Signature recognition method and device, storage medium, electronic equipment and product

Abstract

The invention discloses a signature recognition method and device, a storage medium, electronic equipment and a product, and the method comprises the steps: carrying out the extraction of a signature region in a file through a detection sub-model, and obtaining a first image; performing enhancement processing on the first image by using the preprocessing sub-model to obtain a second image; and obtaining a target signature feature in the second image through the identification sub-model, and determining the similarity between the target signature feature and a preset signature feature in a database, so as to determine a target identification result of the signature region based on the similarity. By adopting the technical scheme, the problem that the signature detection and recognition precision cannot be effectively improved when a complex background and a low-quality image are processed is solved. And furthermore, through cooperative work of the detection sub-model, the enhancement sub-model and the identification sub-model, a signature region is accurately positioned from the document, the image quality is optimized, and signature features are compared and analyzed, so that the authenticity of the signature is efficiently judged.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and more specifically, to a signature recognition method, apparatus, storage medium, electronic device, and product. Background Technology

[0002] In today's digital office and identity verification field, handwritten signatures, as a traditional yet indispensable personal identification tool, still hold a crucial position. Whether for contract signing, financial document confirmation, or legal document authorization, handwritten signatures provide an intuitive and personalized verification method to ensure the authenticity and validity of documents. However, with the continuous evolution of document processing methods and the increasing prevalence of technological applications, traditional signature recognition systems are facing unprecedented challenges. These challenges mainly include processing low-quality images, dealing with complex background interference, and effectively identifying forged signatures. Existing technologies mostly rely on template matching on static images. Although this method can meet basic recognition needs in some situations, its limitations become apparent when faced with blurry, noisy, or diverse signature styles. Quality defects in signature images, background complexity, and the diversity of personalized signature styles lead to a high false recognition rate in practical applications.

[0003] Therefore, in related technologies, there is still no effective solution to the problem of not being able to effectively improve the accuracy of signature detection and recognition when dealing with complex backgrounds and low-quality images. Summary of the Invention

[0004] This application provides a signature recognition method, apparatus, storage medium, electronic device, and product to at least solve the problem in the related art that the accuracy of signature detection and recognition cannot be effectively improved when processing complex backgrounds and low-quality images.

[0005] According to one embodiment of the present application, a signature recognition method is provided, comprising: extracting a signature region in a file using a detection sub-model to obtain a first image; enhancing the first image using a preprocessing sub-model to obtain a second image; obtaining target signature features in the second image using a recognition sub-model, and determining the similarity between the target signature features and preset signature features in a database, so as to determine the target recognition result of the signature region based on the similarity.

[0006] In an exemplary embodiment, obtaining the target signature features in the second image by recognizing a sub-model includes: performing multi-level processing on the second image to generate multiple feature maps; extracting features from the multiple feature maps to obtain multiple types of features, wherein the multiple types of features include at least one of the following: texture features of the second image, shape features of the second image, and handwriting features of the second image; and fusing the multiple types of features to obtain the target signature features.

[0007] In an exemplary embodiment, determining the similarity between a target signature feature and a preset signature feature in a database, and determining the target recognition result of the signature region based on the similarity, includes: calculating the similarity between the target signature feature and the preset signature feature in the database according to a preset feature matching algorithm to obtain a similarity score; and determining the target recognition result of the signature region according to the relationship between the similarity score and a preset similarity threshold.

[0008] In an exemplary embodiment, determining the target recognition result of the signature region based on the relationship between the similarity score and a preset similarity threshold includes: determining that the target recognition result of the signature region is a genuine signature when the similarity score is greater than or equal to the preset similarity threshold; and determining that the target recognition result of the signature region is not a genuine signature when the similarity score is less than the preset similarity threshold.

[0009] In an exemplary embodiment, after determining the similarity between the target signature feature and the preset signature features in the database to determine the target recognition result of the signature region based on the similarity, the method further includes: obtaining a real signature sample corresponding to the file; processing the real signature sample by generating a sub-model to obtain a forged signature sample; training a preset anti-counterfeiting sub-model based on the forged signature sample and the real signature sample to obtain a target anti-counterfeiting sub-model; and verifying the target recognition result based on the target anti-counterfeiting sub-model.

[0010] In an exemplary embodiment, after verifying the target identification result based on the target anti-counterfeiting sub-model, the method further includes: collecting the verification results generated by the target anti-counterfeiting sub-model within a preset period; classifying the verification results to obtain a set of genuine signature samples and a set of forged signature samples; inputting the set of genuine signature samples into the target anti-counterfeiting sub-model for iterative training, and recording abnormal signature sample data during the iterative training process.

[0011] According to another aspect of the embodiments of this application, a signature recognition device is also provided, comprising: a first processing module, configured to extract a signature region in a file through a detection sub-model to obtain a first image; a second processing module, configured to enhance the first image using a preprocessing sub-model to obtain a second image; and a determination module, configured to obtain target signature features in the second image through a recognition sub-model, and determine the similarity between the target signature features and preset signature features in a database, so as to determine the target recognition result of the signature region based on the similarity.

[0012] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, wherein a computer program is stored in the computer-readable storage medium, and the computer program is configured to execute the above-described signature recognition method at runtime.

[0013] According to another aspect of the embodiments of this application, an electronic device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the signature recognition method described above through the computer program.

[0014] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program, wherein the above-described signature recognition method is executed by a processor.

[0015] In this embodiment, a first image is obtained by extracting the signature region in a document using a detection sub-model; a second image is obtained by enhancing the first image using a preprocessing sub-model; and a second image is obtained by acquiring the target signature features in the second image using a recognition sub-model, and determining the similarity between the target signature features and preset signature features in the database, thereby determining the target recognition result of the signature region based on the similarity. This technical solution solves the problem of not being able to effectively improve the accuracy of signature detection and recognition when processing complex backgrounds and low-quality images. Furthermore, through the collaborative work of the three sub-models—detection, enhancement, and recognition—the signature region is accurately located in the document, image quality is optimized, and signature features are compared and analyzed, thereby efficiently determining the authenticity of the signature. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the hardware environment for a signature recognition method according to an embodiment of this application;

[0019] Figure 2 This is a flowchart of a signature recognition method according to an embodiment of this application;

[0020] Figure 3 This is a schematic diagram of a deep learning-based integrated signature detection and anti-counterfeiting system according to an embodiment of this application;

[0021] Figure 4 This is a flowchart of a deep learning-based signature detection and anti-counterfeiting method according to an embodiment of this application;

[0022] Figure 5 This is a structural block diagram of a signature recognition device according to an embodiment of this application. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, apparatus, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, apparatus, or devices.

[0025] According to one aspect of the embodiments of this application, a signature recognition method is provided. This signature recognition method is widely used in whole-house intelligent digital control application scenarios such as smart homes, smart home ecosystems, and intelligencehouse ecosystems. Optionally, in this embodiment, the above-mentioned signature recognition method can be applied to, for example... Figure 1 The hardware environment shown consists of terminal device 102 and server 104. Figure 1 This is a schematic diagram of the hardware environment for a signature recognition method according to an embodiment of this application, such as... Figure 1 As shown, server 104 is connected to terminal device 102 via a network and can be used to provide services (such as application services) to the terminal or clients installed on the terminal. A database can be set up on the server or independently of the server to provide data storage services for server 104. Cloud computing and / or edge computing services can be configured on the server or independently of the server to provide data processing services for server 104.

[0026] The aforementioned network may include, but is not limited to, at least one of the following: wired network, wireless network. The aforementioned wired network may include, but is not limited to, at least one of the following: wide area network, metropolitan area network, local area network. The aforementioned wireless network may include, but is not limited to, at least one of the following: Wi-Fi (Wireless Fidelity), Bluetooth. The terminal device 102 may not be limited to PC, mobile phone, tablet computer, smart air conditioner, smart range hood, smart refrigerator, smart oven, smart stove, smart washing machine, smart water heater, smart washing equipment, smart dishwasher, smart projector, smart TV, smart clothes rack, smart curtains, smart audio-visual equipment, smart socket, smart speaker, smart speaker box, smart fresh air equipment, smart kitchen and bathroom equipment, smart bathroom equipment, smart robot vacuum cleaner, smart window cleaning robot, smart mopping robot, smart air purifier, smart steam oven, smart microwave oven, smart water heater, smart air purifier, smart water dispenser, smart door lock, etc.

[0027] This embodiment provides a signature recognition method applied to the aforementioned terminal device. Figure 2 This is a flowchart of a signature recognition method according to an embodiment of this application, which includes the following steps:

[0028] Step S202: Extract the signature region in the file using the detection sub-model to obtain the first image;

[0029] Optionally, during the construction and training of the detection sub-model, a signature annotation dataset is first collected, containing thousands of images of various documents with handwritten signatures. Each image has a manually annotated bounding box and category label (e.g., "signature") for the signature region. Next, this dataset is used to train the YOLO (You Only Look Once) model, enabling it to learn to identify and locate signature regions from document backgrounds. Specifically, the collected raw images are preprocessed, including resizing, image enhancement, and format conversion. The YOLO model architecture is built, selecting an appropriate pre-trained model as initial weights. The YOLO model includes a backbone network for feature extraction and a detection head for predicting bounding box coordinates and category probabilities. Loss functions are defined, typically including classification loss, bounding box regression loss, and confidence loss, to guide model optimization. The preprocessed data is fed into the model for training. The loss is calculated through forward propagation, and the model weights are updated iteratively during backpropagation until convergence. During training, model performance is monitored, including precision, recall, and mean precision metrics, to ensure the model detects signature regions accurately and comprehensively. After initial training, the model's hyperparameters (such as learning rate, batch size, and regularization terms) are tuned to find the optimal performance configuration. Subsequently, the model is tested using an independent dataset not used in training to evaluate its generalization ability on unknown samples. Once the model achieves satisfactory detection performance, it is deployed to the signature verification system. In the above embodiment, the input to the detection sub-model is a document image containing potential signature regions, and the output is a first image containing the coordinates of the signature region's bounding box and category information. This entire process not only demonstrates the superiority of deep learning in solving complex visual tasks but also highlights the model's ability to continuously optimize and adapt to new scenarios through large-scale training data, laying a solid foundation for efficient and accurate handwritten signature detection.

[0030] Step S204: Enhance the first image using a preprocessing sub-model to obtain the second image;

[0031] Optionally, the aforementioned preprocessing sub-model is a combination of CLAHE (Contrast Limited Adaptive Histogram Equalization), DnCNN (Deep Convolutional Neural Network for Image Denoising), and ESRGAN (Enhanced Super-Resolution Generative Adversarial Networks). It is assumed that after processing by the detection sub-model, a first image of the signature region is obtained, but this image is of low quality, exhibiting insufficient contrast, noise, and blurring. The preprocessing sub-model is used for training and application to improve image quality. The specific training process involves collecting a set of signature image pairs containing high-resolution and low-resolution versions, with the high-resolution image as the target and the low-resolution image as the input. These image pairs are used to train the ESRGAN model, enabling it to learn how to generate high-resolution images from low-resolution images. Before the images are fed into ESRGAN, contrast enhancement is performed using CLAHE, followed by noise removal using DnCNN. After model training, the overall performance of the preprocessing sub-model is evaluated using an independent test set, including contrast enhancement, denoising, and super-resolution reconstruction quality. Once the preprocessing sub-model is trained, it can be deployed in the signature detection and anti-counterfeiting system to perform real-time enhancement processing on the first image, generating a second image with optimized quality. In the above embodiment, the input of the preprocessing sub-model is the first image with the signature region located by the detection sub-model, and the output is the second image processed by contrast enhancement, denoising, and super-resolution reconstruction. Through these steps, the preprocessing sub-model can ensure that even with poor original image quality, it can provide high-quality, detailed signature images for subsequent system operations, thereby improving the performance and reliability of the entire system.

[0032] Step S206: Obtain the target signature features in the second image through the recognition sub-model, and determine the similarity between the target signature features and the preset signature features in the database, so as to determine the target recognition result of the signature region based on the similarity.

[0033] Optionally, if an automated signature verification system for bank checks is developed, when a user submits a check containing a handwritten signature, the system first locates the signature region using a detection sub-model, and then feeds the image of that region into a recognition sub-model. The recognition sub-model, such as ResNet-50 (50-layer Deep Residual Network), extracts features from the signature region image to form a signature feature vector. Next, this feature vector is compared with the corresponding customer signature template in the bank's database using cosine similarity calculation. Based on the calculated similarity value, and through a dynamically adjusted recognition threshold, it is determined whether the signature matches a preset signature.

[0034] Through the above steps, the signature region in the file is extracted and processed using a detection sub-model to obtain a first image; the first image is enhanced using a preprocessing sub-model to obtain a second image; the target signature features in the second image are obtained using a recognition sub-model, and the similarity between the target signature features and preset signature features in the database is determined, so as to determine the target recognition result of the signature region based on the similarity. This technical solution solves the problem of not being able to effectively improve the accuracy of signature detection and recognition when processing complex backgrounds and low-quality images. Furthermore, through the collaborative work of the three sub-models of detection, enhancement, and recognition, the signature region is accurately located in the document, image quality is optimized, and signature features are compared and analyzed, thereby efficiently determining the authenticity of the signature.

[0035] In an exemplary embodiment, obtaining the target signature features in the second image by recognizing a sub-model includes: performing multi-level processing on the second image to generate multiple feature maps; extracting features from the multiple feature maps to obtain multiple types of features, wherein the multiple types of features include at least one of the following: texture features of the second image, shape features of the second image, and handwriting features of the second image; and fusing the multiple types of features to obtain the target signature features.

[0036] Optionally, a deep learning-based signature recognition system can be used to verify the authenticity of a signature on an important document. First, the second image (the signature image to be verified) is preprocessed, including grayscale conversion, size normalization, and enhancement (such as contrast and brightness adjustment) to ensure image quality meets recognition requirements. A deep convolutional neural network (such as ResNet-50) is used to perform multi-level convolutional processing on the preprocessed image, generating a series of feature maps with different levels of abstraction. These feature maps cover multi-level information from local details (such as the thickness and curvature of strokes) to global structure (such as the overall shape and layout of the signature). Different types of features, such as texture, shape, and handwriting, are extracted from the generated feature maps. For example, texture features may involve the surface structure of strokes, such as the distribution of ink; shape features focus on the outline and internal structure of the signature; and handwriting features emphasize the fluidity and continuity of the writing lines. The extracted features of various types are fused to construct a comprehensive signature feature vector. Through the above implementation steps, the complex challenges in signature recognition, such as the diversity of handwriting styles and the influence of the writing environment, can be effectively addressed.

[0037] In an exemplary embodiment, determining the similarity between a target signature feature and a preset signature feature in a database, and determining the target recognition result of the signature region based on the similarity, includes: calculating the similarity between the target signature feature and the preset signature feature in the database according to a preset feature matching algorithm to obtain a similarity score; and determining the target recognition result of the signature region according to the relationship between the similarity score and a preset similarity threshold.

[0038] Optionally, develop signature recognition software for online contract signing. This software needs to determine within seconds whether a user-submitted electronic signature matches a genuine signature associated with their account. The software first uses a recognition sub-model to extract features from the user-submitted electronic signature (second image), converting it into a vector containing texture, shape, and handwriting features—the target signature features. Simultaneously, it retrieves preset signature features associated with the user account from a pre-set database. Next, the software employs an efficient feature matching algorithm based on deep learning's cosine similarity calculation method to quantify the distance between the target signature features and the preset signature features. The calculated similarity score reflects the degree of consistency between the two feature vectors. The relationship between the similarity score and a similarity threshold is used to distinguish whether the signature in the signature area is a genuine signature.

[0039] In an exemplary embodiment, determining the target recognition result of the signature region based on the relationship between the similarity score and a preset similarity threshold includes: determining that the target recognition result of the signature region is a genuine signature when the similarity score is greater than or equal to the preset similarity threshold; and determining that the target recognition result of the signature region is not a genuine signature when the similarity score is less than the preset similarity threshold.

[0040] Optionally, a large financial institution's online banking system needs to verify signatures on various files uploaded by users to ensure the legitimacy of operations. First, using a recognition sub-model, such as a ResNet-50 deep convolutional neural network, features of the signature region to be verified are extracted from the user-uploaded photo. Then, the cosine similarity score between these features and the signature feature template saved during user registration is calculated. When the calculated similarity score is greater than or equal to a similarity threshold (assumed to be 0.85), the signature region is automatically determined to be a genuine signature, allowing the banking transaction or file operation to continue; conversely, if the similarity score is less than 0.85, the operation is blocked, and the user is notified that the signature verification has failed. Through the above embodiment, this similarity threshold-based decision-making mechanism not only enhances security but also ensures operational convenience and efficiency.

[0041] In an exemplary embodiment, after determining the similarity between the target signature feature and the preset signature features in the database to determine the target recognition result of the signature region based on the similarity, the method further includes: obtaining a real signature sample corresponding to the file; processing the real signature sample by generating a sub-model to obtain a forged signature sample; training a preset anti-counterfeiting sub-model based on the forged signature sample and the real signature sample to obtain a target anti-counterfeiting sub-model; and verifying the target recognition result based on the target anti-counterfeiting sub-model.

[0042] Optionally, a financial institution is developing a deep learning-based signature anti-counterfeiting system for verifying electronic contract signatures. It retrieves a large number of verified genuine signature samples from a database, potentially originating from the institution's customers and employees, covering diverse signature styles and writing environments. Using a generative sub-model, based on these genuine signature samples, it generates a series of forged signature samples that are highly similar to genuine signatures but with subtle differences by simulating different forgery processes. This includes altering the thickness, slant, and stroke order of the handwriting to simulate common forgery behaviors. The genuine and forged signature samples are then input into a pre-defined anti-counterfeiting sub-model for training. Through deep learning algorithms, the anti-counterfeiting sub-model learns how to distinguish between these two types of samples, ultimately resulting in an optimized target anti-counterfeiting sub-model. When the system receives an electronic contract signature to be verified, the anti-counterfeiting sub-model performs a deep analysis of the signature to detect any signs of forgery. If the identification results indicate that the signature may be forged, further measures are taken, such as requiring manual review and secondary authentication with the signer, to ensure the security and legitimacy of the document.

[0043] In an exemplary embodiment, after verifying the target identification result based on the target anti-counterfeiting sub-model, the method further includes: collecting the verification results generated by the target anti-counterfeiting sub-model within a preset period; classifying the verification results to obtain a set of genuine signature samples and a set of forged signature samples; inputting the set of genuine signature samples into the target anti-counterfeiting sub-model for iterative training, and recording abnormal signature sample data during the iterative training process.

[0044] Optionally, a deep learning-based signature verification system has been deployed in banks' electronic signature verification services. The system includes a target anti-counterfeiting sub-model for detecting the authenticity of signatures. Each day, this model processes thousands of signature images, outputting recognition decisions and their corresponding confidence scores. The system has a pre-set monthly cycle, automatically collecting and storing all signature images processed by the target anti-counterfeiting sub-model and their recognition results, including metadata such as the recognition score for each image, genuine / counterfeit labels, and processing date. At the end of the month, the system runs a pre-defined classification algorithm to classify the collected verification results into a set of genuine signature samples and a set of counterfeit signature samples. Then, a new iterative training is conducted using the genuine signature sample set to optimize the parameters of the target anti-counterfeiting sub-model and improve its recognition accuracy. During training, whenever the model's confidence score for a signature classification falls below a pre-set threshold, or when misclassification is detected during the testing phase, the system automatically records these abnormal signature sample data, including their images, classification results, and confidence scores, for subsequent deep feature analysis and model parameter adjustment. By implementing the above steps, not only are operational delays caused by bank failures reduced, but user trust in electronic signature services is also significantly improved.

[0045] To better understand the process of the above signature recognition method, the following description of the signature recognition method flow is further illustrated with reference to optional embodiments, but it is not intended to limit the technical solution of the embodiments of this application.

[0046] Among related technologies, existing signature detection and anti-counterfeiting technologies still have significant shortcomings in terms of adaptability to complex scenarios, recognition accuracy, anti-counterfeiting capabilities, system integration, and data generalization capabilities. With the diversification of application scenarios and the advancement of counterfeiting technologies, there is an urgent need to introduce more intelligent, dynamic, and adaptive technologies to address these challenges and improve the overall performance and reliability of the system.

[0047] To address the aforementioned issues, this application proposes a deep learning-based signature detection and anti-counterfeiting method. This method integrates advanced image processing techniques and deep learning models to achieve high-precision detection, recognition, and anti-counterfeiting of handwritten signatures. First, multi-scale feature extraction is performed using the YOLO model and the Feature Pyramid Network (FPN) model to ensure accurate signature region localization even against complex backgrounds. Subsequently, a series of adaptive image enhancement techniques, such as CLAHE, DnCNN, and ESRGAN, are used to improve the quality of the signature image, creating favorable conditions for subsequent recognition and anti-counterfeiting. In the signature recognition stage, a ResNet-50 deep convolutional neural network is used to extract deep-level features of the signature, and cosine similarity matching is used for efficient verification. At the anti-counterfeiting level, a Generative Adversarial Network (GAN) is used to generate counterfeit signature samples, combined with Gabor filters to analyze signature texture features, accurately distinguishing genuine from counterfeit signatures and significantly improving the accuracy of anti-counterfeiting detection. Finally, all modules are optimized and integrated through a multi-task learning strategy to form an integrated workflow, enhancing the system's real-time response capability and adaptability to diverse scenarios.

[0048] Optional, Figure 3 This is a schematic diagram of a deep learning-based integrated signature detection and anti-counterfeiting system according to an embodiment of this application. The system includes at least: a signature detection module 32, a signature recognition module 34, and a signature anti-counterfeiting module 36.

[0049] Optional, signature detection module 32. Uses a deep learning model (YOLO model) to detect signature regions. Shares a feature extraction layer with the signature recognition and anti-counterfeiting module to reduce computational costs.

[0050] Optional, signature recognition module 34. Feature extraction is performed using a pre-trained deep convolutional neural network ResNet-50. Multi-task learning is applied, allowing the recognition module to share features from the detection module while optimizing recognition accuracy.

[0051] Optional, signature anti-counterfeiting module 36. Forged signature samples are generated using GAN to train the anti-counterfeiting model. Gabor filters are used to analyze the signature texture, and overall performance is improved by combining multi-task learning (allowing the anti-counterfeiting module to share feature representations with other modules) and joint optimization.

[0052] Optional, Figure 4 This is a flowchart of a deep learning-based signature detection and anti-counterfeiting method according to an embodiment of this application, which specifically includes the following steps:

[0053] Step 1: Multi-scale Automatic Signature Detection. The YOLO model is used for signature region detection, combined with FPN to extract multi-scale features, ensuring accurate detection of signatures of different sizes and resolutions even in complex backgrounds and with diverse signature styles. The YOLO model quickly identifies potential signature regions due to its fast detection capabilities, while FPN constructs a feature pyramid to ensure the capture of key signature details at different scales.

[0054] Optionally, the YOLO model treats object detection as a regression problem, directly predicting the location and category of bounding boxes on the input image, thus significantly improving detection speed. Its basic detection process can be divided into two main components: Backbone (feature extraction) and Detection Head, as follows:

[0055] YOLO=Backbone(Feature Extraction)+Detection Head;

[0056] Optionally, the Backbone is the initial part of the YOLO model, responsible for extracting useful features from the input image. Typically, this part uses a pre-trained deep convolutional neural network (such as ResNet-50) as its foundation. These networks are pre-trained on large-scale image classification datasets, thus enabling them to effectively extract high-level abstract features from images. The specific process is as follows:

[0057] The original image is input into the YOLO model. The image size can be fixed or dynamically scaled to suit the model's needs. Through a series of convolutional, pooling, and fully connected layers, Backbone extracts image features layer by layer, with each layer building a more abstract and complex feature representation. This ultimately produces a feature map containing multi-level feature information of the image, providing rich feature input for subsequent detection heads.

[0058] Optionally, based on the feature map generated by the Backbone, the Detection Head is responsible for predicting the location and category of the target. This part divides the feature map into a grid, predicting the presence or absence of the target within each grid cell. YOLO divides the final feature map into S×S grids, with each grid cell responsible for predicting the target located within it. The specific process is as follows:

[0059] YOLO divides the final feature map into S×S grids, with each grid cell responsible for predicting the objects within it. Each grid cell predicts multiple bounding boxes, their confidence scores, and the probability of their respective class. The confidence score represents the probability that the bounding box contains an object, as well as the accuracy of the bounding box prediction. The class probability indicates the likelihood that the object within the bounding box belongs to any particular class. Furthermore, the detection head performs Non-Maximum Suppression (NMS) to filter out overlapping bounding boxes, retaining only the most probable detection results.

[0060] Optionally, in the YOLO model, the loss function is the core of optimizing model parameters during training. It ensures that the model can simultaneously learn the classification accuracy, location prediction accuracy, and bounding box confidence of the target. The YOLO loss function consists of three parts: classification loss (L... cls ), positioning loss (L) loc ) and confidence loss (L conf The details are as follows:

[0061] (1) Classification loss (L) cls The classification loss measures the accuracy with which the model predicts the category of each bounding box. In YOLO, each grid cell predicts multiple bounding boxes, and each bounding box, in addition to predicting coordinates and size, also predicts a category vector representing the probability that the detected object belongs to each category. The classification loss is typically calculated using the cross-entropy loss function, which effectively measures the difference between the model's predicted probabilities and the actual categories.

[0062] (2) Positioning loss (L) loc The localization loss is responsible for training the model to predict the center coordinates and size of the bounding box. In the YOLO model, the position and size of the bounding box are represented by the coordinates of the top-left corner of a relative grid cell and the size of the grid cell. The localization loss typically uses mean squared error to measure the difference between the predicted and actual values, ensuring that the model can accurately predict the bounding box position of the target.

[0063] (3) Confidence loss (L) conf The confidence loss is unique to the YOLO model and is used to train the model to estimate the probability that an object is contained within the bounding box. When predicting the bounding box, the YOLO model simultaneously predicts a confidence value, which reflects the model's belief that an object actually exists within the bounding box. The confidence loss is also calculated using mean squared error or cross-entropy loss to ensure that the model can correctly estimate the validity of the bounding box.

[0064] Optionally, YOLO's total loss function is a weighted sum that includes the classification loss, localization loss, and confidence loss mentioned above. The formula for the total loss function is as follows:

[0065] L total =L cls +λ1 L loc +λ2 L conf ;

[0066] Here, λ1 and λ2 are hyperparameters used to adjust the model's balance between classification, localization, and confidence estimation. By adjusting these weights, a suitable balance between accuracy and detection speed is found, ensuring that the model achieves optimal detection performance in various scenarios.

[0067] Step Two: Image Enhancement and Normalization. This includes the following steps:

[0068] Step 1: Enhance the contrast of the signature image using the CLAHE technique. The CLAHE processing flow can be described by the following formula:

[0069] I enhanced =CLAHE(I original (clipLimit, tileGridSize);

[0070] Among them, I original This represents the original signature image; clipLimit is the contrast limit threshold, used to control the intensity of local contrast enhancement to prevent excessive contrast enhancement from causing image distortion; tileGridSize determines the size of the segmented image blocks. By setting an appropriate tileGridSize, the contrast enhancement effect can be affected. Smaller block sizes can capture details better, but may introduce more noise; larger block sizes help reduce noise, but may weaken the detail enhancement effect.

[0071] Step 2: Use DnCNN for image denoising. DnCNN effectively removes noise through a deep convolutional network while preserving image details, ensuring that signature features are clearly visible in the denoised image. The basic formula of DnCNN is as follows:

[0072] I denoised =I noisy -f(I noisy );

[0073] Among them, I noisy It is a signature image containing noise, I denoised It is the image after denoising, and f( ) represents the mapping function of the DnCNN denoising network, which, through learning, can accurately estimate and remove noise components in an image.

[0074] Step 3: Perform super-resolution reconstruction using ESRGAN to generate a high-quality, normalized signature image input. The generation function used by ESRGAN is as follows:

[0075] I SR =G(I LR );

[0076] Among them, G( ) is a generative network, I LR It is a low-resolution image, I SR It is a super-resolution image.

[0077] Step 3: High-precision signature feature extraction and matching. A ResNet-50 deep convolutional neural network model is used for signature feature extraction, and a cosine similarity-based feature matching method is used for signature recognition and authentication. ResNet-50 deep convolutional neural network feature extraction is an advanced feature extraction technique that can capture and summarize deep features in signature images. The extracted features are then compared with a pre-stored signature template library using a cosine similarity matching algorithm. This not only ensures accurate matching but also effectively handles signatures of different styles and languages, ensuring high-precision authentication even in highly variable environments, closely linking the image processing and anti-counterfeiting detection stages.

[0078] Optionally, the residual block formula for ResNet-50 is as follows:

[0079] y=F(x,W i )+x;

[0080] Where F(x,W) i ) represents the residual mapping obtained after passing through a series of convolutional layers and activation functions, where x is the data input to the residual block and y is the output of the residual block.

[0081] Optionally, during feature matching, cosine similarity can be used to measure the similarity between feature vectors:

[0082] ;

[0083] Here, A and B are two feature vectors, A·B represents their dot product (inner product), and ||A|| and ||B|| are their moduli (i.e., the lengths of the vectors). Cosine similarity measures the cosine of the angle between two vectors, ranging from -1 to 1, with values ​​closer to 1 indicating higher similarity.

[0084] Step 4: Deep Anti-counterfeiting Detection of Forged Signatures. A GAN-based forged signature detection model is used to generate forged signature samples, and a Gabor filter is combined for multi-scale analysis of signature texture. Specifically, a forgery detection mechanism is constructed using generative adversarial networks (GANs) and Gabor filters. Game-theoretic training between the GAN generator and discriminator networks generates a series of highly realistic forged signature samples. These samples are used to train the anti-counterfeiting model, enabling it to distinguish forged signatures that are almost indistinguishable from genuine ones. The Gabor filter can be seen as a "magnifying glass," specifically designed to detect texture details at specific directions and scales in an image. It effectively captures important texture information in the signature, thus helping the system more accurately distinguish between genuine and forged signatures. The combined application of GANs and Gabor filters significantly improves the accuracy of forged signature detection, thereby enhancing the overall anti-counterfeiting capability of the system.

[0085] Optionally, the basic mathematical formula for GAN is as follows:

[0086] ;

[0087] Where G represents the generator network, whose purpose is to generate samples that look like real data; D represents the discriminator network, whose role is to distinguish as much as possible which samples were generated by the generator network and which samples came from the real training data; V(D, G) is the value function between the discriminator and the generator; p data (x) represents the true data distribution, p z (z) represents the noise distribution of the input to the generator network; E represents the expected value, i.e. the probability average; x is a data point sampled from the real data distribution; z is a data point sampled from the noise distribution, and the generator network G converts z into a synthetic sample G(z); D(x) and D(G(z)) represent the discriminator's judgment results on the real data x and the generated data G(z), respectively.

[0088] Step 5: Integrated System Design and Optimization. Through a multi-task learning strategy, the three core modules of signature detection, recognition, and anti-counterfeiting are unified, sharing underlying features, reducing redundant computation, and improving overall efficiency. Simultaneously, a complementary approach of fixed thresholds and adaptive adjustment is adopted, enabling dynamic parameter adjustments based on the characteristics of the input signature and environmental changes, ensuring optimal performance in any scenario. This integrated design not only optimizes the collaboration between modules but also improves real-time response capabilities and robustness, ultimately realizing an advanced and practical multi-scale signature detection and anti-counterfeiting integrated system.

[0089] In summary, this application employs a YOLO model combined with FPN for multi-scale signature detection, enabling the system to accurately detect signatures of different sizes and resolutions under complex backgrounds and diverse signature styles. Secondly, the adaptive enhancement and standardization process for signature images significantly improves the quality of the input image. By using CLAHE technology for contrast enhancement, the visibility of the signature is effectively improved in scenes with uneven lighting or low contrast; the introduction of DnCNN technology further removes image noise, enhancing image clarity and detail. Finally, ESRGAN is used for super-resolution reconstruction to generate high-quality standardized signature image input. Regarding signature recognition and feature matching, a ResNet-50 deep convolutional neural network model is used for feature extraction, combined with a cosine similarity-based feature matching method, resulting in excellent performance when processing signatures in different languages ​​and fonts. In terms of anti-counterfeiting, a GAN-based forged signature detection model generates high-quality forged signature samples and combines Gabor filters for multi-scale analysis of signature textures, effectively improving the detection capability for highly similar forged signatures. Finally, a multi-task learning strategy was used to integrate the signature detection, recognition, and anti-counterfeiting modules into a jointly optimized system, further improving the synergistic effect between the modules.

[0090] Through the above description of the embodiments, those skilled in the art can clearly understand that the signature recognition method according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platform. Of course, it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software device. This computer software device is stored in a storage medium (such as ROM / RAM, disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the signature recognition method of the various embodiments of this application.

[0091] This embodiment also provides a signature recognition device for implementing the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, hardware implementations, or a combination of software and hardware, are also possible and contemplated.

[0092] Figure 5 This is a structural block diagram of a signature recognition device according to an embodiment of this application; as shown below. Figure 5 As shown, it includes:

[0093] The first processing module 52 is used to extract the signature region in the file through the detection sub-model to obtain the first image;

[0094] The second processing module 54 is used to enhance the first image using a preprocessing sub-model to obtain a second image;

[0095] The determination module 56 is used to obtain the target signature features in the second image through the recognition sub-model, and determine the similarity between the target signature features and the preset signature features in the database, so as to determine the target recognition result of the signature region based on the similarity.

[0096] Using the aforementioned apparatus, a detection sub-model extracts and processes the signature region in a document to obtain a first image; a preprocessing sub-model enhances the first image to obtain a second image; a recognition sub-model acquires the target signature features in the second image and determines the similarity between the target signature features and preset signature features in a database, thereby determining the target recognition result of the signature region based on the similarity. This technical solution solves the problem of not being able to effectively improve the accuracy of signature detection and recognition when processing complex backgrounds and low-quality images. Furthermore, through the collaborative work of the three sub-models—detection, enhancement, and recognition—the signature region is accurately located in the document, image quality is optimized, and signature features are compared and analyzed, thereby efficiently determining the authenticity of the signature.

[0097] In an exemplary embodiment, the determining module is further configured to perform multi-level processing on the second image to generate multiple feature maps; extract features from the multiple feature maps to obtain multiple types of features, wherein the multiple types of features include at least one of the following: texture features of the second image, shape features of the second image, and handwriting features of the second image; and fuse the multiple types of features to obtain target signature features.

[0098] In an exemplary embodiment, the determining module is further configured to calculate the similarity between the target signature feature and the preset signature feature in the database according to a preset feature matching algorithm, and obtain a similarity score; and determine the target recognition result of the signature region according to the relationship between the similarity score and the preset similarity threshold.

[0099] In an exemplary embodiment, the determining module is further configured to determine that the target recognition result of the signature region is a genuine signature when the similarity score is greater than or equal to a preset similarity threshold; and to determine that the target recognition result of the signature region is not a genuine signature when the similarity score is less than the preset similarity threshold.

[0100] In an exemplary embodiment, the apparatus further includes: a training module, configured to determine the similarity between the target signature feature and a preset signature feature in a database, and after determining the target recognition result of the signature region based on the similarity, obtain a real signature sample corresponding to the file; process the real signature sample by generating a sub-model to obtain a forged signature sample; train a preset anti-counterfeiting sub-model based on the forged signature sample and the real signature sample to obtain a target anti-counterfeiting sub-model; and perform verification processing on the target recognition result based on the target anti-counterfeiting sub-model.

[0101] In an exemplary embodiment, the training module further includes: a collection unit, configured to collect the verification results generated by the target anti-counterfeiting sub-model within a preset period after verifying the target recognition results based on the target anti-counterfeiting sub-model; classify the verification results to obtain a set of genuine signature samples and a set of forged signature samples; input the set of genuine signature samples into the target anti-counterfeiting sub-model for iterative training, and record abnormal signature sample data during the iterative training process.

[0102] Embodiments of this application also provide a storage medium including a stored program, wherein the program executes any of the above signature recognition methods when it is run.

[0103] Optionally, in this embodiment, the storage medium may be configured to store program code for performing the following steps:

[0104] S1, the signature region in the file is extracted and processed by the detection sub-model to obtain the first image;

[0105] S2, use the preprocessing sub-model to enhance the first image to obtain the second image;

[0106] S3, the target signature features in the second image are obtained by recognizing the sub-model, and the similarity between the target signature features and the preset signature features in the database is determined, so as to determine the target recognition result of the signature region based on the similarity.

[0107] Embodiments of this application also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0108] Embodiments of this application also provide a computer program product, including a computer program, wherein the computer program, when executed by a processor, performs the steps in any of the above method embodiments.

[0109] Optionally, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0110] Optionally, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0111] S1, the signature region in the file is extracted and processed by the detection sub-model to obtain the first image;

[0112] S2, use the preprocessing sub-model to enhance the first image to obtain the second image;

[0113] S3, the target signature features in the second image are obtained by recognizing the sub-model, and the similarity between the target signature features and the preset signature features in the database is determined, so as to determine the target recognition result of the signature region based on the similarity.

[0114] Optionally, in this embodiment, the storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical discs.

[0115] Optionally, specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0116] The collection, storage, use, processing, transmission, provision, and disclosure of financial data or user data involved in the technical solution of this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0117] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0118] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0119] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A signature recognition method, characterized in that, include: The signature region in the file is extracted and processed by the detection sub-model to obtain the first image; The first image is enhanced using a preprocessing sub-model to obtain the second image; The target signature features in the second image are obtained by recognizing the sub-model, and the similarity between the target signature features and the preset signature features in the database is determined, so as to determine the target recognition result of the signature region based on the similarity.

2. The signature recognition method according to claim 1, characterized in that, The target signature features in the second image are obtained by recognizing the sub-model, including: The second image is processed in multiple stages to generate multiple feature maps; Feature extraction is performed on the multiple feature maps to obtain multiple types of features, wherein the multiple types of features include at least one of the following: texture features of the second image, shape features of the second image, and handwriting features of the second image; The target signature feature is obtained by fusing the multiple types of features.

3. The signature recognition method according to claim 1, characterized in that, Determining the similarity between the target signature feature and preset signature features in the database, and determining the target recognition result of the signature region based on the similarity, includes: The similarity score is obtained by calculating the similarity between the target signature feature and the preset signature features in the database according to the preset feature matching algorithm. The target recognition result of the signature region is determined based on the relationship between the similarity score and the preset similarity threshold.

4. The signature recognition method according to claim 3, characterized in that, The target recognition result of the signature region is determined based on the relationship between the similarity score and a preset similarity threshold, including: If the similarity score is greater than or equal to the preset similarity threshold, the target recognition result of the signature region is determined to be a genuine signature; If the similarity score is less than the preset similarity threshold, the target recognition result of the signature region is determined to be a non-genuine signature.

5. The signature recognition method according to claim 1, characterized in that, After determining the similarity between the target signature feature and preset signature features in the database, and determining the target recognition result of the signature region based on the similarity, the method further includes: Obtain the actual signature sample corresponding to the file; By generating a sub-model, the real signature sample is processed to obtain a forged signature sample; The preset anti-counterfeiting sub-model is trained based on the forged signature sample and the real signature sample to obtain the target anti-counterfeiting sub-model; The target identification result is verified based on the target anti-counterfeiting sub-model.

6. The signature recognition method according to claim 1, characterized in that, After verifying the target identification result based on the target anti-counterfeiting sub-model, the method further includes: Collect the verification results generated by the target anti-counterfeiting sub-model within a preset period; The test results are classified to obtain a set of genuine signature samples and a set of forged signature samples; The set of real signature samples is input into the target anti-counterfeiting sub-model for iterative training, and abnormal signature sample data during the iterative training process is recorded.

7. A signature recognition device, characterized in that, include: The first processing module is used to extract the signature region in the file through the detection sub-model to obtain the first image; The second processing module is used to enhance the first image using a preprocessing sub-model to obtain the second image; The determination module is used to obtain the target signature features in the second image through the recognition sub-model, and determine the similarity between the target signature features and the preset signature features in the database, so as to determine the target recognition result of the signature region based on the similarity.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, it implements the steps of the signature recognition method according to any one of claims 1 to 6.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the signature recognition method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the signature recognition method according to any one of claims 1 to 6.

Images