Signature identification method and device, equipment and storage medium

Abstract

The invention discloses a signature identification method and device, equipment and a storage medium. The method comprises the following steps: firstly, performing feature extraction on a signature image to be identified by using a deep residual network to obtain a multi-channel signature feature map, and converging the multi-channel signature feature map into a single-channel signature feature vector through a global average pooling layer; then performing feature excitation on the feature vector by using an excitation module to obtain a weight vector, and multiplying each dimension value in the weight vector by a corresponding signature feature map channel to obtain a weighted feature map; and carrying out classification processing on the weighted feature map through a signature classification module, and outputting an identification result for identifying the authenticity of the signature to be identified. Wherein the target signature recognition model is obtained by carrying out transfer learning and model fine adjustment on the basic signature recognition model based on a plurality of real signature images, and the real signature images are reference objects of the to-be-recognized signature images so as to provide standard signature feature samples. According to the method, the accuracy and the processing speed of signature authenticity identification are remarkably improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a signature recognition method, apparatus, device and storage medium. Background Art

[0002] As a unique biometric, handwritten signatures are stable and easily accessible, making them widely used in many industries, including finance. In these fields, handwritten signatures are often considered an effective means of authentication, not only proving the authenticity of documents but also providing an intuitive way to identify users.

[0003] However, despite the numerous advantages of handwritten signatures, authenticity verification methods primarily rely on manual and traditional physical methods, significantly limiting the accuracy and efficiency of signature recognition. Manual verification typically requires professionals to carefully observe and compare signatures. While this method can achieve high accuracy in some cases, it also has significant drawbacks. For example, due to individual differences and variations in writing habits, even experienced experts cannot guarantee accurate identification of genuine signatures every time.

[0004] Traditional physical methods, such as using chemical reagents or magnifying glasses to detect signature details, can assist manual recognition to a certain extent. However, due to their reliance on visual observation and subjective judgment, their results are also unstable and unreliable. Furthermore, with the continuous development of technology, the skills of forging signatures are also constantly improving, making the limitations of traditional methods more prominent.

[0005] Therefore, although handwritten signatures have an irreplaceable position in many scenarios, current identification methods still face the problems of low recognition efficiency and inaccurate recognition. Summary of the Invention

[0006] Based on the above problems, the present application provides a signature recognition method, device, equipment and storage medium, which can improve the accuracy and efficiency of handwritten signature authenticity identification, so as to cope with the challenges of low recognition rate, low efficiency and difficulty in dealing with complex forgery methods of traditional manual review and physical detection methods.

[0007] The embodiments of this application disclose the following technical solutions:

[0008] A signature recognition method, applied to a target signature recognition model, comprising:

[0009] Use the deep residual network to extract features from the signature image to be identified and obtain a multi-channel signature feature map;

[0010] Performing global average pooling on the signature feature map using a global average pooling layer to obtain a single-channel signature feature vector;

[0011] Using an excitation module to perform feature excitation on the signature feature vector to obtain a weight vector;

[0012] Multiplying the value of each dimension in the weight vector by each channel of the signature feature map to obtain a weighted feature map;

[0013] Using a signature classification module to classify the weighted feature map, obtaining a recognition result of the signature image to be recognized; the recognition result is used to indicate the authenticity of the signature in the signature image to be recognized;

[0014] Among them, the target signature recognition model is obtained by performing transfer learning and model fine-tuning on the basic signature recognition model based on multiple real signature images; the real signature image is the reference object of the signature image to be recognized, which is used to provide a standard signature feature sample.

[0015] In a possible implementation, the excitation module includes a dimensionality reduction layer and a dimensionality increase layer;

[0016] The utilizing an excitation module to perform feature excitation on the signature feature vector to obtain a weight vector includes:

[0017] Using the dimensionality reduction layer to perform a dimensionality reduction operation on the signature feature vector to obtain a reduced dimensionality feature vector;

[0018] The dimension-raising layer is used to perform a dimension-raising operation on the dimension-reduced feature vector to obtain the weight vector.

[0019] In one possible implementation, the basic signature recognition model is a pre-trained model obtained by preliminary training using a large-scale signature dataset; the basic signature recognition model includes a basic deep residual network, a basic pooling layer, a basic excitation module and a basic classification module; the basic signature recognition model obtains the target signature recognition model including the deep residual network, the global average pooling layer, the excitation module and the signature classification module through transfer learning and model fine-tuning.

[0020] In a possible implementation, the basic signature recognition model further includes a basic discarding layer; and the target signature recognition model further includes a discarding layer.

[0021] In a possible implementation, the method further includes:

[0022] Performing random feature masking on the weighted feature map using the discarding layer to obtain a masked weighted feature map;

[0023] The method of classifying the weighted feature map using a signature classification module to obtain a recognition result of the signature image to be recognized includes:

[0024] The masked weighted feature map is classified using the signature classification module to obtain the recognition result.

[0025] A signature recognition device, comprising:

[0026] A feature extraction unit is used to extract features from the signature image to be identified using a deep residual network to obtain a multi-channel signature feature map;

[0027] A feature pooling unit, configured to perform global average pooling on the signature feature map using a global average pooling layer to obtain a single-channel signature feature vector;

[0028] A feature excitation unit, configured to perform feature excitation on the signature feature vector using an excitation module to obtain a weight vector;

[0029] A weighting unit, configured to multiply the value of each dimension in the weight vector by each channel of the signature feature map to obtain a weighted feature map;

[0030] a classification unit, configured to classify the weighted feature map using a signature classification module to obtain a recognition result of the signature image to be recognized; the recognition result is used to indicate the authenticity of the signature in the signature image to be recognized;

[0031] Among them, the target signature recognition model is obtained by performing transfer learning and model fine-tuning on the basic signature recognition model based on multiple real signature images; the real signature image is the reference object of the signature image to be recognized, which is used to provide a standard signature feature sample.

[0032] In a possible implementation, the excitation module includes a dimensionality reduction layer and a dimensionality increase layer;

[0033] The characteristic excitation unit specifically includes:

[0034] a dimensionality reduction unit, configured to perform a dimensionality reduction operation on the signature feature vector using the dimensionality reduction layer to obtain a reduced-dimensionality feature vector;

[0035] A dimension increasing unit is used to use the dimension increasing layer to perform a dimension increasing operation on the dimension reduced feature vector to obtain the weight vector.

[0036] In one possible implementation, the basic signature recognition model is a pre-trained model obtained by preliminary training using a large-scale signature dataset; the basic signature recognition model includes a basic deep residual network, a basic pooling layer, a basic excitation module and a basic classification module; the basic signature recognition model obtains the target signature recognition model including the deep residual network, the global average pooling layer, the excitation module and the signature classification module through transfer learning and model fine-tuning.

[0037] A signature recognition device 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, the signature recognition method described above is implemented.

[0038] A computer-readable storage medium stores instructions, which, when executed on a terminal device, enable the terminal device to execute the signature recognition method described above.

[0039] Compared with the existing technology, this application has the following beneficial effects:

[0040] The present application provides a signature recognition method, apparatus, device, and storage medium. Specifically, when executing the signature recognition method provided in the embodiments of the present application, a deep residual network can first be used to perform multi-channel feature extraction on the signature image to be recognized to generate a rich signature feature map. Subsequently, the multi-channel feature map is aggregated into a single-channel signature feature vector through a global average pooling layer. Next, the feature vector is feature-excited using an excitation module to obtain an importance weight vector for each channel. Then, the value of each dimension in the weight vector is multiplied by the corresponding signature feature map channel to obtain a weighted feature map. Finally, the weighted feature map is classified and processed by a signature classification module to output a recognition result that indicates the authenticity of the signature to be recognized. Among them, the target signature recognition model is optimized from the basic signature recognition model based on a large number of real signature images through transfer learning and fine-tuning technology. The real signature images serve as reference samples, providing a standardized feature basis for the recognition process, thereby improving the accuracy and reliability of recognition. The present application automatically extracts and weights key features through a deep learning model, overcoming the errors and efficiency bottlenecks caused by human subjective judgment. At the same time, the use of transfer learning and fine-tuning technology enables the model to better adapt to specific real signature samples, significantly improving the recognition accuracy and robustness, thereby effectively solving the problems of inaccurate recognition and low efficiency of traditional methods, and meeting the needs of modern finance and other industries for large-scale, high-quality signature identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in this embodiment or the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0042] Figure 1 A flow chart of a signature recognition method provided in an embodiment of the present application;

[0043] Figure 2 A schematic diagram of an exemplary application scenario provided in an embodiment of the present application;

[0044] Figure 3 A schematic diagram of the structure of a signature recognition device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0045] To facilitate understanding of the technical solutions provided by the embodiments of the present application, the background technology involved in the embodiments of the present application will be described below.

[0046] As a unique biometric, handwritten signatures have long been widely used in finance and other important fields due to their stability and convenience in individual identity authentication. However, despite the significant value of handwritten signatures in practical applications, their authenticity verification primarily relies on manual review or traditional physical detection methods such as ink analysis and paper testing. These traditional methods, limited by expert experience, visual fatigue, and technical limitations, often result in low recognition accuracy and inefficiency. Furthermore, the manual verification process is time-consuming and costly, making it difficult to meet the needs of modern large-scale, high-frequency signature verification. With the continuous advancement of signature forgery technology, traditional methods are becoming increasingly inadequate when dealing with complex disguises, making them prone to misjudgment or omission, posing potential risks to related industries.

[0047] To address this problem, embodiments of the present application provide a signature recognition method, apparatus, device, and storage medium. First, a deep residual network is used to extract multi-channel features from the signature image to be recognized, generating a rich signature feature map. This multi-channel feature map is then converted into a single-channel signature feature vector using a global average pooling layer. An excitation module then performs feature excitation on this feature vector, calculating a weight vector to reflect the importance of each channel. Each dimension in the weight vector is then multiplied by the corresponding feature map channel to obtain a weighted feature map. Finally, a signature classification module classifies the weighted feature map and outputs a recognition result for determining the authenticity of the signature to be recognized. Furthermore, the target signature recognition model is based on a basic signature recognition model, combined with multiple real signature images, and obtained through transfer learning and model fine-tuning techniques. Real signatures are used as reference samples, thereby improving the model's recognition accuracy and adaptability. This application automatically extracts multi-channel signature feature maps using deep learning technology, and utilizes global average pooling, feature excitation, and weighted feature map processing to accurately determine the authenticity of the signature using a classification module. In addition, the model performs transfer learning and fine-tuning based on a large number of real signature images, which can effectively deal with forged signatures, thus overcoming the problems of low recognition efficiency and poor accuracy of traditional methods.

[0048] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0049] See also Figure 1 This figure is a flow chart of a signature recognition method provided in an embodiment of the present application, which is applied to a target signature recognition model. The target signature recognition model is based on the basic signature recognition model. By introducing multiple real signature images, using transfer learning technology to transfer knowledge, and combining model fine-tuning and optimization, the target signature recognition model can more accurately adapt to the characteristic distribution of the target signature. These real signature images serve as reference samples for the signature to be recognized, providing the model with standardized feature templates, thereby effectively improving the accuracy and reliability of recognition.

[0050] like Figure 1 As shown, the signature recognition method may include steps S101-S105:

[0051] S101: The target signature recognition model uses a deep residual network to extract features from the signature image to be recognized and obtain a multi-channel signature feature map.

[0052] The target signature recognition model implements a deep residual network (DRN) to achieve efficient feature extraction of signature images. Using multi-layer residual modules, DRNs effectively alleviate the vanishing gradient problem in deep networks, thereby improving the model's expressiveness and training effectiveness. Specifically, the network performs layer-by-layer convolution and nonlinear transformations on the input signature image, capturing multi-dimensional information such as the signature's stroke texture and structural morphology. The network ultimately outputs a signature feature map U consisting of multiple channels, where U = [u1, u2, ..., uc], with a width of W, a height of H, and a number of channels of C. These multi-channel feature maps not only richly represent the signature's detailed features but also provide a solid foundation for subsequent feature weighting and classification, helping to improve the accuracy and robustness of signature authentication.

[0053] S102: The target signature recognition model uses a global average pooling layer to perform global average pooling on the signature feature map to obtain a single-channel signature feature vector.

[0054] The target signature recognition model uses a global average pooling layer to process the multi-channel signature feature map U, compressing it from the spatial dimension (height H and width W) to generate a single-channel signature feature vector Z (such as Figure 2 ). Specifically, the input feature map U has dimensions H×W×C, where H and W represent the height and width of the feature map, respectively, and C represents the number of channels. Global average pooling is used to average the feature values ​​of all spatial locations within each channel, compressing the original H×W region into a scalar, thereby concentrating the information of each channel into a single value. Ultimately, this process "squeezes" the original multi-channel two-dimensional feature map into a 1×1×C feature vector Z. This vector effectively preserves the global information of each channel, providing a compact and rich description for subsequent feature excitation and weighting operations, helping to improve the recognition performance and stability of the model.

[0055] In one possible implementation, the global average pooling formula of the global average pooling layer is as follows:

[0056] = ;y=1,2,....,C; Represents the feature vector of the yth channel in the signature feature vector Z; Represents the eigenvalue of the y-th channel in the feature map U at the i-th row and j-th column.

[0057] S103: The target signature recognition model uses an excitation module to perform feature excitation on the signature feature vector to obtain a weight vector.

[0058] The target signature recognition model performs feature excitation on the signature feature vector through the excitation module to generate a weight vector S (such as Figure 2 (as shown). This excitation module automatically captures and evaluates the importance of each channel's features through a series of nonlinear transformations and an attention mechanism, assigning higher weights to task-relevant features and suppressing irrelevant or noisy information. Specifically, the excitation module inputs the signature feature vector, passes it through a fully connected layer and activation function, and outputs a weight vector S that reflects the relative importance of each channel. This weight vector is then used to guide the weighted processing of the feature map, enhancing the model's focus on key features, thereby effectively improving the accuracy and robustness of signature authentication.

[0059] In a possible implementation, the excitation module includes a dimensionality reduction layer and a dimensionality increase layer.

[0060] In a possible implementation, step S103 of the target signature recognition model performs feature excitation on the signature feature vector using the excitation module to obtain a weight vector, including the following operations:

[0061] First, the target signature recognition model uses a dimensionality reduction layer to reduce the dimensionality of the input signature feature vector, compressing high-dimensional features into a low-dimensional representation, thereby effectively extracting key information and reducing computational complexity, resulting in a reduced-dimensional feature vector. The model then uses a dimensionality increase layer to restore the reduced-dimensional feature vector, mapping it back to its original dimensional space. It also enhances the feature representation through nonlinear transformations, ultimately generating a weight vector that reflects the importance of each channel.

[0062] Specifically, the feature vector Z is first reduced to a lower dimensional space through a dimensionality reduction layer (such as a fully connected layer or a convolutional layer). The role of the dimensionality reduction layer is to extract more compact and important features. Assume that the feature vector after dimensionality reduction is Zd, and its dimension is 1×1×D1, where D <C。

[0063] Applying an activation function, such as the Rectified Linear Unit (ReLU) activation function, to the reduced-dimensional feature vector Zd performs a nonlinear transformation. This step can further enhance the key information in the low-dimensional features and suppress negative values.

[0064] Next, the reduced feature vector Zd is returned to its original dimension through a dimensionality-increasing layer (such as a fully connected layer or a deconvolution layer). The dimensionality-increasing layer restores the original dimension of the feature vector and generates a weight vector S with the same dimensions as the original feature vector Z, namely 1×1×C.

[0065] This process enables the excitation module to adaptively adjust the weights of each channel feature, increase the model's attention to key features, and enhance the accuracy and robustness of signature recognition.

[0066] In one possible implementation, the characteristic excitation formula of the excitation module is as follows:

[0067] S= ( ) = ( ); Represents the weighting function; W=[W1,W2]; W1 represents the weight parameter in the dimensionality reduction layer, which is used to compress the dimension of the input feature vector Z; W2 represents the weight parameter in the dimensionality increase layer, which is used to restore the feature vector after dimensionality reduction to its original dimension; σ represents the activation function, usually the Sigmoid function, which is used to map the output to a weight value between 0 and 1; δ represents the ReLU activation function, which is used to introduce nonlinearity and enhance the model's expressiveness.

[0068] S104: The target signature recognition model multiplies the value of each dimension in the weight vector with each channel of the signature feature map to obtain a weighted feature map.

[0069] The target signature recognition model multiplies the value of each dimension in the weight vector S with each channel of the corresponding signature feature map U channel by channel, and weights the importance of different channel features to obtain the weighted feature map X (such as Figure 2 Specifically, the feature map U consists of C channels, each channel is denoted as u y , each element in the weight vector S corresponds to s y , the model will channel u y and weight coefficient s y Multiply them together to get the weighted feature channels. By performing this operation on all channels, a weighted feature map X is formed, that is, X=F scale (U,S)=(u1·s1,u2·s2,...,u C ·s C ), F scale The name of a function that represents a "channel weighting" or "feature scaling" operation, F scale The function is to multiply each channel in the input multi-channel feature map by the corresponding weight coefficient to achieve channel-level feature weighting.

[0070] This process can dynamically adjust the contribution of each channel feature, highlight key features and suppress irrelevant information, thereby enhancing the model's sensitivity to signature image details and structure, and improving recognition accuracy and robustness.

[0071] S105: The target signature recognition model uses a signature classification module to classify the weighted feature map to obtain a recognition result of the signature image to be recognized.

[0072] The target signature recognition model uses the signature classification module to conduct in-depth analysis and processing of the weighted feature map to achieve the classification task of the signature image to be identified. Specifically, after the aforementioned weighting operation, the information of key channels in the feature map is effectively enhanced, allowing the classification module to more accurately capture the subtle differences and key features in the signature. In one possible implementation, the classification module can be composed of multiple fully connected layers, activation functions, and normalization layers. It is responsible for mapping the weighted features into a predefined category space and outputting a probability or judgment result representing the authenticity of the signature. Ultimately, this recognition result is used to determine the authenticity of the input signature image, that is, to confirm whether the signature is legitimate or forged, thereby providing a reliable basis for signature verification and enhancing the security and credibility of the system.

[0073] In one possible implementation, the basic signature recognition model is a pre-trained model obtained through preliminary training on a large-scale signature dataset. This basic model comprises core components such as a basic deep residual network, a basic pooling layer, a basic excitation module, and a basic classification module. Using transfer learning techniques, the model can transfer general features and knowledge learned from large-scale data to new tasks. Subsequently, the basic model is fine-tuned using target signature data to better adapt the model to the specific characteristics of the signature to be recognized. Through this process, the basic signature recognition model evolves into a target signature recognition model, whose structure includes a deep residual network, a global average pooling layer, an excitation module, and a signature classification module, significantly improving the recognition accuracy and robustness of target signature images.

[0074] It's important to note that the basic classification module is a key component of the basic signature recognition model, responsible for mapping extracted and processed signature features into specific classification results. The basic classification module can be composed of several fully connected layers, activation functions, and normalization layers, performing nonlinear transformations and pattern recognition on the input feature vector. The basic classification module's primary function is to determine the signature's category (e.g., authentic or forged) based on the extracted signature features, thereby achieving preliminary signature recognition. By training on large-scale signature datasets, the basic classification module learns effective classification boundaries and discrimination rules, providing a solid foundation for subsequent transfer learning and fine-tuning.

[0075] In one possible implementation, the basic signature recognition model also includes a basic dropout layer (Dropout Layer), which is used to randomly drop some neuronal connections during training to prevent overfitting and improve generalization. Based on this basic model, the target signature recognition model obtained through transfer learning and fine-tuning also includes a dropout layer to maintain training stability and improve recognition performance, thereby enhancing the model's adaptability and robustness to different signature samples.

[0076] In a possible implementation, the method further includes:

[0077] A dropout layer performs random feature masking on the weighted feature map. During the recognition process, some neuron activation values ​​in the weighted feature map are randomly "masked" or "discarded" with a certain probability, resulting in a masked weighted feature map. This operation effectively prevents the model from becoming overly dependent on specific channels or features, enhancing the model's generalization and robustness, and thus improving the recognition accuracy and stability of the target signature recognition model in practical applications.

[0078] In a possible implementation, the method of using the signature classification module to classify the weighted feature map to obtain a recognition result of the signature image to be recognized specifically includes:

[0079] First, the weighted feature map, after random masking by the dropout layer, is fed into the signature classification module. The classification module then performs feature analysis and discrimination based on this masked feature map, ultimately outputting the signature recognition result. This approach enables the model to perform classification based on a more robust feature representation, effectively improving the accuracy and stability of signature authenticity judgments.

[0080] Based on the contents of S101-S105, it can be seen that a deep residual network is first used to extract multi-channel features from the signature image to be identified, obtaining a rich signature feature map. Subsequently, the feature map is compressed into a single-channel feature vector through a global average pooling layer. The feature vector is then weighted and excited using an excitation module to generate a weight vector to highlight the key feature channels. The values ​​of each dimension in the weight vector are then multiplied by the corresponding feature map channel to obtain a weighted feature map. Finally, the weighted feature map is classified by the signature classification module, and a result is output for determining the authenticity of the signature to be identified. Among them, the target signature recognition model is based on the basic signature recognition model and uses transfer learning and model fine-tuning on multiple real signature images to better adapt the model to the characteristics of the signature image to be identified and improve identification accuracy. This application automatically extracts and weights key features through a deep learning model, effectively overcoming the errors and efficiency bottlenecks caused by human subjective judgment in traditional methods. At the same time, the use of transfer learning and fine-tuning techniques enables the model to better adapt to specific real signature samples, thereby significantly improving recognition accuracy and robustness. This not only solves the problems of inaccurate recognition and low efficiency of traditional methods, but also meets the needs of modern finance and other industries for large-scale, high-quality signature authentication.

[0081] See also Figure 3 , Figure 3 This is a schematic diagram of the structure of a signature recognition device provided in an embodiment of the present application. Figure 3 As shown, the signature recognition device includes:

[0082] A feature extraction unit 301 is used to extract features from the signature image to be identified using a deep residual network to obtain a multi-channel signature feature map;

[0083] A feature pooling unit 302 is configured to perform global average pooling on the signature feature map using a global average pooling layer to obtain a single-channel signature feature vector;

[0084] A feature excitation unit 303 is configured to perform feature excitation on the signature feature vector using an excitation module to obtain a weight vector;

[0085] A weighting unit 304 is configured to multiply the value of each dimension in the weight vector by each channel of the signature feature map to obtain a weighted feature map;

[0086] The classification unit 305 is configured to classify the weighted feature map using a signature classification module to obtain a recognition result of the signature image to be recognized; the recognition result is used to indicate the authenticity of the signature in the signature image to be recognized;

[0087] Among them, the target signature recognition model is obtained by performing transfer learning and model fine-tuning on the basic signature recognition model based on multiple real signature images; the real signature image is the reference object of the signature image to be recognized, which is used to provide a standard signature feature sample.

[0088] In a possible implementation, the excitation module includes a dimensionality reduction layer and a dimensionality increase layer.

[0089] In a possible implementation, the feature excitation unit specifically includes:

[0090] a dimensionality reduction unit, configured to perform a dimensionality reduction operation on the signature feature vector using the dimensionality reduction layer to obtain a reduced-dimensionality feature vector;

[0091] A dimension increasing unit is used to use the dimension increasing layer to perform a dimension increasing operation on the dimension reduced feature vector to obtain the weight vector.

[0092] In one possible implementation, the basic signature recognition model is a pre-trained model obtained by preliminary training using a large-scale signature dataset; the basic signature recognition model includes a basic deep residual network, a basic pooling layer, a basic excitation module and a basic classification module; the basic signature recognition model obtains the target signature recognition model including the deep residual network, the global average pooling layer, the excitation module and the signature classification module through transfer learning and model fine-tuning.

[0093] In a possible implementation, the basic signature recognition model further includes a basic discarding layer; and the target signature recognition model further includes a discarding layer.

[0094] In a possible implementation, the apparatus further includes:

[0095] A feature masking unit is used to perform random feature masking on the weighted feature map using the discarding layer to obtain a masked weighted feature map.

[0096] In a possible implementation, the classification unit 305 is specifically configured to:

[0097] The masked weighted feature map is classified using the signature classification module to obtain the recognition result.

[0098] In addition, an embodiment of the present application also provides a signature recognition device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the signature recognition method described above is implemented.

[0099] In addition, an embodiment of the present application further provides a computer-readable storage medium, in which instructions are stored. When the instructions are executed on a terminal device, the terminal device executes the signature recognition method described above.

[0100] This application utilizes deep learning technology to automatically extract key features from signature images and enhances important information through a weighted mechanism, avoiding the errors and efficiency limitations inherent in manual recognition due to subjective judgment. With the help of transfer learning and model fine-tuning, the recognition model can be optimized for real signature samples, improving its adaptability to diverse signature features and recognition stability. This application significantly improves the accuracy and processing speed of signature authenticity verification, effectively overcoming the shortcomings of traditional verification methods in terms of precision and efficiency, and meeting the practical needs of industries such as finance for large-scale, high-quality signature authentication.

[0101] The above is a detailed introduction to the signature recognition method, device, equipment and storage medium provided by the present application. The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the various embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of this application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the claims of this application.

[0102] It should also be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

Claims

1. A signature recognition method, characterized in that: Applied to a target signature recognition model, the method includes: Use the deep residual network to extract features from the signature image to be identified and obtain a multi-channel signature feature map; Performing global average pooling on the signature feature map using a global average pooling layer to obtain a single-channel signature feature vector; Using an excitation module to perform feature excitation on the signature feature vector to obtain a weight vector; Multiplying the value of each dimension in the weight vector by each channel of the signature feature map to obtain a weighted feature map; Using a signature classification module to classify the weighted feature map, obtaining a recognition result of the signature image to be recognized; the recognition result is used to indicate the authenticity of the signature in the signature image to be recognized; Among them, the target signature recognition model is obtained by performing transfer learning and model fine-tuning on the basic signature recognition model based on multiple real signature images; the real signature image is the reference object of the signature image to be recognized, which is used to provide a standard signature feature sample.

2. The method according to claim 1, characterized in that The excitation module includes a dimensionality reduction layer and a dimensionality increase layer; The utilizing an excitation module to perform feature excitation on the signature feature vector to obtain a weight vector includes: Using the dimensionality reduction layer to perform a dimensionality reduction operation on the signature feature vector to obtain a reduced dimensionality feature vector; The dimension-raising layer is used to perform a dimension-raising operation on the dimension-reduced feature vector to obtain the weight vector.

3. The method according to claim 1, characterized in that The basic signature recognition model is a pre-trained model obtained by preliminary training using a large-scale signature dataset; the basic signature recognition model includes a basic deep residual network, a basic pooling layer, a basic excitation module and a basic classification module; The basic signature recognition model obtains the target signature recognition model including the deep residual network, the global average pooling layer, the excitation module and the signature classification module through transfer learning and model fine-tuning.

4. The method according to claim 3, characterized in that The basic signature recognition model further includes a basic discarding layer; the target signature recognition model further includes a discarding layer.

5. The method according to claim 4, characterized in that The method further comprises: Performing random feature masking on the weighted feature map using the discarding layer to obtain a masked weighted feature map; The method of classifying the weighted feature map using a signature classification module to obtain a recognition result of the signature image to be recognized includes: The masked weighted feature map is classified using the signature classification module to obtain the recognition result.

6. A signature recognition device, characterized in that: The device comprises: A feature extraction unit is used to extract features from the signature image to be identified using a deep residual network to obtain a multi-channel signature feature map; A feature pooling unit, configured to perform global average pooling on the signature feature map using a global average pooling layer to obtain a single-channel signature feature vector; A feature excitation unit, configured to perform feature excitation on the signature feature vector using an excitation module to obtain a weight vector; A weighting unit, configured to multiply the value of each dimension in the weight vector by each channel of the signature feature map to obtain a weighted feature map; a classification unit, configured to classify the weighted feature map using a signature classification module to obtain a recognition result of the signature image to be recognized; the recognition result is used to indicate the authenticity of the signature in the signature image to be recognized; Among them, the target signature recognition model is obtained by performing transfer learning and model fine-tuning on the basic signature recognition model based on multiple real signature images; the real signature image is the reference object of the signature image to be recognized, which is used to provide a standard signature feature sample.

7. The device according to claim 6, characterized in that The excitation module includes a dimensionality reduction layer and a dimensionality increase layer; The characteristic excitation unit specifically includes: a dimensionality reduction unit, configured to perform a dimensionality reduction operation on the signature feature vector using the dimensionality reduction layer to obtain a reduced-dimensionality feature vector; A dimension increasing unit is used to use the dimension increasing layer to perform a dimension increasing operation on the dimension reduced feature vector to obtain the weight vector.

8. The device according to claim 6, characterized in that The basic signature recognition model is a pre-trained model obtained by preliminary training using a large-scale signature dataset; the basic signature recognition model includes a basic deep residual network, a basic pooling layer, a basic excitation module and a basic classification module; The basic signature recognition model obtains the target signature recognition model including the deep residual network, the global average pooling layer, the excitation module and the signature classification module through transfer learning and model fine-tuning.

9. A signature recognition device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the signature recognition method according to any one of claims 1 to 5 is implemented.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores instructions, and when the instructions are executed on a terminal device, the terminal device executes the signature recognition method according to any one of claims 1 to 5.

Images