization and guide, subtly optimizing the spatial structure of features without interfering with the core discrimination task, making it more cluster-friendly. By keeping the model parameters of the multi-layer bottom-level encoding layers unchanged during backpropagation, and only updating the model parameters of the multi-layer top-level encoding layers, the classification head network, and the projection head network, the hierarchical transfer learning strategy is precisely implemented. This not only efficiently reuses the general visual knowledge of the pre-trained multi-layer bottom-level encoding layers, but also focuses on fine-tuning the multi-layer top-level encoding layers, effectively preventing overfitting under small sample conditions and improving the model's generalization ability.

[0083] S104. Using a density-based unsupervised clustering algorithm, cluster the target signature images according to the depth feature vector of each target signature image to obtain at least one cluster.

[0084] Specifically, due to the existence of unknown true cluster numbers and low-quality samples caused by unclear scanning, incomplete writing, and severe contamination, algorithms that require a preset number of clusters (such as the K-means algorithm) were abandoned, and density-based unsupervised clustering algorithms (such as the DBSCAN algorithm) were selected instead.

[0085] Because the DBSCAN algorithm is highly sensitive to feature scale, directly using the original features (i.e., the depth feature vectors of each target signature image) will cause dimensions with larger numerical ranges to dominate in distance calculations, leading to distorted clustering results. To address this, methods such as StandardScaler are used to standardize all depth feature vectors: the mean and standard deviation are calculated separately for each feature dimension, and then the mean is subtracted from each feature value in that dimension before being divided by the standard deviation. This process transforms each feature dimension into a standard distribution with a mean of 0 and a variance of 1. This eliminates scale differences between dimensions, ensuring that all feature dimensions have fair weights in subsequent distance metrics, thus guaranteeing the fairness and accuracy of the clustering results.

[0086] The DBSCAN algorithm clusters target signature images based on their density in the feature space. It defines two key parameters: neighborhood radius and minimum number of samples. The DBSCAN algorithm divides data points into three categories: core points (points with a sufficient number of neighbors), boundary points (points falling within the neighborhood of core points but lacking sufficient neighbors), and noise points (points that are neither core nor boundary points). Starting from a core point, it continuously searches for samples whose density is achievable, thus forming a cluster and automatically discovering cluster structures of arbitrary shapes. This algorithm does not require a preset number of clusters and can automatically divide sample clusters based on feature density, perfectly adapting to business scenarios with an unknown number of categories. Simultaneously, the algorithm can identify abnormal noise samples in the dataset (such as blurry or incomplete signature images), closely matching the actual data distribution characteristics of signature images.

[0087] The clustering process does not rely on quantitative evaluation indicators. Instead, the core hyperparameter ε (neighborhood radius, eps) is fine-tuned through manual experience. Combined with the feature distribution characteristics of the signature images and the business visual verification results, the parameters are iteratively adjusted to the optimal value. Finally, the feature clusters are accurately divided on the target dataset, resulting in an ideal unsupervised clustering effect with highly clustered intra-class samples, significantly separated inter-class samples, and clustering results that closely match the actual business.

[0088] S105. For each cluster, perform text recognition on each target signature image within the cluster to obtain the signature text corresponding to each target signature image within the cluster.

[0089] Specifically, optical character recognition technology can be used to perform text recognition on the target signature images within each cluster, thereby extracting the signature text contained therein.

[0090] To efficiently and accurately extract name information from images, the PaddleOCR (version 3.3.1) engine can be used. It is an integrated tool library with a complete built-in OCR processing workflow (including text detection, text recognition, and orientation classification). For each cluster, it iterates through every target signature image within the cluster, calling the PaddleOCR engine to perform text detection and recognition, ultimately outputting a list of signature texts corresponding to that cluster, containing all the recognition results.

[0091] S106. If the signature text corresponding to each target signature image within a cluster contains at least two different names, then it is determined that there is a proxy signature behavior.

[0092] Specifically, in an ideal business scenario, each cluster formed based on the similarity of handwriting physical features should theoretically correspond to a unique signer identity; that is, the names written on all signature images within that cluster should be identical. If a significant difference exists, it indicates that proxy signing may have occurred.

[0093] In one implementation, before performing text recognition on each target signature image within a cluster, the relevant hyperparameters for text detection are adjusted according to at least one of the following parameter adjustment strategies: Lower the text region detection threshold (det_db_thresh); Lower the detection box threshold (det_db_box_thresh); Increase the unclipping ratio of the detection box (det_db_unclip_ratio); Enable morphological dilation operation (use_dilation).

[0094] Lowering the text region detection threshold allows handwriting pixels that exhibit weak signals due to light writing pressure, faint ink, unclear scanning, or background interference to still exceed the new threshold, even though their corresponding activation values ​​are low, and thus be initially marked as candidate text points. This directly improves the perception and recall capabilities for low-contrast, blurred, or fine-stroke areas, which is fundamental to ensuring that no potential signature handwriting areas are missed.

[0095] By lowering the detection box threshold, even if the overall average confidence of the initially detected text region is low due to discontinuous strokes, internal breaks, or local blurring, the region will still be identified as a valid text detection box and output as long as the new threshold is reached. This effectively addresses the uneven and discontinuous signals within strokes caused by common issues in handwritten signatures, such as ligatures, slashes, and pauses. It prevents the erroneous splitting of text regions belonging to the same signature into multiple boxes or their direct discarding due to poor local quality, thus ensuring that each signature image can be completely and correctly bounded, providing a structurally complete input for subsequent text recognition.

[0096] Increasing the anti-clipping ratio of the detection box allows the system to expand outward from the initially detected core text area when forming the final text detection box. This ensures that the generated text bounding box can fully cover the flowing, elegant stroke edges commonly found in handwritten signatures, avoiding character clipping caused by an overly tight box.

[0097] Enabling morphological dilation, by expanding the feature map, helps to connect internal breaks in characters caused by poor image quality, natural discontinuities in strokes, or writing style, linking discrete stroke pixels into coherent text regions. This is especially important for cursive writing or lightly written signatures.

[0098] By adjusting the relevant hyperparameters of text detection using at least one of the above parameter adjustment strategies, the text content in the signature image can be located and extracted as completely and accurately as possible.

[0099] Assuming that these four parameter adjustment strategies work together, they can more reliably locate the signature text region from signature images with complex backgrounds, blurred strokes, or noise interference, and form accurate text bounding boxes, thereby providing high-quality, complete text information input for subsequent high-precision text recognition and forged signature logic judgment.

[0100] This invention provides a method for identifying forged signatures. First, a set of unlabeled signature images is acquired and standardized preprocessed to ensure the uniformity of model input, a prerequisite for automated batch processing. Then, a feature extraction model based on a visual transformer and incorporating a hierarchical transfer learning strategy is employed. During training, the model parameters of the pre-trained multi-layered bottom-level encoding layers are frozen, while the parameters of the multi-layered top-level encoding layers are fine-tuned. This effectively reduces the risk of overfitting in small-sample scenarios and allows the model's training computation and gradient updates to focus on learning higher-order semantic features in the later multi-layered top-level encoding layers. Next, a density-based unsupervised clustering algorithm is used to cluster these feature vectors. Without pre-setting the number of categories, signatures with similar handwriting are automatically grouped into the same cluster. This accurately corresponds to the business reality that the same person may sign multiple documents, achieving a preliminary physical identification of the writer's identity. Furthermore, text recognition is performed for each cluster to extract the signature text, completing the conversion from handwriting images to semantic information. Finally, by checking whether at least two different names appear within the same handwriting cluster, a rigid logical contradiction is established between the handwriting originating from the same person and the signature belonging to different people. This contradiction directly and objectively reveals the act of signing on behalf of others, transforming the complex problem of handwriting identification into a calculable and verifiable consistency verification problem. Thus, without the need for any pre-stored signature templates or authenticity marking data, it enables efficient, automatic, and highly reliable identification of the specific violation pattern of one person signing on behalf of multiple people.

[0101] Please refer to Figure 5 , Figure 5 This is a schematic diagram of the signature identification device provided by the present invention. Figure 5 As shown, the device includes: The acquisition module 501 is used to acquire multiple original signature images to be identified, forming a set of unlabeled signature images; The preprocessing module 502 is used to perform standardization preprocessing on each original signature image in the signature image set to obtain the target signature image; The extraction module 503 is used to input the target signature images into the feature extraction model trained by transfer learning to extract the depth feature vector of each target signature image; wherein, the feature extraction model is built based on the visual transformer model; during model training, the model parameters of the pre-trained multi-layer bottom coding layer are frozen and the model parameters of the multi-layer top coding layer are fine-tuned. Clustering module 504 is used to cluster each target signature image based on the depth feature vector of each target signature image using a density-based unsupervised clustering algorithm to obtain at least one cluster. The recognition module 505 is used to perform text recognition on each target signature image within each cluster for each cluster, and obtain the signature text corresponding to each target signature image within the cluster. The determination module 506 is used to determine that there is a proxy signing behavior if the signature text corresponding to each target signature image in the cluster contains at least two different names.

[0102] In some embodiments, the signature identification device of the present invention can be implemented in a combination of hardware and software. As an example, the signature identification device of the present invention can be a processor in the form of a hardware decoding processor, which is programmed to execute the signature identification method of the present invention. For example, the processor in the form of a hardware decoding processor can be one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0103] The modules described in the embodiments of this invention can be implemented in software or hardware. The names of the modules are not, in some cases, limiting the scope of the module itself.

[0104] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-mentioned methods for identifying counterfeit signatures. That is, an electronic device according to an embodiment of the present invention may include, but is not limited to: a processor and a memory; the memory is used to store the computer program; the processor is used to execute the method for identifying counterfeit signatures shown in any embodiment of the present invention by calling the computer program.

[0105] In one alternative embodiment, an electronic device is provided, such as Figure 6 As shown, Figure 6The illustrated electronic device 4000 includes a processor 4001 and a memory 4003. The processor 4001 and the memory 4003 are connected, for example, via a bus 4002. Optionally, the electronic device 4000 may further include a transceiver 4004, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 4004 is not limited to one type, and the structure of the electronic device 4000 does not constitute a limitation on the embodiments of the present invention.

[0106] Processor 4001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. Processor 4001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0107] Bus 4002 may include a path for transmitting information between the aforementioned components. Bus 4002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 4002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The bus 4002 is represented by only one thick line, but this does not mean that there is only one bus or one type of bus.

[0108] The memory 4003 may be ROM (Read Only Memory) or other types of static storage devices capable of storing static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices capable of storing information and instructions, or EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0109] The memory 4003 stores application code (computer program) for executing the present invention, and its execution is controlled by the processor 4001. The processor 4001 executes the application code stored in the memory 4003 to implement the content shown in the foregoing method embodiments.

[0110] Among them, electronic devices can also be terminal devices, which can be any device that can install applications, including at least one of smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, smart TVs, and smart in-vehicle devices.

[0111] It should be noted that, Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0112] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-described methods for identifying counterfeit signatures.

[0113] Alternatively, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, a floppy disk, and an optical data storage device, etc.

[0114] In an exemplary embodiment, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform the aforementioned signature identification.

[0115] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0116] It should be understood that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0117] The computer-readable storage medium provided in this invention can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EEPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0118] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the method shown in the above embodiments.

[0119] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0120] It should be noted that the terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and represent a limitation on a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of this application described herein can be implemented in an order other than that shown or described.

[0121] Those skilled in the art will recognize that this invention can be implemented as a system, method, or computer program product. Therefore, this invention can be specifically implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, this invention can also be implemented as a computer program product contained in one or more computer-readable media, which includes computer-readable program code.

[0122] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for identifying proxy signatures, characterized in that, include: Multiple raw signature images to be identified are obtained to form an unlabeled signature image set; For each original signature image in the signature image set, a standardization preprocessing is performed on the original signature image to obtain the target signature image; Each of the target signature images is input into a feature extraction model trained by transfer learning to extract the deep feature vector of each target signature image; wherein, the feature extraction model is constructed based on a visual transformer model; during model training, the model parameters of the pre-trained multi-layer bottom coding layer are frozen, and the model parameters of the multi-layer top coding layer are fine-tuned. Using a density-based unsupervised clustering algorithm, the target signature images are clustered according to their depth feature vectors to obtain at least one cluster. For each cluster, text recognition is performed on each target signature image within the cluster to obtain the signature text corresponding to each target signature image within the cluster. If the signature text corresponding to each target signature image within the cluster contains at least two different names, then it is determined that there is a proxy signature.

2. The method for identifying proxy signatures according to claim 1, characterized in that, The feature extraction model includes a basic feature extraction network and a dual-output head network; The basic feature extraction network is a visual transformer model with the original classification head removed; The dual-output head network includes a classification head network and a projection head network. The classification head network is used to predict the writer's identity category based on the output feature vector of the basic feature extraction network, and the projection head network is used to map the output feature vector of the basic feature extraction network to a contrastive feature space for unsupervised clustering.

3. The method for identifying proxy signatures according to claim 2, characterized in that, The classification head network includes a compression layer, an activation layer, and a random deactivation layer; wherein, the compression layer is used for progressive feature dimension compression, the activation layer uses Gaussian error linear units as activation functions, and the random deactivation layer is used to randomly discard some output features of the activation layer with a preset probability during training.

4. The method for identifying proxy signatures according to claim 2, characterized in that, The projection head network includes a linear transformation layer, a batch normalization layer, and an activation layer connected in sequence; wherein, the activation layer uses a rectified linear unit as the activation function.

5. The method for identifying proxy signatures according to any one of claims 2 to 4, characterized in that, The feature extraction model is trained by transfer learning through the following iterative steps until the model converges. Obtain signature image samples from multiple different writers to form a training dataset; The training dataset is input into the feature extraction model for forward propagation to obtain the identity category prediction result and the contrast feature vector. The cross-entropy loss is calculated based on the identity category prediction results, and the contrast loss is calculated based on the contrast feature vector. The total loss is obtained by weighted summation of the cross-entropy loss and the contrastive loss, wherein the weight of the contrastive loss is lower than the weight of the cross-entropy loss. Backpropagation is performed based on the total loss; wherein, during backpropagation, the model parameters of the pre-trained multi-layer bottom coding layer are kept unchanged, and the model parameters of the multi-layer top coding layer, the classification head network, and the projection head network are updated based on the gradient of this iteration.

6. The method for identifying proxy signatures according to claim 1, characterized in that, Before performing text recognition on each of the target signature images within the cluster, the relevant hyperparameters of text detection are adjusted according to at least one of the following parameter adjustment strategies: Lower the text region detection threshold; Lower the detection frame threshold; Increase the reverse clipping ratio of the detection frame; Enable morphological expansion operation.

7. The method for identifying proxy signatures according to claim 1, characterized in that, The standardization preprocessing of the original signature image to obtain the target signature image includes: The original signature image is decoded and standardized to obtain a bitmap format image; The bitmap format image is geometrically corrected and sized to obtain a standard-sized image; The standard-sized image is subjected to denoising and feature enhancement processing to obtain the target signature image.

8. A signature identification device, characterized in that, include: The acquisition module is used to acquire multiple raw signature images to be identified, forming a set of unlabeled signature images; The preprocessing module is used to perform standardization preprocessing on each original signature image in the signature image set to obtain a target signature image; The extraction module is used to input each of the target signature images into a feature extraction model trained by transfer learning, and extract the depth feature vector of each of the target signature images; wherein, the feature extraction model is constructed based on a visual transformer model; during model training, the model parameters of the pre-trained multi-layer bottom coding layer are frozen, and the model parameters of the multi-layer top coding layer are fine-tuned. The clustering module is used to cluster the target signature images according to the depth feature vector of each target signature image using a density-based unsupervised clustering algorithm to obtain at least one cluster. The recognition module is used to perform text recognition on each target signature image within each cluster for each cluster, so as to obtain the signature text corresponding to each target signature image within the cluster. The determination module is used to determine that there is a proxy signing behavior if the signature text corresponding to each of the target signature images in the cluster contains at least two different names.

9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that the processor, when executing the computer program, implements the signature identification method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the proxy signature identification method as described in any one of claims 1 to 7.

Images