Signature identification method based on book track characteristics

Abstract

The invention provides a signature authentic identification method based on book track characteristics, and aims to improve the accuracy and efficiency of signature authentic identification. According to the method, firstly, historical signature pictures of signature owners are collected to serve as a positive sample set, and current signature pictures are shot to serve as pictures to be verified; through an image text segmentation network, text pixel point segmentation is carried out on the pictures, so that complex background interference is resisted. And carrying out multi-direction feature extraction on the mask image by adopting a track fusion perception algorithm to obtain line book track features. By comprehensively considering the feature distance between the positive sample feature set and the to-be-verified feature and using the self-attention trajectory anomaly perception network for deep analysis, the method not only ensures the high similarity between the current signature and the historical signature of the affiliation, but also effectively prevents the imitation and counterfeiting of the signature of the affiliation, and improves the verification accuracy of the affiliation. And double guarantee of accuracy and safety is realized.

Description

Technical Field:

[0001] The present invention relates to the field of signature authentication, and specifically to a signature authentication method based on the running script trajectory features. Background Art:

[0002] In today's society where digitalization and networking are developing day by day, the application of electronic signatures is becoming more and more widespread, especially in the fields of e-commerce, legal document processing, distance education, etc. However, with the increasing use of electronic signatures, the verification of their security and authenticity has become particularly important. Traditional signature verification methods usually rely on manual authentication, which is not only time-consuming and laborious, but also the accuracy is easily affected by subjective judgment.

[0003] In recent years, in order to improve the efficiency and accuracy of signature authentication, many automatic signature verification technologies based on image processing and machine learning have emerged. These technologies usually involve feature extraction, comparison, and analysis of signature images. For example, invention CN202110144380.6 uses a pre-set genuine and fake signature decision model to identify the authenticity of signature images. This method relies on a pre-trained model, and the accuracy may decrease when dealing with unseen signature styles or signatures in complex backgrounds. The generalization ability of the model and its adaptability to environmental changes are the main challenges. Invention CN201610075319.X improves the accuracy of electronic signature authentication by increasing the amount of information collection, such as combining the writing information of signature information and verification information. However, this method may increase the complexity of user operations and affect the user experience. At the same time, the increased amount of information collection may arouse users' concerns about privacy.

[0004] The existing technologies usually focus on improving the accuracy of image processing algorithms or increasing information collection to assist the verification process. However, these methods still face challenges in dealing with highly complex signatures, countering forgery techniques, and protecting user privacy. Especially in terms of the generalization ability of the genuine and fake signature decision model and the user experience, the existing methods have not provided a comprehensive and balanced solution. Summary of the Invention:

[0005] Aiming at the deficiencies of the existing technologies, the present invention proposes a signature authentication method based on the running script trajectory features, which specifically includes the following steps to complete the authentication of the authenticity of signature pictures: S1: Collect the previous signature pictures of the signature owner as the positive sample set; take the current signature picture as the picture to be verified; S2: Use an image text segmentation network to perform text pixel point segmentation operations on each picture in the positive sample set to form a collection of positive sample mask pictures; S3: Use the same image text segmentation network in S2 to perform text pixel point segmentation operations on the picture to be verified to form a mask picture to be verified; S4: Use a trajectory fusion perception algorithm to extract features from each mask picture in the collection of positive sample mask pictures to form a collection of positive sample features;

[0006] S5: Use the same trajectory fusion perception algorithm in S4 to extract features from the mask graph to be verified, and form the feature to be verified; S6: Calculate the feature distances between each positive sample feature in the positive sample feature set and the feature to be verified, form a distance list, and determine whether the minimum value of the distance list is greater than the threshold. If it is less, execute S7. If it is greater, output "forged signature"; S7: Use a self-attention trajectory anomaly perception network to detect whether the feature to be verified is abnormal. If it is abnormal, output "forged signature". If it is not, output "correct signature".

[0007] As an alternative solution of the present invention, the positive sample set in step S1 is composed of 4 signature pictures of the owner in the past, and the pictures have no noise interference; the verification picture in step S1 is the signature picture in the current file to be verified, and the picture includes noise interference of background text, foreground seal and watermark.

[0008] As an alternative solution of the present invention, an image text segmentation network described in steps S2 and S3 specifically includes the following structure: the input of the network is a signature picture, and the output is a mask graph. From the input end to the output end, it successively includes: sampling layer 1, downsampling layer 2, downsampling layer 3, convolutional layer, upsampling layer 1, upsampling layer 2, upsampling layer 3; among them, downsampling layer 1 and upsampling layer 3, downsampling layer 2 and upsampling layer 2, downsampling layer 3 and upsampling layer 1 are merged through a resizing operation respectively; among them, the downsampling layer is composed of a pooling layer, a convolutional layer and a Relu activation; among them, the upsampling layer is composed of a transposed convolution, a convolutional layer and a Relu activation.

[0009] As an alternative solution of the present invention, in the trajectory fusion perception algorithm described in steps S4 and S5, a multi-directional perception convolution is specifically used. The multi-directions include: upper left, upper right, lower left, lower right, up, down, left, and right. It is characterized in that: the upper left direction perception convolution is specifically a 3*3 convolution kernel with weights increasing from the lower left corner to the upper left corner, and is used to capture the trajectory of the handwriting from the lower left to the upper left; the upper right direction perception convolution is specifically a 3*3 convolution kernel with weights increasing from the lower right corner to the upper right corner, and is used to capture the trajectory of the handwriting from the lower right to the upper right; the lower left direction perception convolution is specifically a 3*3 convolution kernel with weights increasing from the upper left corner to the lower left corner, and is used to capture the trajectory of the handwriting from the upper left to the lower left; the lower right direction perception convolution is specifically a 3*3 convolution kernel with weights increasing from the upper right corner to the lower right corner, and is used to capture the trajectory of the handwriting from the upper right to the lower right; the up direction perception convolution is specifically a 3*3 convolution kernel with weights increasing from top to bottom, and is used to capture the trajectory of the handwriting from top to bottom; the down direction perception convolution is specifically a 3*3 convolution kernel with weights increasing from bottom to top, and is used to capture the trajectory of the handwriting from bottom to top; the left direction perception convolution is specifically a 3*3 convolution kernel with weights increasing from left to right, and is used to capture the trajectory of the handwriting from left to right; the right direction perception convolution is specifically a 3*3 convolution kernel with weights increasing from right to left, and is used to capture the trajectory of the handwriting from right to left.

[0010] As an alternative solution of the present invention, the trajectory fusion perception algorithm specifically extracts features from the mask graph through the following steps:

[0011] T1: Input the mask graph, and through the convolution operations of the upper left direction perception convolution, upper right direction perception convolution, lower left direction perception convolution, lower right direction perception convolution, up direction perception convolution, down direction perception convolution, left direction perception convolution, and right convolution direction perception convolution, obtain the trajectory features in 8 directions, and each direction is consistent with the direction to which the perception convolution belongs;

[0012] T2: Merge the trajectory features in 8 directions in the last dimension to obtain the merged trajectory features;

[0013] T3: Expand the merged trajectory features in one dimension to obtain the positive sample features or the features to be verified.

[0014] As an alternative solution of the present invention, the specific calculation process of the feature distance in S6 is as follows:

[0015] S6-1: Input the positive sample feature A and the feature B to be verified, where A = (a1, a2,..., a n ), B = (b1, b2,..., b n ), where a and b respectively represent the elements of the positive sample feature A and the feature B to be verified, and n represents the feature length;

[0016] S6-2: Calculate the dot product of the corresponding elements of the positive sample feature A and the feature B to be verified. The calculation formula is as follows:

[0017]

[0018] S6-3: Calculate the norms of the corresponding elements of the positive sample feature A and the feature B to be verified. The calculation formula is as follows:

[0019]

[0020]

[0021] S6-4: Calculate the feature distance through the dot product and the norm. The calculation formula is as follows:

[0022]

[0023] Among them, the feature distance value will be between 0 and 2. The closer the value is to 0, the more similar the features are.

[0024] As an alternative solution of the present invention, the self-attention trajectory anomaly perception network in S7 specifically includes the following structure: The input of this network is the feature to be verified, and the output is 0 and 1, where 0 represents "not abnormal" and 1 represents "abnormal"; from the input to the output of this network, it successively includes a rightward LSTM, a leftward LSTM, and a self-attention feature fusion layer. The output of the rightward LSTM contains K hidden states, and the output of the leftward LSTM updates these K hidden states; the rightward LSTM contains K time steps internally, and the leftward LSTM also contains K time steps internally. The structures of the rightward LSTM and the leftward LSTM are completely opposite.

[0025] Compared with the related prior art, compared with the prior art, the present application proposal has the following main technical advantages: The beneficial effects of the present invention are:

[0026] Improve the discrimination accuracy: The present invention adopts advanced image processing technology and deep learning models, which can more accurately extract and analyze the stroke trajectory features in the signature image, thereby greatly improving the discrimination accuracy of signature authenticity.

[0027] Powerful generalization ability: Compared with the prior art, the algorithm and model design of the present invention consider various different signature styles and conditions. Four previous signature pictures of the owner are used as a control set. As long as any one is matched, it is a valid signature, with stronger generalization ability.

[0028] Ability to handle complex backgrounds: The present invention extracts the pixel-level region of the signature through a segmentation algorithm, and can effectively analyze and process signature images with complex backgrounds, such as background text, seals, and watermarks, which is a significant challenge in traditional methods.

[0029] User privacy protection: Considering the sensitivity of personal signatures, the processing method of the present invention analyzes all picture content, and all operations use background programs to ensure that personal data will not be leaked or misused during signature authentication. Brief description of the drawings:

[0030] Figure 1 is the flowchart of the method provided by the present invention;

[0031] Figure 2 is the structural diagram of an image text segmentation network provided by the present invention;

[0032] Figure 3 is the structural diagram of a self-attention trajectory anomaly perception network provided by the present invention;

[0033] Figure 4 is an example of the positive sample set of the embodiment provided by the present invention;

[0034] Figure 5 is an example of the pattern to be verified in the embodiment provided by the present invention;

[0035] Figure 6 is an example of the positive sample mask pattern of the embodiment provided by the present invention;

[0036] Figure 7 is an example of the mask pattern to be verified in the embodiment provided by the present invention;

[0037] Figure 8 is an example of the multi-directional perception convolution of the embodiment provided by the present invention. Detailed implementation manners:

[0038] The present invention will be further described below in conjunction with the drawings and embodiments. However, the present invention can be implemented in many different ways and should not be construed as limited to the embodiments shown; on the contrary, these embodiments provide implementation manners that meet the applicable legal requirements for those skilled in the art.

[0039] Embodiment 1: This embodiment details a signature forgery detection method based on the running script trajectory features. This method uses advanced image processing techniques and deep learning models to authenticate the authenticity of signature pictures. The entire process involves the collection of signature samples, feature extraction, feature comparison, and anomaly detection. The application scenario of this embodiment is to authenticate a forged signature that is very similar.

[0040] Such as Figure 1As shown in the figure, a signature authentication method based on the trajectory features of running script proposed by the present invention specifically includes the following steps to authenticate the authenticity of the signature picture:

[0041] S1: As Figure 4 shown, collect the previous signature pictures of the signature owner as the positive sample set; as Figure 5 shown, take a picture of the current signature as the picture to be verified; as Figure 4 shown, the positive sample set consists of 4 previous signature pictures of the owner, and the pictures have no noise interference; as Figure 5 shown, the verification picture is the signature picture in the current file to be verified, and the picture includes noise interference such as background text, foreground seal and watermark.

[0042] S2: Use an image text segmentation network to perform text pixel point segmentation operations on each picture in the Figure 4 shown positive sample set to form a positive sample mask map collection as Figure 6 shown;

[0043] As Figure 2 shown, the specific structure of the image text segmentation network is as follows. The input of the network is the signature picture, and the output is the mask map. From the input end to the output end, it includes: sampling layer 1, downsampling layer 2, downsampling layer 3, convolutional layer, upsampling layer 1, upsampling layer 2, upsampling layer 3; among them, downsampling layer 1 and upsampling layer 3, downsampling layer 2 and upsampling layer 2, downsampling layer 3 and upsampling layer 1 are merged through a resize operation respectively; among them, the downsampling layer consists of a pooling layer, a convolutional layer and a Relu activation; among them, the upsampling layer consists of a transposed convolution, a convolutional layer and a Relu activation.

[0044] S3: As Figure 2 shown, use an image text segmentation network to perform text pixel point segmentation operations on the Figure 5 shown picture to be verified to form a mask map to be verified as Figure 7 shown;

[0045] S4: Use a trajectory fusion perception algorithm to extract features from each mask map in the positive sample mask map collection to form a positive sample feature collection;

[0046] This trajectory fusion perception algorithm specifically uses a multi-directional perception convolution. The multi-directions include: upper left, upper right, lower left, lower right, up, down, left, right. Its characteristics are:

[0047] As Figure 8 A shown, the upper left direction perception convolution is specifically a 3*3 convolution kernel with weights increasing from the lower left corner to the upper left corner, which is used to capture the trajectory of the handwriting from the lower left to the upper left;

[0048] As Figure 8As shown in Figure B, the upper - right - direction perception convolution is specifically a 3×3 convolution kernel with weights increasing from the lower - right corner to the upper - right corner, which is used to capture the trajectory of the handwriting from the lower - right to the upper - right;

[0049] As Figure 8 shown in Figure C, the lower - left - direction perception convolution is specifically a 3×3 convolution kernel with weights increasing from the upper - left corner to the lower - left corner, which is used to capture the trajectory of the handwriting from the upper - left to the lower - left;

[0050] As Figure 8 shown in Figure D, the lower - right - direction perception convolution is specifically a 3×3 convolution kernel with weights increasing from the upper - right corner to the lower - right corner, which is used to capture the trajectory of the handwriting from the upper - right to the lower - right;

[0051] As Figure 8 shown in Figure E, the upward - direction perception convolution is specifically a 3×3 convolution kernel with weights increasing from top to bottom, which is used to capture the trajectory of the handwriting from top to bottom;

[0052] As Figure 8 shown in Figure F, the downward - direction perception convolution is specifically a 3×3 convolution kernel with weights increasing from bottom to top, which is used to capture the trajectory of the handwriting from bottom to top;

[0053] As Figure 8 shown in Figure G, the left - direction perception convolution is specifically a 3×3 convolution kernel with weights increasing from left to right, which is used to capture the trajectory of the handwriting from left to right;

[0054] As Figure 8 shown in Figure H, the right - direction perception convolution is specifically a 3×3 convolution kernel with weights increasing from right to left, which is used to capture the trajectory of the handwriting from right to left.

[0055] The specific steps for feature extraction are as follows:

[0056] T1: Input the mask image. After convolution operations of the upper - left - direction perception convolution, upper - right - direction perception convolution, lower - left - direction perception convolution, lower - right - direction perception convolution, upward - direction perception convolution, downward - direction perception convolution, left - direction perception convolution, and right - direction perception convolution, obtain the trajectory features in 8 directions, and each direction is consistent with the direction to which the perception convolution belongs;

[0057] T2: Merge the trajectory features in 8 directions in the last dimension to obtain the merged trajectory features;

[0058] T3: Expand the merged trajectory features in one dimension to obtain the positive sample features or features to be verified.

[0059] S6: Calculate the feature distances between each positive sample feature in the positive sample feature set and the feature to be verified, forming a distance list, and determine whether the minimum value of the distance list is greater than the threshold. If it is less, execute S7; if it is greater, output "Forged signature". The specific steps for calculating the threshold are as follows:

[0060] S6-1: Input positive sample feature A and the feature to be verified B, where A = (a1, a2,..., a n ), B = (b1, b2,..., b n ), where a and b respectively represent the elements of positive sample feature A and the feature to be verified B, and n represents the feature length;

[0061] S6-2: Calculate the dot product of the corresponding elements of positive sample feature A and the feature to be verified B. The calculation formula is:

[0062]

[0063] S6-3: Calculate the norm lengths of the corresponding elements of positive sample feature A and the feature to be verified B. The calculation formula is:

[0064]

[0065]

[0066] S6-4: Calculate the feature distance through the dot product and the norm length. The calculation formula is:

[0067]

[0068] Among them, the feature distance value will be between 0 and 2. The closer the value is to 0, the more similar the features are.

[0069] In the embodiment, if the feature distance between the positive sample feature set Figure 4 C and the feature of the picture to be verified is less than 0.2, then execute S7.

[0070] S7: As Figure 3 shown, use a self-attention trajectory anomaly perception network to detect whether the feature to be verified is abnormal. Determine whether it is abnormal. If it is, output "Forged signature"; if not, output "Correct signature".

[0071] The structure of the self-attention trajectory anomaly perception network is as follows: The input of this network is the feature to be verified, and the output is 0 and 1, where 0 represents "not an anomaly" and 1 represents "is an anomaly". From the input to the output of this network, it successively includes a rightward LSTM, a leftward LSTM, and a self-attention feature fusion layer. The output of the rightward LSTM contains K hidden states, and the output of the leftward LSTM updates these K hidden states. The rightward LSTM contains K time steps internally, and the leftward LSTM also contains K time steps internally. The structures of the rightward LSTM and the leftward LSTM are exactly opposite.

[0072] In this embodiment, the running script trajectory of the "A" in "ZHANGSAN" is abnormal, and the output of the self-attention trajectory anomaly perception network is "yes".

[0073] Although there is a certain similarity between the image to be verified and the positive sample feature image judged by similarity, through the perception of the running script trajectory feature, the present invention accurately identifies this sample as a forged signature, preventing further losses caused by contract fraud.

[0074] This embodiment details the entire process from image acquisition to the final judgment of signature authenticity. The key technologies involved include image text segmentation, the use of the trajectory fusion perception algorithm, the calculation method of feature distance, and the application of the self-attention trajectory anomaly perception network. This embodiment aims to provide an effective signature forgery detection method applicable to various scenarios with signature verification requirements.

[0075] The above embodiments only represent several implementation manners of the present invention. Their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A signature authentication method based on the trajectory features of running script, a signature forgery detection method based on the trajectory of running script, characterized in that, The method comprises the following steps S1: Collect the signature owner's previous signature pictures as the positive sample set; take the current signature picture as the picture to be verified; S2: Use an image text segmentation network to perform text pixel segmentation operations on each image in the positive sample set to form a collection of positive sample mask images; S3: Use the same image text segmentation network in S2 to perform text pixel segmentation on the image to be verified to form a mask image to be verified; S4: Use a trajectory fusion perception algorithm to extract features from each mask image in the positive sample mask image collection to form a positive sample feature collection; S5: Use the same trajectory fusion perception algorithm in S4 to extract features from the mask image to be verified to form features to be verified; S6: Calculate the feature distance between each positive sample feature and the feature to be verified in the positive sample feature set, form a distance list, and determine whether the minimum value of the distance list is greater than the threshold. If it is less than, execute S7; if it is greater, output "forged signature"; S7: Use a self-attention trajectory anomaly perception network to detect whether the feature to be verified is abnormal and determine whether it is abnormal. If so, output "forged signature"; if not, output "correct signature".

2. The signature forgery detection method based on the running script trajectory according to claim 1, wherein, The positive sample set in step S1 is composed of 4 previous signature pictures of the owner, and the pictures are free of any noise interference; the verification picture in step S1 is the signature picture in the current document to be verified, and the picture includes background text, foreground seal and watermark noise interference.

3. A signature forgery detection method based on running script trajectories according to claim 1, characterized in that, The image text segmentation network described in steps S2 and S3 specifically comprises the following structure: the input of the network is a signature image, the output is a mask image, and the network includes from the input end to the output end: sampling layer 1, downsampling layer 2, downsampling layer 3, convolution layer, upsampling layer 1, upsampling layer 2, upsampling layer 3; wherein the downsampling layer 1 and upsampling layer 3, the downsampling layer 2 and upsampling layer 2, the downsampling layer 3 and upsampling layer 1 are respectively merged through a resize operation; wherein the downsampling layer is composed of a pooling layer, a convolution layer and a Relu activation; wherein the upsampling layer is composed of a deconvolution, a convolution layer and a Relu activation.

4. A signature forgery detection method based on running script trajectory according to claim 1, characterized in that The trajectory fusion perception algorithm described in steps S4 and S5 specifically uses a multi-directional perception convolution, and the multi-directions include: upper left, upper right, lower left, lower right, upper, lower, left, and right, and is characterized by: The upper left direction aware convolution is a 3*3 convolution kernel with increasing weights from the lower left corner to the upper left corner, which is used to capture the trajectory of the handwriting from the lower left corner to the upper left corner. The upper right direction aware convolution is a 3*3 convolution kernel with increasing weights from the lower right corner to the upper right corner, which is used to capture the trajectory of the handwriting from the lower right corner to the upper right corner. The lower left direction aware convolution is a 3*3 convolution kernel with increasing weights from the upper left corner to the lower left corner, which is used to capture the trajectory of the handwriting from the upper left corner to the lower left corner. The lower right direction aware convolution is a 3*3 convolution kernel with increasing weights from the upper right corner to the lower right corner, which is used to capture the trajectory of the handwriting from the upper right corner to the lower right corner. The upward-aware convolution is a 3*3 convolution kernel with increasing weights from top to bottom, which is used to capture the trajectory of handwriting from top to bottom. Downward direction perception convolution, specifically a 3*3 convolution kernel with weights increasing from bottom to top, used to capture the trajectory of handwriting from bottom to top; Left direction perception convolution, specifically a 3*3 convolution kernel with weights increasing from left to right, used to capture the trajectory of handwriting from left to right; Right direction perception convolution, specifically a 3*3 convolution kernel with weights increasing from right to left, used to capture the trajectory of handwriting from right to left.

5. The signature forgery detection method based on running script trajectory according to claim 4, wherein, The described trajectory fusion perception algorithm specifically extracts features from the mask image through the following steps: T1: Input the mask image, and perform convolution operations of the upper left direction perception convolution, upper right direction perception convolution, lower left direction perception convolution, lower right direction perception convolution, upper direction perception convolution, lower direction perception convolution, left direction perception convolution, and right convolution direction perception convolution to obtain trajectory features in 8 directions, and each direction is consistent with the direction to which the perception convolution belongs; T2: Merge the trajectory features in 8 directions in the last dimension to obtain the merged trajectory features; T3: Expand the merged trajectory features in one dimension to obtain positive sample features or features to be verified.

6. The signature forgery detection method based on running script trajectory according to claim 1, wherein, The specific calculation process of the feature distance in S6 is as follows: S6-1: Input the positive sample feature A and the feature B to be verified, where A = (a1, a2,..., a n ), B = (b1, b2,..., b n ), where a and b represent the elements of the positive sample feature A and the feature B to be verified respectively, and n represents the feature length; S6-2: Calculate the dot product of the corresponding elements of the positive sample feature A and the feature B to be verified. The calculation formula is: S6-3: Calculate the norm lengths of the corresponding elements of the positive sample feature A and the feature B to be verified. The calculation formula is: S6-4: Calculate the feature distance through the dot product and the norm length. The calculation formula is: Among them, the feature distance value will be between 0 and 2, and the closer the value is to 0, the more similar the features are.

7. A signature forgery detection method based on the running script trajectory according to claim 1, characterized in that The self-attention trajectory anomaly perception network described in S7 specifically includes the following structure: The input of this network is the feature to be verified, and the output is 0 and 1, where 0 represents "not an anomaly" and 1 represents "is an anomaly"; from the input to the output of this network, it successively includes a rightward LSTM, a leftward LSTM, and a self-attention feature fusion layer. The output of the rightward LSTM contains K hidden states, and the output of the leftward LSTM updates the K hidden states; among them, the rightward LSTM contains K time steps inside, and the leftward LSTM also contains K time steps inside. The structures of the rightward LSTM and the leftward LSTM are completely opposite, where K represents the total length of the time steps and the hidden states.

Images