Rhythm feature-based forgery signature detection method and device, and storage medium

Abstract

The application discloses a signature recognition technology based on rhythm characteristics, obtains sample signature discrete track points and verification signature discrete track points, adopts speed logarithmic segmentation to obtain sample signature strokes and corresponding verification signature strokes, calculates rhythm characteristics of the sample signature strokes and the corresponding verification signature strokes, carries out distance measurement according to the rhythm characteristics to obtain total rhythm characteristic distance, and thus determines whether the verification signature is a signature of the signer or a copied signature. The application proposes rhythm characteristics with strokes as basic granularity, does not depend on a large amount of training data, is not limited to any language, character or handwriting form, processes any type of signal wave, satisfies translation stretching, rotation and sampling mode invariance of signature track based on a more scientific and reasonable normalization mode and measurement mode, supports cross-device verification, and can efficiently and accurately distinguish between copying and imitation.

Description

Technical Field

[0001] This invention relates to the field of computer information processing technology, specifically to the identification and detection of counterfeit electronic signatures, solving the problem of difficulty in identifying highly forged signatures in online signature comparison. Background Technology

[0002] In the field of signature authentication, accurately and efficiently identifying highly forged signatures has always been a very difficult problem. Forged signatures are generally indistinguishable from the original sample signature in terms of image, and may even be more similar than the signature's other genuine signatures. For such highly forged signatures, conventional stroke outline features, structural layout features, and handwriting style features are insufficient for effectively distinguishing between positive and negative samples. For online signature data, dynamic features such as signature speed, acceleration, and pressure magnitude are commonly used for differentiation. However, these features have limited discriminative power, and in complex verification scenarios such as mobile devices and cross-device collaborations, the inconsistent sampling frequency, resolution, and pressure sensing methods of devices make it impossible to accurately compare absolute speed, acceleration, and pressure magnitude. Chinese invention patent CN 111310546A, entitled "An Extraction and Authentication Method for Writing Rhythm Features in Online Handwriting Authentication," uses big data statistics on the writing rhythm speed of individual Chinese characters to construct a probability density function of rhythm features at the single-character level. The Chinese invention patent application with publication number CN 111259726 A, entitled "An Extraction and Authentication Method of Stroke Dynamic Features in Online Handwriting Authentication", requires pre-defining the types of various signal waves and classifying and measuring the distribution statistics (such as mean, variance, maximum and minimum) of the original signal waves of the handwriting signature.

[0003] The rhythmic features of the aforementioned existing signature recognition technology are based on single-character granularity, requiring extensive training, pre-definition of various signal wave types, and extensive classification and statistical work. It is difficult to accurately compare and recognize signatures with different text forms and complex strokes. Summary of the Invention

[0004] This invention addresses the aforementioned shortcomings of existing technologies by acquiring signature rhythm features at the stroke level. It does not rely on a large amount of training data, nor is it limited to any language, script, or handwriting form. It can process any type of signal wave and satisfies the invariance of translation, stretching, rotation, and sampling methods of signature trajectories based on a more scientific and reasonable normalization and measurement method. It also supports cross-device verification.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is to propose a method for detecting counterfeit signatures based on rhythm features. This method preprocesses the original signature data to obtain discrete trajectory points of the sample signature and the verification signature, determines the alignment path between the sample signature and the verification signature, uses logarithmic velocity segmentation to obtain the strokes of the sample signature and their corresponding verification signature, calculates the rhythm features of the sample signature and the corresponding verification signature in multiple dimensions using strokes as the basic unit, performs distance measurement on each dimension of the rhythm features to obtain the distance of each dimension of rhythm features, determines the total rhythm feature distance between the sample signature and the verification signature based on one or more of the rhythm feature distances in each dimension, and determines whether the verification signature is the signature of the signatory or a counterfeit signature based on the total rhythm feature distance.

[0006] Further preferably, the total rhythm feature distance includes: one of the following features: Fourier transform feature distance, discrete cosine transform feature distance, paragraph feature distance, time distribution feature distance, and pause feature distance, or a linear weighted average of all or part of the feature distances in each dimension mentioned above, as the total rhythm feature distance; Fourier feature distance is calculated using Euclidean distance, discrete cosine transform feature distance is calculated using Manhattan distance, paragraph feature distance is calculated using information entropy difference, time distribution feature distance is calculated using relative entropy, and pause feature distance is calculated using logarithmic distance.

[0007] Further optimization involves linearly weighting the feature distances across each dimension, according to the formula:

[0008] D(S,T)=λ fft D fft (S,T)(S,T)+λ dct D dct (S,T)+λ seg D seg (S,T)+λ time D time (S,T)+λ hold D hold (S, T)

[0009] The total distance of rhythmic features, D(S, T), is obtained.

[0010] Further preferably, the preprocessing specifically includes: aligning the sampled signature trajectory S and the verified signature trajectory T through dynamic time planning; dividing the sampled signature trajectory S into K strokes according to the logarithm of velocity; and then finding the verified signature segmentation strokes corresponding to the sampled signature segmentation strokes according to the alignment path, wherein the sampled signature trajectory S = {S i |i=0,1,...,K-1}, the i-th stroke of the sample signature S i ={s ij |j = 0, 1, ..., m i-1}, verify the signature trajectory T={T i |i = 0, 1, ..., K-1}, verify the i-th stroke T of the signature. i ={t ij |j = 0, 1, ..., n i -1}, K is the number of segments, m i and n i These represent the number of dots in the i-th stroke of the sample signature and the verification signature, respectively.

[0011] Further optimization involves obtaining the Fourier transform feature distance by performing a Fast Fourier Transform on each stroke segment of the sampled signature trajectory and the verified signature trajectory to obtain Fourier transform features. These features are then standardized to obtain standardized Fourier transform features. Based on the standardized Fourier transform features, the Euclidean distance is used to calculate the Fourier feature distance between corresponding stroke segments of the sampled signature and the verified signature. The weighted average of the feature distances between strokes is taken as the Fourier feature distance D between the signatures. fft (S, T).

[0012] Further optimization involves selecting each stroke S... i The Fourier transform features F are obtained by performing Fast Fourier Transform on each of them. i ,

[0013] S i ={s ij |j = 0, 1, ..., m i -1}→F i ={f ik |k = 0, 1, ..., m i -1},

[0014] The standardized Fourier transform features of the sample signature strokes are as follows: The Fourier transform features of the signature strokes are verified as follows: According to the formula:

[0015] Calculate the Fourier feature distance d between the strokes of the standardized sample signature and the verified signature. fft (S i T i According to the formula: Calculate the Fourier feature distance between the sample signature and the verification signature, where, f ik g ik The Fourier transform features of the i-th stroke k points are used to segment the signature for the retention and verification of the signature trajectory.

[0016] Further optimization was performed on the sample signature stroke segment S. i Perform a cosine transform and standardize to obtain the standardized cosine features. For the corresponding verification signature stroke T i Perform a discrete cosine transform and standardize to obtain the standardized cosine features. Using the Manhattan distance formula: Calculate the discrete cosine feature distance between the sample signature and the verified signature strokes, and then calculate the cosine feature distance d. dct (S i T i Weighted average, according to the formula: The discrete cosine feature distance D between the sampled signature and the verified signature is obtained. dct (S,T).

[0017] Further optimization involves obtaining paragraph features based on the log-normal distribution characteristics of the signature stroke trajectory, and then verifying the signature segmentation stroke T. i Divide the data into segments using the logarithm of the velocity, and calculate the segment Seg(T) for each logarithm of the velocity. i The information entropy of ) is used as the verification signature segmentation stroke T. i Paragraph feature distance d seg (S i T i ), where Seg(T) i )={π ij |j = 0, 1, ..., k i -1}, k i π is the number of segments. ij The percentage of writing time for the j-th segment within the subdivision of the i-th signature stroke.

[0018] Further optimization involves obtaining pause features based on continuous repeating points in the signature stroke segment, and then segmenting the sample signature and verification signature into strokes S. i and T i Consecutive repeating points are merged, and the number of consecutive repeating points for each non-repeating point is recorded, according to the formula: Obtain the peak feature H(S) of Si i ), Get T i The feature of the pause point H(T) i ), call the formula: d hold (S i T i )=H(T i )-H(S i ) Calculate the distance d between the strokes corresponding to the sample signature and the verification signature. hold (S i T i The distance between signatures is calculated by weighted average of the distances between strokes, where l i-s , li-t To retain the signature sample and verify the number of non-repeating dots in the signature stroke segment, δ ij-s δ ij-t The number of repetitions of the j-th non-repeating point of the i-th stroke in the signature for the sample signature and the signature verification is given. α < 1 represents the natural pause rate of the starting point of each stroke.

[0019] The present invention also claims protection for a computer-readable storage medium having a computer program stored thereon, which can be loaded and run by a processor to perform the above-described method for detecting counterfeit signatures based on rhythm features.

[0020] The present invention also claims protection for an electronic device comprising: one or more processors; a memory; and one or more application programs stored in the memory and configured to be loaded and run by the one or more processors to perform the above-described rhythm-based signature detection method.

[0021] This invention utilizes the rhythmic features of online signature handwriting to identify forged signatures. It proposes rhythmic features with strokes as the basic granularity and provides different extraction and measurement methods for each dimension of features. It does not rely on large amounts of training data and is not limited to any particular language or handwriting form; it can handle any type of signal wave. Based on a more scientific and reasonable normalization and measurement method, it satisfies the invariance of signature trajectory translation, stretching, rotation, and sampling methods, and supports cross-device verification. The signature identification system based on these rhythmic features can not only accurately and efficiently identify highly forged signatures but also support handwriting verification in complex scenarios such as cross-device and cross-handwriting types. Therefore, this invention can extract rhythmic features from the dynamic signals of signature handwriting to identify highly forged signatures. Attached Figure Description

[0022] Figure 1 Flowchart of a signature comparison method based on rhythm features according to an embodiment of the present invention;

[0023] Figure 2 Flowchart of rhythm feature extraction according to an embodiment of the present invention;

[0024] Figure 3 A flowchart illustrating the rhythm feature measurement in an embodiment of the present invention. Detailed Implementation

[0025] This invention addresses the task of identifying highly forged signatures in electronic signature systems by proposing a method and measurement mechanism for extracting writing rhythm features in various dimensions of the online signature trajectory. Based on these rhythm features, similarity calculations are performed on the verified signatures for final identification.

[0026] To facilitate a clear understanding of the present invention and to make the technical problems it addresses, its solutions, and advantages more apparent, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. In the following description, specific details such as particular configurations and components are provided merely to aid in a comprehensive understanding of the embodiments of the present invention. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the invention. Furthermore, for clarity and brevity, descriptions of known functions and structures have been omitted. It should be understood that the embodiments are merely illustrative of the invention and not intended to limit its scope.

[0027] like Figure 1 The flowchart shown is a signature comparison method based on rhythm features according to the present invention. The method involves preprocessing the original signature data to obtain discrete trajectory points of the sample signature and the verification signature. The DTW algorithm is used to obtain the alignment path between the sample signature and the verification signature. The sample signature is segmented using velocity logarithmic division to obtain the sample signature strokes. Based on the sample signature strokes and the alignment path, the verification signature strokes corresponding to the sample signature are obtained through the discrete trajectory points of the verification signature. Feature extraction is performed on the sample signature strokes and the corresponding verification signature strokes to obtain the rhythm features of the sample signature and the verification signature. Distance measurement is performed based on the rhythm features to obtain the rhythm feature distance between the signatures. Specifically, the method includes:

[0028] In the preprocessing stage, the sample signature trajectory S and the verification signature trajectory T are first aligned using Dynamic Time Planning (DTW). Then, the sample signature trajectory S is divided into K strokes based on the logarithm of velocity (the segmentation points can be the troughs or peaks of the logarithm of velocity, or other methods can be used to segment the strokes). Finally, the strokes that correspond one-to-one with the verification signature are found based on the DTW alignment path. Wherein, the sample signature trajectory S = {S... i |i=0,1,...,K-1}, the i-th stroke of the sample signature S i ={s ij |j = 0, 1, ..., m i -1}, the signature trajectory to be verified T={T i |i = 0, 1, ..., K-1}, verify the i-th stroke T of the signature. i ={t ij |j = 0, 1, ..., n i -1}, K is the number of segments, where s ij t ij For the purpose of retaining the signature and verifying the position of the j-th stroke of the i-th segment of the signature, m i and n i These represent the number of dots in the i-th stroke of the sample signature and the verification signature, respectively.

[0029] After preprocessing and segmenting the handwriting of the sample signature and the verification signature into strokes, rhythmic features of multiple dimensions are extracted using strokes as the basic unit. The distance between the signatures is then measured based on these rhythmic features. Rhythmic features may include: Fourier transform features, discrete cosine transform features, paragraph features, time distribution features, and pause features. The total rhythmic feature distance between the verification signature and the sample signature is determined based on the rhythmic feature distance, and the authenticity of the verification signature is determined based on this total rhythmic feature distance.

[0030] like Figure 2 The diagram shows the flowchart for rhythm feature extraction. Based on the acquired segmented stroke trajectory, dynamic signal extraction is used to obtain stroke feature signals such as speed, acceleration, and pressure of the segmented strokes. These stroke feature signals are processed to obtain Fourier transform features, discrete cosine transform features, and segment features, respectively. Stroke time distribution statistics are performed on the segmented stroke trajectory to obtain time distribution features, and continuous repetition point frequency statistics are performed on the segmented stroke trajectory to obtain pause features. The rhythm features are obtained by weighting some or all of the features obtained from the segmented stroke trajectory. The following example illustrates the method for obtaining features of each dimension of the segmented stroke trajectory; other methods well-known to those skilled in the art can also be used to obtain relevant features.

[0031] Fourier transform is used to obtain Fourier transform features. The Fourier transform features of each stroke segment are obtained by transforming the coordinates of each stroke trajectory in the sample signature. A Fast Fourier Transform (FFT) is performed on each stroke segment after segmenting the sample signature trajectory to obtain the Fourier transform features. Specifically, the formula is called:

[0032]

[0033] The Fourier transform feature f of the frequency k of the i-th segment of the sampled signature trajectory is obtained. ik Each stroke S i The Fourier transform features F are obtained by performing Fast Fourier Transform on each of them. i ,

[0034] S i ={s ij |j = 0, 1, ..., m i -1}→F i ={f ik |k = 0, 1, ..., m i -1}:

[0035] Among them, s ij The feature of point j on the trajectory of stroke i, s ij =x ij +iy ij It can represent the x and y coordinates of the j-th point of the i-th stroke (x, y, y). ij y ijThe plural form of ) s ij It can also represent the complex form of the pressure, velocity, or acceleration characteristics of the j-th point in the i-th stroke, where j represents the j-th point in the stroke trajectory. The Fourier transform is a transformation from the time domain to the frequency domain, where k represents the frequency.

[0036] The Fourier transform features are standardized, and the first term f i0 The magnitude of the trajectory coordinates is meaningless and can therefore be discarded or directly assigned a value of 0. Discarding the angles in the complex domain of the Fourier transform features and retaining only their magnitude as features ensures rotation invariance. Normalizing the remaining terms using the 1-norm ensures scaling invariance and time sampling invariance. Therefore, the Fourier transform features are standardized, and the standardized Fourier transform features are expressed as:

[0037] in, For each stroke segment of the verified signature, resample it to the same length as the corresponding stroke segment of the sampled signature, and then calculate its Fourier transform feature. For example, if the i-th stroke T of the verified signature... i Resampling to S i The lengths are equal, and then the standardized Fourier transform features are calculated using the same method as the aforementioned sample signatures. Right now

[0038]

[0039] in,

[0040] The Fourier feature distance between the sampled signature and the verified signature is calculated using Euclidean distance based on the standardized Fourier transform features, and then standardized using the L2 norm, specifically according to the formula:

[0041]

[0042] Calculate the Fourier feature distance d between the strokes of the standardized sample signature and the verified signature. fft (S i T i The Fourier feature distance between two signatures can be calculated based on the Fourier feature distance between the strokes. A weighted average of the feature distances between the strokes can be used as the Fourier feature distance D between the signatures. fft (S, T), that is, according to the formula: Calculate the Fourier feature distance between the sampled signature and the verified signature. The above calculation of the Fourier feature distance uses the coordinates of the signature stroke sampling as a feature. Similarly, other time-domain signals such as signature speed, acceleration, and pressure can also be used to calculate the Fourier feature distance between signatures.

[0043] Discrete cosine transform (DCT) is performed on the trajectory features of each segmented signature stroke to obtain DCT features. Specifically, for each segmented signature stroke, the xy components (e.g., the horizontal and vertical coordinates of the segmented stroke trajectory points) are subjected to DCT. i The cosine characteristic (P) of the stroke segment is obtained by performing a cosine transform. i Q i The cosine transform of the coordinates of any coordinate point in the direction trajectory feature point of the stroke segment is used to obtain the cosine transform feature of the k-th point in the i-th stroke segment (p). ik ,q ik ),

[0044]

[0045]

[0046]

[0047] P i ={p ik |k = 0, 1, ..., m i -1}

[0048] Q i ={q ik |k = 0, 1, ..., m i -1}

[0049] Where: (x ij ,y ij ) represents the coordinates of the feature point of the trajectory in the direction of point j in the i-th stroke segment.

[0050] Due to the high-frequency sparsity of DCT, high-frequency terms can be discarded or directly assigned a value of 0. Similar to the standardization method of Fourier transform, the discrete cosine transform features are standardized to obtain the standardized discrete cosine transform features of the signature stroke segments. That The elements in are respectively Specifically, according to the formula:

[0051]

[0052]

[0053] Calculate the standardized discrete cosine transform features of the i-th signature stroke segment. elements in

[0054] For verifying any segment T of the signature i First, resample to the corresponding segment's sample signature S. iThe lengths are equal, and then the discrete cosine transform characteristics are calculated according to the above method and standardized to obtain the standardized discrete cosine transform characteristics. Obtain the discrete cosine transform (DCT) features of all stroke segments, and calculate the DCT feature distance between the sampled signature and the verified signature strokes based on the DCT features. The Manhattan distance can be used to calculate the DCT feature distance between the sampled signature and the verified signature strokes; specifically, the formula is:

[0055]

[0056] Calculate the discrete cosine characteristic distance d dct (S i T i () is used as the distance between strokes.

[0057] The discrete cosine feature distance between signatures is a weighted average of the distances between strokes, i.e., the formula is:

[0058] Calculate the discrete cosine feature distance D between signatures dct (S, T).

[0059] This feature can also be applied to other time-domain signals such as signature speed, acceleration, and pressure.

[0060] Segment features are obtained based on the log-normal distribution of signature stroke trajectory characteristics. The following explanation uses stroke speed features as an example, which are derived from the log-normal distribution of stroke speed. Due to the segmentation method based on the logarithm of speed described above, S... i The logarithm of the velocity follows an approximately normal distribution. If the signature T is verified as a forgery, based on its image features and the i-th segment S... i Alignment verification stroke T i The logarithm of writing speed will produce multiple consecutive normal distributions due to differences in writing proficiency or intonation. In other words, different writing speeds of genuine and counterfeit signatures will produce completely different "paragraphs." Consider the verification stroke T. i Continue segmenting using the logarithm of velocity, assuming we are verifying stroke T. i The result of further segmentation is Seg(T) i )for:

[0061] Seg(T i )={π ij |j = 0, 1, ..., k i -1}

[0062] Where k i Let π be the number of segments in the i-th segment. ij Let j be the percentage of writing time for the j-th segment. According to the formula:

[0063]

[0064] Calculate and verify stroke T i Seg(T) segmentation i The information entropy of the segment can represent the degree of disorder in the segment, serving as a verification parameter for stroke T. i Paragraph features L(T) i Information entropy represents the degree of uncertainty. Therefore, if there is only one definite segment, the information entropy is at its minimum value of 0, since S... i Since it can only be divided into segments, its information entropy is always 0. Therefore, the segment feature distance d between strokes is... seg (S i T i ) is L(T i ):d seg (S i T i )=L(T i )

[0065] The paragraph feature distance is obtained by weighting the distance between strokes. The paragraph feature distance between signatures is the weighted average of the distances between strokes.

[0066]

[0067] This feature can also be applied to other time-domain signals such as signature acceleration and pressure. That is, similar methods can be used to calculate and obtain paragraph features through trajectory feature signals such as signature acceleration and pressure.

[0068] The number of dots (m) in the strokes of the segmented signature based on the sample signature and the verification signature. i n i The time distribution characteristics Seg(S) and Seg(T) of the sample signature and the verification signature are determined by the number of segments K. The difference in the proportion of time spent on each stroke of the sample signature and the verification signature is calculated based on these time distribution characteristics. If it is a forgery, less time will be spent on familiar and simple strokes, and more time will be spent on unfamiliar and complex strokes.

[0069] Using the segmented representation method of the above paragraph features, the time distribution characteristics of the sample signature and the verification signature are represented as follows:

[0070] Seg(S)={p i |i = 0, 1, ..., K-1}

[0071] Seg(T)={q i |i = 0, 1, ..., K-1}

[0072] Among them, (p i q i These are the time distribution feature elements of the sample signature and the verification signature, respectively:

[0073]

[0074] The temporal distribution feature distance D between signatures time (S, T) can be calculated using relative entropy, specifically according to the formula:

[0075]

[0076] Calculate the temporal distribution feature distance between the sample signature and the verification signature.

[0077] Based on the continuous repetition of points in the signature strokes, pause features are obtained, and the distance between the pause features of the sample signature and the verified signature is calculated. Forgers may pause at certain points in the signature due to unfamiliarity or other factors. Therefore, the more pauses and the longer the pauses, the more likely it is a forged signature. First, the sample signature is divided into strokes S. i Merge consecutively repeating points and record the number of consecutive repetitions of each non-repeating point (Unique(S)). i ): Unique(S i )={(∈ ij δ ij |j = 0, 1, ..., l i -1}

[0078] Among them, l i The number of non-repeating points, ∈ ij For the j-th non-repeating point of the i-th stroke, δ ij For ∈ ij The number of repetitions.

[0079] According to the formula:

[0080] Get S i The feature of the pause point H(S) i ), S i The feature of the pause point H(S) i The ) indicates the degree of pause. Similarly, for any segment T of the signature verification... i First, resample to the corresponding segment's sample signature S. i The strokes are of equal length, and then the pause feature H(T) of each stroke is calculated using the method described above. i According to the formula:

[0081] Get T i The feature of the pause point H(T) i ), indicating the degree of pause. Among them, l i-s , l i-t To retain the signature sample and verify the number of non-repeating dots in the signature stroke segment, δ ij-s δij-t For signature verification and signature retention, the number of repetitions of the j-th non-repeating point of the i-th stroke segment is given, where α < 1 represents the natural pause rate at the starting point of each stroke segment, and δ i0 Let be the number of times the first non-repeating point of the i-th stroke is repeated.

[0082] Statistics show that even in genuine signatures, there is a natural pause at the point where each stroke begins, thus reducing the duration of the pause. Based on the pause characteristics of a certain segment of the sample signature and the corresponding pause characteristics of the verified signature, the formula is applied:

[0083] d hold (S i T i )=H(T i )-H(S i )

[0084] Calculate the distance d between the strokes corresponding to the sample signature and the verification signature. hold (S i T i ).

[0085] The distance D between signatures is a feature of the pause feature. hold (S, T) is the weighted average of the distances between strokes:

[0086]

[0087] like Figure 3 The flowchart shown is a rhythm feature measurement flowchart of the present invention. Based on the Fourier transform features, discrete cosine transform features, paragraph features, time distribution features, and pause features obtained above, the corresponding feature distances are calculated to determine the total rhythm feature distance between the verification signature and the sample signature. Specifically, the time distribution feature distance is obtained using relative entropy, the Fourier feature distance is calculated using Euclidean distance, the paragraph feature distance is calculated using information entropy difference, the discrete cosine transform feature distance is calculated using Manhattan distance, and the pause feature distance is calculated using logarithmic distance. A certain feature distance can be directly used as the total rhythm feature distance, or all or part of the above feature distances can be linearly weighted to obtain the total rhythm feature distance between the sample signature and the verification signature.

[0088] For example, the total rhythm feature distance D(S, T) can be obtained by linearly weighting the feature distances of each dimension, specifically as follows:

[0089] D(S,T)=λ fft D fft (S,T)(S,T)+λ dct D dct (S,T)+λ seg D seg (S,T)+λtime D time (S,T)+λ hold D hold (S, T)

[0090] Where, λ fft , λ dct , λ seg , λ time , λ hold These represent the importance of Fourier features, discrete cosine transform features, paragraph features, time distribution features, and pause features, respectively. The importance of each feature can be set based on experiments and experience, satisfying λ. fft +λ dct +λ seg +λ time +λ hold =1. D fft (S,T)(S,T),D dct (S, T), D seg (S, T), D time (S, T), D hold (S, T) represent the Fourier transform feature distance, discrete cosine transform feature distance, paragraph feature distance, time distribution feature distance, and pause feature distance, respectively. The signature verification is quantized based on the total rhythm feature distance. When the total rhythm feature distance exceeds a specified threshold, the verified signature is determined to be a forged signature.

[0091] This invention analyzes the rhythmic changes and potential fluctuations of the dynamic signals of signature handwriting to extract rhythmic features for identifying highly forged signatures. Since it is difficult for forgers to obtain the dynamic trajectory information of the original signature and to imitate its writing rhythm, this invention uses these features as rhythmic characteristics, providing a powerful means of distinguishing between forged and counterfeit signatures.

Claims

1. A method for detecting counterfeit signatures based on rhythmic features, characterized in that, The original signature data is preprocessed to obtain discrete trajectory points of the sampled signature and the verified signature. The sampled signature trajectory is then aligned using dynamic time planning. and verify signature trace The sample signature trajectory is determined based on the logarithm of the velocity. Cut into The strokes are then used to find the verification signature segmentation strokes corresponding to the segmentation strokes of the sample signature based on the alignment path. The sample signature trajectory is also included. The i-th stroke of the signature sample Verify signature trace Verify the i-th stroke of the signature. , For the number of segments, and The signatures for the sample and the signatures for verification are respectively the first and second signatures. The number of strokes in a segment is used as the basic unit to calculate the rhythmic features of the sample signature strokes and the corresponding verification signature strokes in multiple dimensions. Distance measurements are performed on the rhythmic features in each dimension to obtain the rhythmic feature distance for each dimension. The total rhythmic feature distance between the sample signature and the verification signature is determined based on one or more of the rhythmic feature distances in each dimension. The total rhythmic feature distance is used to determine whether the verification signature is the signature of the signatory or a forged signature. The total rhythmic feature distance includes: Fourier transform feature distance, discrete cosine transform feature distance, segment feature distance, time distribution feature distance, and pause feature distance as the total rhythmic feature distance, or a linear weighted average of all or part of the above feature distances as the total rhythmic feature distance. The Fourier transform feature distance is calculated using Euclidean distance, the discrete cosine transform feature distance is calculated using Manhattan distance, the segment feature distance is calculated using information entropy difference, the time distribution feature distance is calculated using relative entropy, and the pause feature distance is calculated using logarithmic distance.

2. The method according to claim 1, characterized in that, Linear weighting is applied to the feature distances of each dimension, according to the formula: Obtain the total distance of rhythmic features ,in, , , , , These represent the importance of Fourier transform features, discrete cosine transform features, paragraph features, time distribution features, and pause features, respectively. These represent the Fourier transform feature distance, discrete cosine transform feature distance, paragraph feature distance, time distribution feature distance, and pause feature distance, respectively.

3. The method according to claim 1 or 2, characterized in that, Obtaining the Fourier transform feature distance involves: performing a Fast Fourier Transform on each stroke segment of the sampled signature trajectory and the verified signature trajectory to obtain Fourier transform features; standardizing these Fourier transform features to obtain standardized Fourier transform features; calculating the Fourier feature distance between corresponding stroke segments of the sampled signature and the verified signature using Euclidean distance based on the standardized Fourier transform features; and finally, using a weighted average of the feature distances between strokes as the Fourier feature distance between signatures. Specifically, each stroke segment is subjected to a Fast Fourier Transform to obtain Fourier Transform features, which are then standardized to obtain the standardized Fourier Transform features of the sampled signature strokes: For each stroke segment of the verified signature, resample to the same length as the corresponding stroke segment of the sampled signature, and determine its standardized Fourier transform features as follows: According to the formula: Calculate the Fourier transform feature distance between signature strokes According to the formula: Calculate the Fourier transform feature distance between the sampled signature and the verified signature, where, , , , The standard Fourier transform feature vector of the k-th frequency of the i-th stroke is used to segment the signature for the sample signature and to verify the signature trajectory.

4. The method according to any one of claims 1-3, characterized in that, For the signature strokes of the sample Perform a cosine transform and standardize to obtain the standardized cosine features. ), for the corresponding verification signature strokes Perform a discrete cosine transform and standardize to obtain the standardized cosine features. The Manhattan distance call formula is used: Calculate the discrete cosine feature distance between the sample signature and the verified signature strokes, and then calculate the discrete cosine feature distance. Weighted average, according to the formula: Obtain the discrete cosine feature distance between the sampled signature and the verified signature. , where K is the number of segments.

5. The method according to any one of claims 1-3, characterized in that, Segment features are obtained based on the log-normal distribution characteristics of signature stroke trajectories to verify signature segmentation. Divide the segments using the logarithm of velocity as follows: The distance of paragraph features between the strokes of the sample signature and the verified signature Information entropy According to the formula: Calculate and verify strokes Segmentation Information entropy According to the formula: The paragraph feature distance is obtained by weighting the distance between strokes. ,in, Let i be the number of segments for the i-th stroke. For the i-th segment of the signature strokes after further segmentation, the first segment is the i-th segment of the signature strokes. The percentage of writing time for each paragraph.

6. The method according to any one of claims 1-3, characterized in that, Based on the continuous repetition of points in the signature stroke segment, the pause feature is obtained, and the sample signature and the verification signature are segmented by stroke. and Consecutive repeating points are merged, and the number of consecutive repeating points for each non-repeating point is recorded, according to the formula: Get pause features , Get pause features Call the formula: Calculate the distance between the strokes of the sample signature and the verification signature. The distance between signatures is calculated by weighted averaging of the distances between strokes, where... , To retain the signature sample and verify the number of non-repeating dots in the signature stroke segment, , For sample signature, verify the signature, the i-th stroke of the signature The number of repetitions, This indicates the natural pause rate at the beginning of each stroke segment. .

7. A computer-readable storage medium, characterized in that, It stores a computer program that can be loaded and run by a processor to perform the rhythm-based signature detection method according to any one of claims 1 to 6.

8. An electronic device, characterized in that, The electronic device includes: one or more processors; a memory; and one or more applications stored in the memory and configured to be loaded and run by the one or more processors to perform the rhythm-based signature detection method according to any one of claims 1 to 6.

Images