Convenient signature authentication method for deep learning

Abstract

The invention relates to the technical field of deep learning and artificial intelligence, in particular to a convenient signature authentication method for deep learning. Comprising the steps of model training, data acquisition, image preprocessing, feature extraction and comparison, decision making and the like, and a deep learning model is deployed on a terminal, so that signatures of personnel in products can be quickly and accurately verified, the legality and authenticity of transactions are ensured, the security of the transactions is improved, and the products are protected from being influenced by forged signatures. In the feature extraction stage, deep mining is directly carried out through the feature learning network, multi-scale features of signatures are extracted, subtle differences are accurately captured, the training process is simplified, and the accuracy of feature extraction is improved. The network architecture of the model ensures that feature vectors are compared in the same feature space, the accuracy of similarity evaluation is improved, the features of real and forged signatures are learned and discriminated by comparing loss or triple loss, and the method is more direct and efficient.

Description

Technical Field

[0001] The present invention relates to the technical fields of deep learning and artificial intelligence, and particularly to a portable signature authentication method for deep learning. Background Art

[0002] With the continuous advancement of the digitalization process, the demand for product authenticity verification has been increasing. Especially in the field of fast-moving consumer goods, traditional manual verification methods can no longer meet the requirements of efficient and accurate authenticity verification, as they are inefficient and vulnerable to subjective factors. Therefore, there is an urgent need in the market for an automated and efficient authenticity verification technology to ensure product authenticity and enhance the overall security of the supply chain.

[0003] Currently, the existing technologies face the following challenges: 1. Real-time requirement: On the production line of fast-moving consumer goods, there is a very high requirement for real-time performance, while traditional authenticity verification methods often fail to meet this standard, thus restricting the improvement of production efficiency. 2. Environmental adaptability: Since product packaging may vary due to different production lines, materials, and printing processes, this poses a higher challenge to the adaptability and robustness of authenticity verification algorithms. 3. Difficulty in data collection: In the fast-moving consumer goods industry, collecting a large number of representative genuine and fake product packaging samples to train the authentication model is a daunting task, especially under the premise of ensuring intellectual property protection and data security. 4. Hardware resource limitations: On embedded terminal devices, limited computing power and storage space become constraints for deploying and running complex models.

[0004] In the prior art, such as Publication No. CN109344856A, an offline signature authentication method based on multi-layer discriminative feature learning, which uses a multi-layer discriminative feature learning neural network model and a binary classification SVM model for feature extraction and classification. The preprocessing includes Gaussian smoothing, OTSU binarization, etc. The processing is relatively simple, and the SVM model is used for the final authenticity identification, and its authenticity verification method is not efficient enough. Such as Publication No. JP2012099157A5, an image recognition extraction device, which calculates the feature quantities of different regions in the image and quantifies them into image identifiers; the process includes steps such as calculating regional feature quantities, determining thresholds, and quantization, and the efficiency for real-time signature authentication is insufficient. Such as Publication No. WO2023272993A1, a picture recognition method, device, equipment, and readable storage medium, which uses isomorphic branches and knowledge collaboration for auxiliary training to improve the performance of the feature extraction model. An isomorphic branch is introduced during the training process and the knowledge collaboration loss value is calculated. The training process includes steps such as extracting an isomorphic branch from the main network of the model, calculating the maximum inter-class distance and the minimum inter-class distance, etc. In the recognition stage, the recognition result is determined by calculating the vector distance between the image features and the labeled image features in the query dataset, and its authenticity verification method is not accurate enough. Summary of the Invention

[0005] In the process of authenticating the authenticity of the above products, the present invention solves the problems of insufficient real-time requirements, insufficient environmental adaptability, difficult data acquisition, and hardware resource limitations. A convenient signature authentication method based on deep learning is proposed. By deploying a deep learning model on the terminal, it can quickly and accurately verify the signatures of personnel in the products, ensure the legality and authenticity of transactions, improve the security of transactions, and protect the products from being affected by forged signatures. The technical solution is as follows:

[0006] A convenient signature authentication method based on deep learning includes the following steps:

[0007] S1. Train the real signature images and forged signature images through a feature learning network model;

[0008] S2. Collect the signature images to be authenticated through a data acquisition device;

[0009] S3. Preprocess the images to be authenticated in step S2 to meet the input requirements of the network model;

[0010] S4. Extract multi-scale features of the images to be authenticated through the feature learning network to generate a first feature vector of a fixed length; the multi-scale features include any one of the features of stroke thickness, curve change, local texture, and overall contour shape;

[0011] S5. Input the first feature vector of the images to be authenticated and the first feature vector of the real signature images into the trained network model to generate their respective corresponding second feature vectors. Evaluate the similarity by comparing the distances of the two second feature vectors, and judge the authenticity of the signature according to the similarity threshold.

[0012] Further, in step S2, collecting the images to be authenticated includes:

[0013] S21. Monitor touch events, capture the contact coordinates (x, y) and touch states;

[0014] S22. Use the drawing interface to draw the signature trajectory on the canvas in real time;

[0015] S23. After the signature is completed, render the content of the canvas as a bitmap and convert it into an image format or a byte array.

[0016] Further, step S4 includes:

[0017] S41. Capture multi-scale information of the signature through the feature learning network;

[0018] S42. Adjust the scale and threshold of the spectral amplitude of the feature learning network to balance the details and global features, and extract the multi-scale features of the signature; the spectral amplitude is the intensity of the change of the image features;

[0019] S43. Map the extracted multi-scale features to different dimensions of space to generate the first feature vector of a fixed length.

[0020] Further, the S1 step includes:

[0021] S11. Collect the genuine signatures of users as positive samples and the forged signatures of other users as negative samples for model training;

[0022] S12. Extract the multi-scale features of the positive samples and the negative samples to generate the first feature vector of a fixed length;

[0023] S13. Use a two-tower structure with shared weights to extract the first feature vector, and learn the features for discriminating genuine and forged signatures through contrastive loss or triplet loss to generate the second feature vector;

[0024] S14. Compare the similarity between the positive samples and the negative samples through the second feature vector to obtain a similarity threshold.

[0025] Further, the similarity comparison in the S14 step includes the following steps:

[0026] S141. For the group of genuine signatures, the goal of the feature learning network is to maximize their similarity;

[0027] S142. For the group of genuine signatures and forged signatures, the goal of the feature learning network is to minimize their similarity.

[0028] Further, the preprocessing in the S3 step is: Convert the image to be authenticated into a grayscale image and scale it to the model input size.

[0029] Even further, the preprocessing in the S3 step includes: Grayscale conversion, binarization, denoising, and normalization processing;

[0030] S31. Convert the image to be authenticated into a grayscale image and perform binarization;

[0031] S32. Extract the main features of the handwritten signature in the image and remove background noise;

[0032] S33. Apply Gaussian blur to smooth the image and remove salt-and-pepper noise through median filtering;

[0033] S34. Use morphological operations to eliminate small noise points and connected interferences;

[0034] S35. Scale the pixel values to meet the model requirements and adjust the size of the image to be authenticated.

[0035] Further, the S5 step includes:

[0036] S51. Input the first feature vectors of the image to be authenticated and the first feature vectors of the genuine signature images into two sub-networks with shared weights respectively to generate corresponding second feature vectors.

[0037] S52. Calculate the distance between the two second feature vectors to obtain the similarity.

[0038] S53. If the similarity score is lower than the similarity threshold, determine that the signature is a genuine signature; otherwise, determine that it is a forged signature.

[0039] Furthermore, the feature learning network is a Scattering wavelet transform or a CNN network, and the training model is a Siamese network architecture.

[0040] Furthermore, the CNN network includes a convolutional layer and a fully connected layer.

[0041] The present invention has the following beneficial effects:

[0042] 1. The portable signature authentication method based on deep learning according to the present invention has a process including model training, data collection, image preprocessing, feature extraction and comparison, decision-making, etc. In the feature extraction stage, it directly deeply mines through the feature learning network to extract multi-scale features of the signature, accurately captures subtle differences, simplifies the training process and improves the accuracy of feature extraction. The network architecture of the model ensures that the feature vectors are compared in the same feature space, improves the accuracy of similarity evaluation, and learns to distinguish the features of genuine and forged signatures through contrastive loss or triplet loss, and the method is more direct and efficient. Through the normalization process in the preprocessing stage, the data is unified, which is conducive to the model to learn and extract features, denoises and highlights key features, and improves the accuracy of feature extraction. Map the extracted multi-scale features to different dimensions of space to generate fixed-length feature vectors, which can map the multi-level features of the signature image into a unified vector form for better subsequent similarity comparison and analysis. Finally, the authentication is achieved by comparing the similarity of the feature vectors, and the technical means are more advanced, adaptable, fast, and more in line with the actual needs of signature authentication.

[0043] 2. For the portable signature authentication method based on deep learning of the present invention, it is difficult to collect a large number of real and forged signature samples. The signature authentication system of the present invention can still work effectively with limited training samples, reducing the dependence on a large number of training samples. Moreover, since an individual's signature may vary due to various factors such as writing speed, strength, and fatigue level, the present invention can adapt to these natural variations and accurately identify deliberately forged signatures. In addition, the model of the present invention is both lightweight and efficient, reducing the demand for hardware resources and enabling it to run smoothly on these devices without burdening the device's performance.

[0044] 3. For the portable signature authentication method based on deep learning of the present invention, the multi-scale information of the signature is captured through a feature learning network, which can effectively extract the detailed features and global features in the signature image. By adjusting the threshold, noise and weak features are filtered to ensure the extraction of important features.

[0045] 4. For the portable signature authentication method based on deep learning of the present invention, preprocessing is performed on the signature image, including denoising, morphological operations, etc., to process the image data more comprehensively, effectively improving the accuracy of feature extraction, making the input dimension match the model, and ensuring that the data can be correctly input into the target model for subsequent processing.

[0046] 5. For the portable signature authentication method based on deep learning of the present invention, by combining the Scattering wavelet transform and the Siamese network architecture, the multi-scale features of the signature are directly extracted through the Scattering wavelet transform first, and then the similarity of the feature vectors is compared using the Siamese network, which can achieve more accurate signature authentication. The Scattering wavelet transform introduces non-linear operations and multi-layer structures compared with the wavelet transform, enhancing the stability and robustness of feature expression. During the feature extraction process of the signature image, by adjusting the scale information, the model can control the sensitivity of the features in the transform, enabling different levels of details to be captured at different scales. By effectively balancing the global and local information of the multi-scale information, it is ensured that the features reflecting the individual differences in writing, such as stroke connection form, strength change, writing speed, etc., can be accurately extracted, thus showing unique distinguishing features in signature comparison and avoiding the situation of blurring or overfitting. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a schematic flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0048] The signature of the present invention is not just a signature representing a name. In essence, it is a combination of biological behavior characteristics (physiological habits) and personalized symbols (cultural expressions). The similarity score is used to evaluate the matching degree between the signature image to be authenticated and the genuine signature sample by their similarity. The connection between similarity and score is based on the fact that the distance or similarity metric value reflects the difference or similarity between the signature to be authenticated and the genuine signature. For example, the smaller the Euclidean distance, the closer the two feature vectors are, and the more similar the signatures are. However, similarity does not directly mean that the signature is genuine, because similarity only indicates the degree of similarity between two signatures and cannot directly determine whether it is a valid signature. The present invention uses the distance between two feature vectors to set the similarity threshold judgment criterion. When the calculated distance is lower than this threshold, the system will determine that the signature is a genuine signature; otherwise, it will be determined as a forged signature. The following will combine the attached Figure 1 figures and specific embodiments to elaborate on the present invention in detail.

[0049] Embodiment 1.

[0050] A convenient signature authentication method based on deep learning, with the feature learning network being the Scattering wavelet transform, the training model being the Siamese network architecture, and the data acquisition device being a portable device. The Scattering wavelet transform simulates a certain "scattering" of "waves" through the process of layer-by-layer wavelet filtering and pooling to describe the change of information distribution. The amplitude of the Scattering wavelet transform is represented by the spectral amplitude after wavelet transform, and the threshold is adjusted to filter noise / weak features, and the threshold is obtained through experiments. By appropriately adjusting the scale and threshold of the Scattering wavelet, these dynamic changes can be captured, balancing details and global features, so as to accurately extract the multi-scale features of the signature. For example, the change in writing speed will affect the distribution of the feature spectrum, and the high-frequency components increase during fast writing; the change in force is reflected as the intensity difference in different regions in the amplitude of the Scattering wavelet; the stroke irregularity caused by fatigue will be reflected in the wavelet spectrum through local texture and contour changes. The Siamese network architecture is a neural network specifically used to compare the similarity of two input samples. The Siamese network consists of two identical sub-networks, which share the same weights and parameters. The signature image to be authenticated and the genuine signature sample are respectively input into the two sub-networks of the Siamese network to extract their feature vectors. These feature vectors are extracted through hierarchical structures such as convolutional layers and fully connected layers in the network, and are usually represented as high-dimensional vectors, with each element capturing different feature information in the image. It includes the following steps:

[0051] S1, use the Siamese network to train the feature learning network model on the genuine signature image and the forged signature image to extract features and obtain the similarity threshold.

[0052] S2. Collect the signature image to be authenticated through a data acquisition device.

[0053] S3. Preprocess the image to be authenticated to meet the input requirements of the network model; specifically, convert the image to be authenticated into a grayscale image and scale it to the model input size.

[0054] S4. Extract multi-scale features of the image to be authenticated through the feature learning network to generate a first feature vector of a fixed length; the multi-scale features include any one of the features of stroke thickness, curve change, local texture, and overall contour shape.

[0055] S5. Input the first feature vector of the image to be authenticated and the first feature vector of the genuine signature image into two sub-networks sharing weights in the Siamese network respectively to generate their corresponding second feature vectors, compare the similarity of the two second feature vectors through distance measurement, and judge the authenticity of the signature according to the similarity threshold.

[0056] The collection of the image to be authenticated in step S2 includes:

[0057] S21. Listen for touch events, capture the touch point coordinates (x, y) and touch states (start, move, end), and record continuous trajectory data.

[0058] S22. Use a drawing interface such as Canvas.drawPath() to draw the signature trajectory on the canvas in real time and dynamically display the signature effect.

[0059] S23. After the signature is completed, render the canvas content as a bitmap, write the signature trajectory into the bitmap by creating an image object and combining with Canvas, and convert the bitmap into an image format or byte array through methods such as Bitmap.compress().

[0060] Step S4 includes:

[0061] S41. Capture multi-scale information of the signature through the feature learning network. This process effectively extracts the detailed features and global features in the signature image through analysis at different scales.

[0062] S42. Adjust the scale and threshold of the spectral amplitude of the feature learning network to balance the detailed and global features and extract the multi-scale features of the signature; the spectral amplitude is the intensity of image feature change, and noise and weak features are filtered by adjusting the threshold to ensure the extraction of important features.

[0063] S43. Map the extracted multi-scale features to different dimensions of space to generate the first feature vector of a fixed length. Through this process, the multi-level features of the signature image can be mapped into a unified vector form for subsequent comparison and analysis.

[0064] Step S1 includes:

[0065] S11. Collect the genuine signatures of users as positive samples and the forged signatures of other users as negative samples for model training.

[0066] S12. Extract the multi-scale features of the positive samples and the negative samples to generate the first feature vector of a fixed length.

[0067] S13. Use a two-tower structure with shared weights to extract the first feature vector. Through contrastive loss or triplet loss, learn the features for discriminating genuine and forged signatures to generate the second feature vector.

[0068] S14. Compare the similarity between the positive samples and the negative samples through the feature vectors to obtain a similarity threshold.

[0069] S141. For the group of genuine signatures, the goal of the feature learning network is to maximize their similarity.

[0070] S142. For the group of genuine signatures and forged signatures, the goal of the feature learning network is to minimize their similarity.

[0071] Step S5 includes:

[0072] S51. Input the first feature vector of the image to be authenticated and the first feature vector of the genuine signature image into two sub-networks with shared weights respectively to generate corresponding second feature vectors.

[0073] S52. Calculate the distance between the two feature vectors to obtain the similarity.

[0074] S53. If the similarity score is lower than the similarity threshold, then determine that the signature is a genuine signature; otherwise, determine that it is a forged signature.

[0075] Embodiment 2.

[0076] A convenient signature authentication method based on deep learning. On the basis of Embodiment 1, optimize the acquisition accuracy of the data acquisition device and further optimize the preprocessing in Step S3, including the following steps:

[0077] S1. Use a Siamese network to train a feature learning network model for genuine signature images and forged signature images.

[0078] S2. Collect the signature image to be authenticated through a data acquisition device. The data acquisition device is equipped with or linked to a high-precision touch screen, which can be used to collect the touch points during the user's signature process and ensure accurate recording of the minute changes during the writing process. The data acquisition accuracy should be set relatively high, such as a sampling rate of 100 Hz or higher.

[0079] S3. Preprocess the image to be authenticated in step S2 to meet the input requirements of the network model.

[0080] S4. Extract multi-scale features of the image to be authenticated through the feature learning network to generate a first feature vector of a fixed length; the multi-scale features include: stroke thickness, curve variation, local texture, and overall contour shape.

[0081] S5. Input the feature vector of the image to be authenticated and the first feature vector of the genuine signature image into two sub-networks of the trained network model with shared weights respectively to generate their corresponding second feature vectors. Evaluate the similarity by comparing the distances of the two second feature vectors, and judge the authenticity of the signature according to the similarity threshold.

[0082] Collecting the image to be authenticated in step S2 includes:

[0083] S21. Listen for touch events in the blank area or signature canvas of the data acquisition device, capture the contact coordinates (x, y) and touch states (start, move, end) during the user's writing, and record the continuous trajectory data; the touch states, such as the pressure, speed, and trajectory during the touch process. In particular, the pressure sensing function can effectively distinguish the changes in writing force, and the real-time recording of the touch timestamp can capture the writing speed. The point intervals are smaller during fast writing and longer during slow writing. At the same time, continuous collection of signature data can reflect the fatigue degree of the writer. When fatigued, the strokes become irregular or lack smoothness.

[0084] S22. Use a drawing interface such as Canvas.drawPath() to draw the signature trajectory on the canvas in real time and dynamically display the signature effect.

[0085] S23. After the signature is completed, render the canvas content as a bitmap. Write the signature trajectory into the bitmap by creating an image object and combining with Canvas, and convert the bitmap into an image format or byte array through methods such as Bitmap.compress().

[0086] The preprocessing in step S3 includes: grayscale conversion, binarization, denoising, and normalization processing.

[0087] S31. Convert the image to be authenticated into a grayscale image for binarization.

[0088] S32. Extract the main features of the handwritten signature in the image and remove background noise.

[0089] S33. Apply Gaussian blur to smooth the image and remove salt-and-pepper noise through median filtering.

[0090] S34. Use morphological operations, such as opening and closing operations, to eliminate small noise points and connected interferences.

[0091] S35. Normalize the data to 0 - 1, scale the pixel values to meet the requirements of the model, adjust the size of the image to be authenticated, and ensure that the input dimension matches the model, so as to ensure that the data can be correctly input into the target model for subsequent processing.

[0092] The S4 step includes:

[0093] S41. Capture the multi-scale information of the signature through the feature learning network. Through analysis at different scales, the detailed features and global features in the signature image are effectively extracted.

[0094] S42. Adjust the scale and threshold of the spectral amplitude of the feature learning network to balance the detailed features and global features, and extract the multi-scale features of the signature; the spectral amplitude is the intensity of the change of image features. By adjusting the threshold, noise and weak features are filtered to ensure the extraction of important features.

[0095] S43. Map the extracted multi-scale features to different dimensions in space to generate the first feature vector with a fixed length. Through this process, the multi-level features of the signature image can be mapped into a unified vector form for subsequent comparison and analysis.

[0096] The S1 step includes:

[0097] S11. Collect the real signatures of users as positive samples and the forged signatures of other users as negative samples for model training.

[0098] S12. Extract the multi-scale features of the positive samples and the negative samples to generate the first feature vector with a fixed length.

[0099] S13. Use a two-tower structure with shared weights to extract the first feature vector. Through contrastive loss or triplet loss, learn the features for discriminating real and forged signatures, and generate the second feature vector.

[0100] S14. Compare the similarity between the positive samples and the negative samples through the feature vectors to obtain the similarity threshold.

[0101] S141. For the real signature group, the goal of the feature learning network is to maximize their similarity.

[0102] S142, the real signature and the forged signature group, and the goal of the feature learning network is to minimize their similarity.

[0103] The similarity comparison in step S13 includes the following steps:

[0104] Step S5 includes:

[0105] S51, input the first feature vector of the image to be authenticated and the first feature vector of the real signature image into two sub-networks sharing weights of the network model respectively to generate corresponding second feature vectors.

[0106] S52, calculate the distance between the two feature vectors to obtain the similarity.

[0107] S53, if the similarity score is lower than the similarity threshold, it is determined that the signature is a real signature, otherwise it is determined to be a forged signature.

[0108] Embodiment III.

[0109] A convenient signature authentication method based on deep learning, the feature learning network is a CNN network, the training model is a Siamese network architecture, and the data acquisition device is a device connected or provided with a high-precision touch screen that can be used to collect the touch points during the user's signature process. It includes the following steps:

[0110] S1, use the Siamese network to train the feature learning network model for real signature images and forged signature images.

[0111] S2, collect the signature image to be authenticated through the data acquisition device.

[0112] S3, preprocess the image to be authenticated in step S2 to meet the input requirements of the network model.

[0113] S4, extract multi-scale features of the signature image through the CNN network to generate a fixed-length feature vector; the multi-scale features include any one of stroke thickness, curve change, local texture and overall contour shape; the CNN network includes a convolutional layer and a fully connected layer.

[0114] S5, input the feature vector of the image to be authenticated and the first feature vector of the real signature image into two sub-networks sharing weights of the trained Siamese network respectively. Inside the network, a series of hierarchical calculations are performed to generate second feature vectors. The similarity is evaluated by comparing the distances between the two second feature vectors, and the authenticity of the signature is judged according to the similarity threshold.

[0115] The collection of the signature image to be authenticated in step S2 includes:

[0116] S21, Monitor touch events, capture the coordinates (x, y) of the touch point and the touch state (start, move, end), and record the continuous trajectory data.

[0117] S22, Use the drawing interface to draw the signature trajectory on the canvas in real time and dynamically display the signature effect.

[0118] S23, After the signature is completed, render the content of the canvas as a bitmap and convert the bitmap into an image format or a byte array.

[0119] The preprocessing described in step S3 includes grayscale conversion, binarization, denoising, and normalization processing.

[0120] S31, Convert the image to be authenticated into a grayscale image for binarization.

[0121] S32, Extract the main features of the handwritten signature in the image and remove the background noise.

[0122] S33, Apply Gaussian blur to smooth the image and remove salt-and-pepper noise through median filtering.

[0123] S34, Use morphological operations to eliminate small noise points and connected interferences.

[0124] S35, Scale the pixel values to the requirements of the model and adjust the size of the image to be authenticated.

[0125] Step S4 includes:

[0126] S41, Extract multi-scale information of the signature image through the CNN network. This process performs multi-scale processing on the image through convolutional layers and can automatically extract the detailed features and global features in the image. Through convolutional kernels at different levels, the CNN network can effectively capture various details of the signature image.

[0127] S42, Adjust the parameters of the convolutional kernels and pooling layers of the CNN network to balance the detailed features and global features and extract the multi-scale features of the signature; the output feature map of the convolutional layer represents the feature changes of the image. Through appropriate pooling operations, dimensionality reduction and feature aggregation are performed to highlight the key features of the image and filter out unnecessary information.

[0128] S43, Map the extracted multi-scale features to different dimensions in space to generate the first feature vector of a fixed length. Through this process, the CNN network converts the multi-level features of the signature image into a unified feature vector for subsequent comparison and analysis.

[0129] Step S1 includes:

[0130] S11, Collect the multi-scale features of the user's real signature as positive samples and the multi-scale features of other users' forged signatures as negative samples for model training.

[0131] S12. Extract the multi-scale features of the positive samples and the negative samples to generate the feature vectors of a fixed length.

[0132] S13. Through contrastive loss or triplet loss, the network learns the features for discriminating real and forged signatures.

[0133] S14. Compare the similarity between the positive samples and the negative samples through the feature vectors to obtain a similarity threshold.

[0134] S141. For the real signature group, the goal of the feature learning network is to maximize their similarity.

[0135] S142. For the real signature and the forged signature group, the goal of the feature learning network is to minimize their similarity. The similarity comparison in step S13 includes the following steps:

[0136] Step S5 includes:

[0137] S51. Respectively input the first feature vector of the image to be identified and the first feature vector of the real signature image into two sub-networks with shared weights of the trained Siamese network to generate corresponding second feature vectors.

[0138] S52. Calculate the distance between the two feature vectors to obtain the similarity.

[0139] S53. If the similarity score is lower than the similarity threshold, it is determined that the signature is a real signature; otherwise, it is determined to be a forged signature.

[0140] The above are only embodiments of the present invention. The selection of the embodiment solutions is only for better understanding the content of the invention, and does not limit the patent scope of the present invention. Any equivalent structural or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be similarly included in the patent protection scope of the present invention.

Claims

1. A deep learning-based convenient signature authentication method, characterized in that: The following steps are involved: S1, train the real signature image and the forged signature image through the feature learning network model; S2, collecting the signature image to be authenticated through a data collection device; S3, preprocessing the image to be identified in step S2 to meet the input requirements of the network model; S4, extracting multi-scale features of the image to be identified through the feature learning network to generate a first feature vector of fixed length; the multi-scale features include: any one of stroke thickness, curve change, local texture and overall contour morphology; S5, inputting the first feature vector of the image to be identified and the first feature vector of the real signature image into the trained network model respectively to generate the second feature vectors corresponding to each other, evaluating the similarity by comparing the distance between the two second feature vectors, and judging the authenticity of the signature according to the similarity threshold.

2. According to claim 1, a deep learning portable signature authentication method is characterized in that: The acquisition of the image to be identified in step S2 includes: S21, monitor touch events, capture touch point coordinates (x, y) and touch status; S22, using a drawing interface to draw the signature track on the canvas in real time; S23, after the signing is completed, the canvas content is rendered into a bitmap and converted into an image format or a byte array.

3. According to claim 1, a deep learning portable signature authentication method is characterized in that: The S4 step includes: S41, capturing multi-scale information of the signature through the feature learning network; S42, adjusting the scale and threshold of the spectrum amplitude of the feature learning network, balancing details and global features, and extracting multi-scale features of the signature; the spectrum amplitude is the intensity of the change of image features; S43: Map the extracted multiple scale features to different dimensions of space to generate the first feature vector of fixed length.

4. According to claim 1, a deep learning portable signature authentication method is characterized in that: The S1 step includes: S11, collect the user's real signature as positive sample and other users' forged signature as negative sample for model training; S12, extracting multi-scale features of the positive sample and the negative sample to generate the first feature vector of fixed length; S13, extracting the first feature vector by using a dual-tower structure with shared weights, learning features for distinguishing between real and forged signatures through contrast loss or triplet loss, and generating the second feature vector; S14: Compare the similarity between the positive sample and the negative sample using the second feature vector to obtain a similarity threshold.

5. A deep learning portable signature authentication method according to claim 4, characterized in that: The similarity comparison in step S14 includes the following steps: S141, a group of real signatures, the goal of the feature learning network is to maximize their similarity; S142, a group of real signatures and forged signatures, the goal of the feature learning network is to minimize their similarity.

6. A deep learning portable signature authentication method according to claim 1, characterized in that: The preprocessing of step S3 is: converting the image to be identified into a grayscale image and scaling it to the model input size.

7. A deep learning portable signature authentication method according to claim 1, characterized in that: Furthermore, the preprocessing of step S3 includes: grayscale, binarization, denoising and normalization processing; S31, converting the image to be identified into a grayscale image for binarization; S32, extracting the main features of the handwritten signature in the image and removing background noise; S33, applying Gaussian blur to smooth the image and removing salt and pepper noise by median filtering; S34, uses morphological operations to remove small noise points and connected interference; S35, scaling the pixel values ​​to the model requirements, and adjusting the size of the image to be identified.

8. A deep learning portable signature authentication method according to claim 1, characterized in that: Step S5 includes: S51, inputting the first feature vector of the image to be identified and the first feature vector of the real signature image into two sub-networks with shared weights respectively to generate corresponding second feature vectors; S52, calculating the distance between the two second feature vectors to obtain the similarity; S53: If the similarity score is lower than the similarity threshold, the signature is determined to be a genuine signature; otherwise, it is determined to be a forged signature.

9. A deep learning portable signature authentication method according to claim 1, characterized in that: The feature learning network is a Scattering wavelet transform or a CNN network, and the training model is a Siamese network architecture.

10. A deep learning portable signature authentication method according to claim 9, characterized in that: The CNN network includes a convolutional layer and a fully connected layer.

Images