CN113779643A - Signature handwriting recognition system and method based on pre-training technology and storage medium - Google Patents
Signature handwriting recognition system and method based on pre-training technology and storage medium Download PDFInfo
- Publication number
- CN113779643A CN113779643A CN202111122613.9A CN202111122613A CN113779643A CN 113779643 A CN113779643 A CN 113779643A CN 202111122613 A CN202111122613 A CN 202111122613A CN 113779643 A CN113779643 A CN 113779643A
- Authority
- CN
- China
- Prior art keywords
- signature
- data
- sample
- encoder
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000012549 training Methods 0.000 title claims abstract description 68
- 238000000034 method Methods 0.000 title claims abstract description 39
- 238000005516 engineering process Methods 0.000 title claims abstract description 14
- 238000012952 Resampling Methods 0.000 claims abstract description 7
- 230000006870 function Effects 0.000 claims description 54
- 239000013598 vector Substances 0.000 claims description 25
- 238000003062 neural network model Methods 0.000 claims description 13
- 238000007781 pre-processing Methods 0.000 claims description 10
- 238000013528 artificial neural network Methods 0.000 claims description 9
- 230000000873 masking effect Effects 0.000 claims description 7
- 238000010606 normalization Methods 0.000 claims description 6
- 238000013459 approach Methods 0.000 claims description 3
- 238000004590 computer program Methods 0.000 claims description 3
- 238000009795 derivation Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 238000011478 gradient descent method Methods 0.000 description 4
- 238000012545 processing Methods 0.000 description 4
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000004422 calculation algorithm Methods 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000012512 characterization method Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000013526 transfer learning Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F21/60—Protecting data
- G06F21/64—Protecting data integrity, e.g. using checksums, certificates or signatures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Security & Cryptography (AREA)
- Bioethics (AREA)
- Computer Hardware Design (AREA)
- Collating Specific Patterns (AREA)
- Character Discrimination (AREA)
Abstract
An electronic signature recognition system and method based on a pre-training technology relate to an electronic signature handwriting recognition technology, and the system and method are used for collecting electronic signature stroke related characteristic data, and acquiring signature sample data and signature to-be-detected data; the method comprises the steps of resampling, normalizing and discretizing signature sample data and data to be signed to obtain data with fixed length, respectively inputting the signature sample data and the data to be signed with the fixed length into a sample encoder, determining common characteristics and difference characteristics of the signature sample data and the data to be signed by the encoder through a cosine similarity loss function, and determining that the signature sample data and the data to be signed come from the same signer or different signers. The invention is widely used in places such as e-commerce, e-government affairs and the like needing signature identification.
Description
Technical Field
The invention relates to the technical field of computer information processing, in particular to a handwritten electronic signature comparison method based on a neural network pre-training technology.
Background
The traditional document and protocol signing depends on signature, the signature is often also a judicial identification mode, and the electronic handwritten signature is brought into a remote evidence obtaining system in the court trial process, so the accuracy authentication of a signature comparison algorithm brings convenience to the judicial justice, and the appearance of a new signature comparison authentication technology is more and more urgent. In the handwritten electronic signature comparison neighborhood, the same person can write different signature tracks called as intra-individual difference, and a forger with worry can write extremely similar tracks called as inter-individual identity, so that difficulty is caused in image signature comparison.
Chinese patent application publication No. CN112560636A, "a handwritten signature comparison method and system based on image recognition" proposes signature comparison by a method of comparing with an image in a database, which is poor in generalization capability. The proposed method uses too single characteristics to fully utilize the characteristics acquired by the acquisition plate, such as pressure, rotation angle, etc. The Chinese patent application with the publication number of CN109409254A discloses an electronic contract handwritten signature identification method based on a twin neural network, and discloses an electronic contract handwritten signature identification method based on a twin neural network, wherein a plurality of times of user handwritten signature background reservation is carried out in a real-name authentication link, and a handwritten signature is stored as an image file; selecting two of the handwritten signature image files and one of the handwritten signature image files which is not the user, inputting the two handwritten signature image files and the one of the handwritten signature image files into a twin neural network model for vectorization calculation to obtain a loss function, and obtaining a training set cost function through the loss function; training other handwritten signature image files of the user through an optimization algorithm, and outputting an identification model; when signing online, the handwritten signature on the touch screen input device is instantly stored as an image, the image is input into the identification model, the identification result is output, and the storage and identification work of the handwritten signature is completed. However, the method adopts the image file to form the training set, does not use a large amount of characteristic information of the handwriting, and has low recognition accuracy.
CN 113158887 a is an electronic signature authentication method and device for improving the identification accuracy of electronic signatures. Collecting signature information including a timestamp, coordinate information and pressure information of each pixel point of an electronic signature track; restoring the signature track on the signature track layer according to the coordinate information of the signature track; acquiring a pressure value from the pressure information, and setting the pressure value on a pixel point corresponding to the coordinate information of the signature track on the signature pressure layer; combining the signature track layer and the signature pressure layer to generate a feature map; inputting the characteristic graph into a trained convolutional neural network, randomly selecting the characteristic graphs corresponding to two signature information from a signature information data set to form a data pair, marking the data pair, distinguishing whether the two signature information in the data pair are both correct signatures through marking, building a convolutional neural network model, comparing the characteristic graph with the signature template graph of the signature through the convolutional neural network, and outputting an authentication result. Although the patent application fully utilizes the collected characteristics of the electronic version, supervised learning consumes a large amount of manpower and resources, and the industrial cost is greatly increased.
Most electronic signature comparison technologies are handwriting around electronic signatures, an image processing mode is adopted, but the same signer can sign different signatures at different time, a counterfeiter with worry can also forge highly similar signatures, along with the development of the technology, the electronic signature board can capture more features, and a large amount of information is omitted through an authentication mode of image signature comparison. In the conventional method, such as deephsv, idn and the like, image data is used as input data, handwriting comparison is used as a two-classification processing problem based on images, and a large number of positive and negative sample learning models are needed. However, it is very difficult to collect high quality negative sample data, resulting in a low accuracy of signature identification.
In addition, the conventional handwritten electronic signature comparison method classifies input sample images and statement images, and the recognition method is not suitable for large-scale handwriting retrieval tasks and other applications due to reasons such as calculated amount, so that the conventional handwriting comparison technology based on images mainly aims at modeling track features and does not fully utilize collected rich multi-dimensional writing feature information. The neural network model with high complexity has stronger identification capability, but the neural network model with high complexity easily causes the problem of overfitting on small data, and the lack of negative samples causes the signature identification accuracy to be difficult to improve.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides an electronic signature comparison method based on a pre-training technology, which can solve the problem of overfitting of a small amount of sample data on a large model, and uses a method for enhancing noise data for the purpose of data sufficiency to solve the problem of lack of negative samples.
The invention provides an electronic signature handwriting recognition system based on a pre-training technology, which comprises a data acquisition module, a data preprocessing module, a signature handwriting pre-training module and a signature handwriting characteristic recognition module, wherein the data acquisition module acquires electronic signature stroke related characteristic data to acquire signature sample data; the data preprocessing module carries out resampling, normalization and discretization on the collected positive sample data to generate negative sample data N for electronic signature handwriting pre-training, a neural network is trained in an autoregressive pre-training mode of the positive sample data P based on a mask, and a neural network model after the pre-training is finished serves as an encoder E; the signature handwriting pre-training module trains an electronic signature handwriting pre-training model by combining a classification loss function with positive sample data and a noise regression loss function with negative sample data respectively to obtain a sample encoder 1 and a sample encoder 2 which have the same structure; and updating the weight parameters of the two encoders until a convergence weight is obtained, and constructing a twin network model of the sample encoder which comprises two same structures and shares the weight as a signature handwriting characteristic identification module.
Further, from the sample encoder 1 output vector y and the sample encoder 2 output vector y ', using the cosine distance Loss _ match ═ (cos (y, y'))2And constructing a classification loss function: loss _ positive ═(1-cos(y,y'))2And noise regression function: loss _ noise ═ 1-cos (y', y) - μ)2. Updating the weight parameters of the two weight-shared encoders further comprises calling a formula using a classification loss function
Calculating the partial derivative of the current weight coefficient W of the encoder 1, and calling a formula W through a preset learning rate lrnew1＝w-lr×W1' update encoder 1 weight parameters until convergence. Calling a formula using a noise regression function
Calculating the partial derivative of the current weight coefficient W of the encoder 2, and calling the formula W according to the preset learning rate lrnew2＝w-lr×W2' update encoder 2 weight parameters until convergence; and obtaining the weight parameters of the two sample encoders, and obtaining the sharing weight of the sharing weight sample encoder in the signature handwriting characteristic identification module by a gradient average or addition synchronization method for the convergence weight parameters of the two encoders.
The training of the neural network in the autoregressive pre-training mode further comprises: and (3) masking partial data sequence by using random mask to the original characteristic sequence of the acquired signature handwriting, inputting the partial data sequence into a neural network model for pre-training, outputting the pre-training characteristic sequence of the signature handwriting by the neural network model, enabling the model to output the pre-training characteristic sequence to approach the original characteristic sequence, and completing the pre-training.
The invention also provides an electronic signature handwriting recognition method based on the pre-training technology, which comprises the steps that a data acquisition module acquires the relevant characteristic data of the electronic signature handwriting, and signature sample data and signature to-be-detected data are acquired; the data preprocessing module is used for resampling, normalizing and discretizing the signature sample data and the signature to-be-detected data to obtain signature sample data and signature to-be-detected data with fixed lengths; respectively training an electronic signature handwriting pre-training model by using a classification loss function in combination with positive sample data and using a noise regression function in combination with negative sample data, and constructing a signature identification module comprising a sample encoder 1 and a sample encoder 2 which have the same structure and share a weight; the method comprises the steps that signature sample data and to-be-detected signature data with fixed lengths are input into a signature identification module, a sample encoder 1 and a sample encoder 2 respectively determine the common characteristic and the difference characteristic of the signature sample data and the to-be-detected signature data by utilizing a classification loss function and a noise regression function, and the signature sample data and the to-be-detected signature data are determined to be from the same signer or different signers.
Further, from the sample encoder 1 output vector y and the sample encoder 2 output vector y ', using the cosine distance Loss _ match ═ (cos (y, y'))2Determine Loss function Loss _ positive (1-cos (y, y'))2And acquiring the common characteristic of the signature sample characteristic and the signature to-be-detected characteristic.
Further, based on the output vector y of the sample encoder 1 and the output vector y 'of the sample encoder 2, using the formula Loss _ noise ═ 1-cos (y', y) -C)2And determining a noise regression function, and acquiring the difference characteristic between the signature sample characteristic and the signature to-be-detected characteristic.
And obtaining a weight parameter by performing bias derivation on the weight parameters W of the sample encoder 1 and the sample encoder 2 through a classification loss function and a noise regression function, and obtaining a shared weight for the obtained weight parameters of the two sample encoders through a gradient average or addition synchronization method.
Reducing the loss values of the two loss functions by adopting a gradient descent method, obtaining a parameter W 'by performing partial derivation on a weight parameter W of an encoder of the signature identification module, and obtaining a parameter W' according to a formula WnewThe weight parameters of the encoder are updated W-lrW'.
The invention also claims a computer-readable storage medium on which a computer program is stored which can be loaded and executed by a processor to perform the above-described method.
The single handwriting vector is obtained through the representation learning, and the vector is easy to use and more convenient in the electronic signature comparison stage. Abundant multi-dimensional writing characteristic information collected is fully utilized, a small amount of positive and negative sample data can be utilized, the problem that small data easily causes overfitting is solved, and signature identification accuracy is improved.
Drawings
FIG. 1 is a schematic diagram of an electronic signature comparison system;
FIG. 2 is a schematic diagram of a pre-trained electronic signature recognition model;
FIG. 3 is a graph of a handwriting pre-training model trained using a classification loss function in conjunction with positive sample data;
FIG. 4 trains a handwriting pre-training model using a noise regression loss function in conjunction with negative sample data.
Detailed Description
Fig. 1 is a schematic diagram of a signature recognition system according to the present invention, which includes a data acquisition module, a data preprocessing module, a data pre-training module, and a signature feature recognition module. The data acquisition module acquires the original handwriting sequence characteristic data of the electronic signature as positive sample data, wherein the sequence characteristic data comprise handwriting characteristic information such as an abscissa X, an ordinate Y, pressure P, signature process duration T and the like of the original handwriting of the signature; the preprocessing module resamples the signature original handwriting characteristic data, the data length is fixed to the same length, normalization and discretization are carried out, and a characteristic data sequence with a fixed length is obtained; the data pre-training module comprises a model pre-training mode in an autoregressive mode and a model pre-training mode in a noise regression learning mode based on a large amount of sample data, and then the low-sample transfer learning is achieved; and the signature characteristic identification module compares and identifies the original handwriting characteristics of the random signature with the original handwriting characteristics of the original signature, and determines the real identity of the signer according to whether the two original handwriting characteristics are consistent.
Fig. 2 is a schematic diagram of a pre-trained electronic signature recognition model.
Firstly, a signature handwriting pre-training module trains an electronic signature handwriting pre-training model by respectively using a classification loss function and a noise regression loss function in combination with positive sample data to obtain a sample encoder 1 and a sample encoder 2 with the same structure, and a twin network model of the sample encoder which comprises the same structure and shares weight is constructed as a signature handwriting feature recognition module. The two encoders can be regarded as a copy of the same model due to the weight sharing. Gradient finding due to two encoder inputs in backward propagationDifferent gradients are obtained, and the gradients of the two encoders are considered together when updating the network, and the common gradient is obtained by averaging or adding, and then the model is updated. Receiving data of a data acquisition module and generated noise enhancement data, converting the corresponding data into vectors, sending the respective converted vectors into an encoder 1 and an encoder 2, outputting twin network learning handwriting characterization formed by the encoder 1 and the encoder 2 through an output layer, and bringing the output of the output layer into a noise regression function Loss _ noise ═ 1-cos (y', y) -mu)2In the medium-pre-training signature feature recognition model, differences of data vectors are output through pre-training encoders 1 and 2, but commonalities of signatures of the same signer cannot be learned, so that a classification Loss function Loss _ positive (1-cos (y, y')) with zero noise is adopted2And pre-training a signature feature recognition model, wherein the encoder 1 and the encoder 2 output loss values between data feature vectors, and the loss values represent the difference between signature sample data and signature to-be-detected data. The noise regression function and the classification loss function are reduced by a gradient descent method so that the loss values of the two are close to 0. The specific gradient descent method is that the classification loss function and the noise regression function can be used for solving the partial derivative W 'of the current weight parameter of the twin network, and the weight parameter of the encoder is updated according to the formula Wnew-W-lrW' by setting a certain learning rate lr until the weight parameter is converged.
The data acquisition module acquires various handwritten signature characteristic information and data, and can adopt the following data acquisition modes: c + + sign board drive; JS H5 mobile terminal, Android APP mobile terminal. The data acquisition module can acquire sequence data of a plurality of characteristics such as a signature abscissa X, a signature ordinate Y, a signature pressure P, a signature time T, an angular speed and the like.
In the data preprocessing stage, the raw sequence data is subjected to normalization (such as batch normalization and layer normalization), resampling, discretization, data enhancement, data generation and other operations to obtain preprocessed data. The length of the generated sequence data is different due to different writing time of writers, the input length adopts a fixed value, original data is resampled, original signature data is fixed to a fixed length, when the data length of the sampling signature data sequence is smaller than the input length, upsampling is carried out in a zero padding mode, and when the data length is larger than the input length, the data length of the sampling data sequence is downsampled to the fixed length of the input length. Obtaining positive sample data with fixed length, adding Gaussian random noise to the obtained fixed length data sequence as negative sample data
And training a neural network by adopting an autoregressive pretraining mode based on a mask, wherein a trained neural network model is used as a coding module, and the coding module comprises two signature recognition encoders with the same structure. The neural network model can be based on cnn, lstm, transformer models or their superposition. And (3) masking partial data sequence by using random mask to the original characteristic sequence of the acquired signature handwriting, inputting the partial data sequence into a neural network model for pre-training, and outputting the pre-training characteristic sequence of the signature handwriting by the neural network model so as to finish pre-training when the pre-training characteristic sequence output by the model approaches to the original characteristic sequence.
The training of the neural network in the autoregressive pretraining manner specifically includes using the autoregressive manner to complete pretraining, so that the target of the model output data sequence is an input data sequence, and the data processed by the preprocessing module exists in a sequence form, for example, X ═ X (X ═ X)1、x2...xn) For the abscissa of the written trace, Y ═ Y1,y2...yn) As the ordinate of the signature script, P ═ P1,p2,..pn) To sign the pressure of the handwriting, T ═ T1,t2,...tn) The writing time is. Masking partial data by random masking, masking partial sequences in the data sequence characteristics after masking processing, wherein the corresponding characteristic data are as follows: x '((X1, X2.. mask _ token.. xn)), Y' ((Y1, y2... mask _ token.. yn), P '((P1, p2... mask _ token.. pn), T' ((T1, T2.. mask _ token.. tn)). Partial features in the sequence are covered by a mask _ token, and various feature data sequences with partial feature values covered by masks are input into a pre-training model. That is, the output sequence is a signature sequence X ═ X (X1, X2,. X.. xn), and a signature sequence y (y ═ X.. xn)1,y2...y...yn) Characteristic sequence (p)1,p2.._p...pn)，(t1,t2,...tn) And finishing pre-training.
The pre-trained model is used as an encoder of an electronic signature handwriting pre-training model to construct a signature feature recognition model, and the signature feature recognition model comprises two pre-training encoders (an encoder 1 and an encoder 2) with the same structure. And updating the weight parameters of the pre-training signature recognition encoder by adopting a Loss function Loss _ noise of noise regression and a classification Loss function Loss _ positive in a gradient descending manner, and taking the convergence weight obtained after updating as the shared weight of the encoder 1 and the encoder 2.
And adding Gaussian random noise C (C is a noise vector with one dimension the same as the length of the sample sequence) into a data sequence B obtained after the original sample is preprocessed to obtain a comparison data sequence B '═ B + C, and inputting B and B' into two encoders through an input layer 1 and an input layer 2 respectively.
The signature characteristic identification module is established based on the encoder, the signature characteristic identification module comprises an input layer, an encoder 1, an encoder 2 and an output layer, the input layer 1 and the input layer 2 respectively acquire signature handwriting sample data B acquired by the data acquisition module, and comparison data B 'of random noise is added to the sample data, the data sequence is converted into vectors and is sent to the encoder 1 and the encoder 2, the data are encoded through the encoder 1 and the encoder 2, the signature sample data B enters the encoder 1 through the input layer 1, the sample data B is encoded to output a characteristic vector y, the random noise is added to the signature sample data to generate comparison signature data B', the comparison signature data B 'enters the encoder 2 through the input layer 2, and the corresponding characteristic vector y' is encoded and output. Pre-training an encoder through a noise enhancement learning and noise regression learning mode to obtain difference characteristic values and loss characteristic values of the signature sample data and the signature comparison data, and determining the difference and the commonality of the signature sample data and the signature comparison data according to the difference characteristic values and the loss characteristic values, so as to distinguish whether the signature comparison handwriting is the handwriting of a signer.
FIG. 3 illustrates training a handwriting pre-training model using a classification loss function in conjunction with positive sample data.
And the signature handwriting pre-training module trains the signature recognition encoder by combining the classification loss function with the positive sample data to obtain a sample encoder 1.
Fig. 4 shows that the noise regression function is used to train the signature recognition encoder with the negative sample data to obtain the sample encoder 2.
An encoder of a signature handwriting recognition module is trained by adopting positive samples, an input layer 1 of the signature handwriting recognition module receives input of a positive sample set B ═ B1, B2.. Bn, n is the number of the positive sample sets, an input layer 2 receives input of a negative sample set B '═ B' 1, B '2., B'm }, m is the number of the negative sample sets, namely, B and B 'are from two signatures of different signers or different signatures of the same signer, the input layer converts input feature sequences into feature vectors V1 and V2, then vectors converted by the input layer 1 and the input layer 2 are sent into the encoder 1 and the encoder 2, the vectors V1 and V2 are output after encoding processing, output layer output vectors y and y', y are output of the encoder 1, y 'is output of the encoder 2, and the output layer obtains output vectors y and y'. And respectively solving partial derivatives of the updated weight parameters of the encoder 1 and the encoder 2 by adopting a noise regression function and a loss function, reducing the difference characteristic value and the loss characteristic value by using a gradient descent method until a convergence condition is reached, updating the weight parameters of two twin networks sharing the weight in the signature characteristic identification module, and obtaining the weight parameters of the two encoders in the signature characteristic identification module.
Establishing different training tasks according to different values of the noise C, and when C is 0, according to a formula: loss _ positive (1-cos (y, y'))2Calculating a classification loss function, and determining common characteristics of signature comparison data and signature sample data; when C is not zero, the formula is called: loss _ noise ═ 1-cos (y', y) - μ)2A noise regression function is calculated and the difference characteristic of the signature comparison data and the signature sample data is determined, where μ represents a value of the noise C. Respectively updating parameters of the encoder 1 and the encoder 2 by using a classification loss function and a noise regression function, and calling formulas by using the classification loss function
The current weight coefficient w of the encoder 1 is subjected to partial derivation, and a formula is called by using a noise regression function
Obtaining the partial derivatives of the current weight coefficient W of the encoder 2 to obtain the weight parameters of the two sample encoders, and calling the formula W according to the preset learning rate lrnewW-lrW' until convergence. (value W which can be derived from the partial derivation as described above1' and W2' separately substituting the formula to obtain updated weight parameters. Such as calling formula Wnew1＝w-lr×W1' update encoder 1 weight parameters, call formula Wnew2＝w-lr×W2Update encoder 2 weight parameters) update the weight parameters until a convergence weight is obtained. The condition for the loss to reach the predetermined value may be preset as a convergence condition, depending on the specific task. And obtaining shared weight by the two obtained convergence weight parameters through a gradient average or addition synchronization method, wherein the shared weight is used as the weight parameter of the encoders of the two twin network structures in the signature feature recognition model.
Calling formulas using classification loss functions
The current weight coefficient w of the encoder 1 is subjected to partial derivation, and a formula is called by using a noise regression function
Calculating a partial derivative of the current weight coefficient W of the encoder 2, and obtaining convergence through a preset learning rate lr; and obtaining the weight parameters of the two sample encoders, and obtaining the sharing weight of the sharing weight sample encoder in the signature handwriting characteristic identification module by a gradient average or addition synchronization method for the convergence weight parameters of the two encoders.
For the encoder receiving the input of the negative sample set, adopting the training and updating of the encoder weight parameters based on the noise regression Loss function to obtain the difference characteristic values of the signature sample characteristic and the signature comparison characteristic, and obtaining the difference characteristic value according to the noise regression function Loss _ noise ═ 1-cos(y',y)-μ)2And determining the difference characteristic Loss _ noise to obtain the difference between the signature sample characteristic and the signature comparison characteristic. Calling a formula using a noise regression function
Calculating the biased reciprocal of the current weight coefficient W of the encoder, and calling the formula W through the preset learning rate lrnewThe weight parameter is updated to obtain the convergence weight W-lrW'.
Although the difference of data can be learned through noise regression learning, the commonalities of signatures of the same signer cannot be learned, the noise C is set to be zero, and a classification Loss function Loss _ positive ═ 1-cos (y, y')/is adopted for an encoder receiving a positive sample set)2And updating the encoder, obtaining Loss characteristics Loss _ positive of the sample characteristics and the comparison characteristics, and determining the common characteristics of the signature sample data and the signature detection data. Calling formulas using classification loss functions
Calculating the inverse bias of the parameter W, and calling the formula W through a preset learning rate lrnewThe weight parameter is updated to obtain the convergence weight W-lrW'.
Updating the weight parameters of the signature comparison model according to a preset learning rate to obtain a convergence weight, reducing the loss value by adopting a shared gradient (average or addition) reduction method on the weight coefficients obtained by the noise regression loss function and the classification loss function to enable the loss values of the two to be close to 0, updating the weight parameters of the encoders 1 and 2, and obtaining the loss value when the convergence weight is the final difference characteristic value and the final loss characteristic value. Thereby constructing a twin network model containing two same structures and sharing a weight sample encoder as a signature feature identification module.
The method comprises the steps of collecting electronic signature stroke related characteristic data, and acquiring signature sample data and signature to-be-detected data; the method comprises the steps of resampling, normalizing and discretizing signature sample data and data to be detected of a signature to obtain data with fixed length, inputting a signature data sequence to be detected and a sample signature data sequence with fixed length into a sample encoder 1 and a sample encoder 2 from an input layer 1 and an input layer 2 respectively, determining common characteristics and difference characteristics of the signature sample data and the data to be detected of the signature by the two sample encoders respectively through a cosine distance loss function and a noise regression function, and determining whether the signature sample data and the data to be detected of the signature come from signature data of the same signer or signature data of different signers. The invention is widely used in places such as e-commerce, e-government affairs and the like needing signature identification.
The invention also claims a computer-readable storage medium having stored thereon a computer program which can be loaded and executed by a processor to perform the method as described herein above.
Claims (12)
1. An electronic signature handwriting recognition system based on a pre-training technology is characterized by comprising a data acquisition module, a data preprocessing module, a signature handwriting pre-training module and a signature handwriting characteristic recognition module, wherein the data acquisition module acquires electronic signature stroke related characteristic data to acquire signature positive sample data; the data preprocessing module carries out resampling, normalization and discretization on the collected positive sample data to generate negative sample data for electronic signature handwriting pre-training, a neural network is trained in an autoregressive pre-training mode of the positive sample data based on a mask, and a neural network model after the pre-training is completed serves as an encoder; the signature handwriting pre-training module trains an electronic signature handwriting pre-training model by combining a classification loss function with positive sample data and a noise regression function with negative sample data respectively to obtain a sample encoder 1 and a sample encoder 2 which have the same structure; and updating the weight parameters of the two sample encoders until a convergence weight is obtained, and constructing a twin network model of the two sample encoders with the same structure and sharing the weight as a signature handwriting feature recognition module.
2. The system of claim 1, wherein the cosine distance Loss _ compare (cos (y, y')) is used based on the sample encoder 1 output vector y and the sample encoder 2 output vector y2And constructing a classification loss function: loss _ positive (1-cos (y, y'))2And noise regression function:
Loss_noise＝(1-cos(y',y)-μ)2。
3. the system of claim 1, wherein the formula is invoked using a classification loss function
Calculating the partial derivative of the current weight coefficient W of the sample encoder 1, and calling the formula W through the preset learning rate lrnew1＝w-lr×W1' update encoder 1 weight parameters until convergence.
4. The system of claim 1, wherein the formula is invoked using a noise regression function
Calculating the partial derivative of the current weight coefficient W of the sample encoder 2, and calling the formula W through a preset learning rate lrnew2＝w-lr×W2' update encoder 2 weight parameters until convergence.
5. The system according to claim 3 or 4, wherein the shared weight of the shared weight sample encoder in the signature handwriting characteristic identification module is obtained by gradient average or addition synchronization method for the convergence weight parameters of the two sample encoders.
6. The system of one of claims 1 or 5, the training of the neural network in an autoregressive pre-training mode further comprising: and (3) masking partial data sequence by using random mask to the original characteristic sequence of the acquired signature handwriting, inputting the partial data sequence into a neural network model for pre-training, outputting the pre-training characteristic sequence of the signature handwriting by the neural network model, enabling the pre-training characteristic sequence output by the model to approach the original characteristic sequence, and completing the pre-training.
7. A method for recognizing electronic signature handwriting based on a pre-training technology is characterized in that a data acquisition module acquires relevant characteristic data of the electronic signature handwriting, and acquires signature sample data and signature to-be-detected data; the data preprocessing module is used for resampling, normalizing and discretizing the signature sample data and the signature to-be-detected data to obtain signature sample data and signature to-be-detected data with fixed lengths; respectively training an electronic signature handwriting pre-training model by using a classification loss function in combination with positive sample data and using a noise regression function in combination with negative sample data, and constructing a signature identification module comprising a sample encoder 1 and a sample encoder 2 which have the same structure and share a weight; the method comprises the steps that signature sample data and to-be-detected signature data with fixed lengths are input into a signature identification module, a sample encoder 1 and a sample encoder 2 respectively determine the common characteristic and the difference characteristic of the signature sample data and the to-be-detected signature data by utilizing a classification loss function and a noise regression function, and the signature sample data and the to-be-detected signature data are determined to be from the same signer or different signers.
8. The method of claim 7, wherein, based on the sample encoder 1 output vector y and the sample encoder 2 output vector y', based on the classification loss function: loss _ positive (1-cos (y, y'))2Acquiring common characteristics Loss _ positive of the signature sample characteristics and the signature to-be-detected characteristics, and according to a noise regression function: loss _ noise ═ 1-cos (y', y) - μ)2And acquiring the difference characteristic Loss _ noise of the signature sample characteristic and the signature to-be-detected characteristic.
9. The method of claim 7, wherein the formula is invoked using a classification loss function
Calculating the partial derivative of the current weight coefficient W of the encoder 1, and calling a formula W through a preset learning rate lrnew1＝w-lr×W1' update encoder 1 weight parameters until convergence.
10. The method of claim 7, wherein the formula is invoked using a noise regression function
Calculating the partial derivative of the current weight coefficient W of the encoder 2, and calling the formula W according to the preset learning rate lrnew2＝w-lr×W2' update encoder 2 weight parameters until convergence.
11. The method as claimed in claim 9 or 10, wherein the shared weight of the shared weight sample encoder in the signature handwriting feature recognition module is obtained by gradient average or addition synchronization method for the convergence weight parameters of the two encoders.
12. A computer-readable storage medium, having stored thereon a computer program which can be loaded and executed by a processor to perform the method of any one of claims 7 to 11.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111122613.9A CN113779643B (en) | 2021-09-24 | 2021-09-24 | Signature handwriting recognition system and method based on pre-training technology and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111122613.9A CN113779643B (en) | 2021-09-24 | 2021-09-24 | Signature handwriting recognition system and method based on pre-training technology and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113779643A true CN113779643A (en) | 2021-12-10 |
| CN113779643B CN113779643B (en) | 2022-07-01 |
Family
ID=78853250
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111122613.9A Active CN113779643B (en) | 2021-09-24 | 2021-09-24 | Signature handwriting recognition system and method based on pre-training technology and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113779643B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114862648A (en) * | 2022-05-27 | 2022-08-05 | 晋城市大锐金马工程设计咨询有限公司 | Document adopting A, B two documents for cross watermark encryption |
| CN117475519A (en) * | 2023-12-26 | 2024-01-30 | 厦门理工学院 | Off-line handwriting identification method based on integration of twin network and multiple channels |
| CN117476240A (en) * | 2023-12-28 | 2024-01-30 | 中国科学院自动化研究所 | Disease prediction method and device with few samples |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109409254A (en) * | 2018-10-10 | 2019-03-01 | 成都优易数据有限公司 | A kind of electronic contract handwritten signature identification method based on twin neural network |
| CN111079785A (en) * | 2019-11-11 | 2020-04-28 | 深圳云天励飞技术有限公司 | Image identification method and device and terminal equipment |
| US20210141897A1 (en) * | 2019-11-11 | 2021-05-13 | Microsoft Technology Licensing, Llc | Detecting unknown malicious content in computer systems |
| CN113408702A (en) * | 2021-06-23 | 2021-09-17 | 腾讯音乐娱乐科技（深圳）有限公司 | Music neural network model pre-training method, electronic device and storage medium |
-
2021
- 2021-09-24 CN CN202111122613.9A patent/CN113779643B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109409254A (en) * | 2018-10-10 | 2019-03-01 | 成都优易数据有限公司 | A kind of electronic contract handwritten signature identification method based on twin neural network |
| CN111079785A (en) * | 2019-11-11 | 2020-04-28 | 深圳云天励飞技术有限公司 | Image identification method and device and terminal equipment |
| US20210141897A1 (en) * | 2019-11-11 | 2021-05-13 | Microsoft Technology Licensing, Llc | Detecting unknown malicious content in computer systems |
| CN113408702A (en) * | 2021-06-23 | 2021-09-17 | 腾讯音乐娱乐科技（深圳）有限公司 | Music neural network model pre-training method, electronic device and storage medium |
Non-Patent Citations (1)
| Title |
|---|
| 刘力: "《离线手写签名认证算法研究》", 《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114862648A (en) * | 2022-05-27 | 2022-08-05 | 晋城市大锐金马工程设计咨询有限公司 | Document adopting A, B two documents for cross watermark encryption |
| CN117475519A (en) * | 2023-12-26 | 2024-01-30 | 厦门理工学院 | Off-line handwriting identification method based on integration of twin network and multiple channels |
| CN117475519B (en) * | 2023-12-26 | 2024-03-12 | 厦门理工学院 | Off-line handwriting identification method based on integration of twin network and multiple channels |
| CN117476240A (en) * | 2023-12-28 | 2024-01-30 | 中国科学院自动化研究所 | Disease prediction method and device with few samples |
| CN117476240B (en) * | 2023-12-28 | 2024-04-05 | 中国科学院自动化研究所 | Disease prediction method and device with few samples |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113779643B (en) | 2022-07-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113779643B (en) | Signature handwriting recognition system and method based on pre-training technology and storage medium | |
| CN110728209B (en) | Gesture recognition method and device, electronic equipment and storage medium | |
| CN110135249B (en) | Human behavior identification method based on time attention mechanism and LSTM (least Square TM) | |
| CN105550658A (en) | Face comparison method based on high-dimensional LBP (Local Binary Patterns) and convolutional neural network feature fusion | |
| US7068821B2 (en) | Information processing method and apparatus | |
| CN111191564A (en) | Multi-pose face emotion recognition method and system based on multi-angle neural network | |
| Alsaadi | Study on most popular behavioral biometrics, advantages, disadvantages and recent applications: A review | |
| CN102938065A (en) | Facial feature extraction method and face recognition method based on large-scale image data | |
| CN109800643A (en) | A kind of personal identification method of living body faces multi-angle | |
| Qiao et al. | Offline signature verification using online handwriting registration | |
| CN110188669B (en) | Air handwritten character track recovery method based on attention mechanism | |
| CN111626246B (en) | Face alignment method under mask shielding | |
| Hasan et al. | Multi-level feature fusion for robust pose-based gait recognition using RNN | |
| Deore et al. | A survey on offline signature recognition and verification schemes | |
| Li et al. | Adaptive deep feature fusion for continuous authentication with data augmentation | |
| Qin et al. | Finger-vein quality assessment based on deep features from grayscale and binary images | |
| Jiang et al. | PointFace: Point set based feature learning for 3D face recognition | |
| CN114612980A (en) | Deformed face detection based on multi-azimuth fusion attention | |
| Behera et al. | Air signature recognition using deep convolutional neural network-based sequential model | |
| CN115761814A (en) | System for detecting emotion in real time according to human body posture | |
| CN113435398B (en) | Signature feature identification method, system, equipment and storage medium based on mask pre-training model | |
| Li et al. | Memory-augmented autoencoder based continuous authentication on smartphones with conditional transformer gans | |
| CN115393876A (en) | Online signature identification method based on neural network | |
| CN114373091A (en) | Gait recognition method based on deep learning fusion SVM | |
| Wang et al. | Three-dimensional lip motion network for text-independent speaker recognition |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CP03 | Change of name, title or address | ||
| CP03 | Change of name, title or address |
Address after: 401121 no.17-1, building 13, No.106, west section of Jinkai Avenue, Yubei District, Chongqing Patentee after: Chongqing Sign Digital Technology Co.,Ltd. Country or region after: China Address before: 401121 no.17-1, building 13, No.106, west section of Jinkai Avenue, Yubei District, Chongqing Patentee before: CHONGQING AOS ONLINE INFORMATION TECHNOLOGY CO.,LTD. Country or region before: China |