Contract seal identification method based on machine learning

Abstract

The invention provides a contract seal recognition method based on machine learning, and relates to the technical field of machine learning, and the method comprises the steps: obtaining a contract seal image in real time, and carrying out the image screening, and obtaining a target seal image; performing analysis and feature extraction on the target seal image to obtain target image features; and on the basis of a first seal identification result obtained by performing pattern identification on the target image features by introducing a machine learning algorithm, verifying and comparing a current contract seal with a reference seal, and judging whether the current contract is legal or not. The method comprises the steps of performing image screening, analysis processing and feature extraction on an image of a contract seal acquired in real time to obtain a target image feature; on the basis of the first seal recognition result obtained by introducing a machine learning algorithm to perform pattern recognition on the target image features, the current contract seal and the reference seal are verified and compared, whether the current contract is legal or not is judged, the authenticity and the type of the contract seal can be quickly and accurately recognized, and then the legality of the current contract is determined.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a contract seal recognition method based on machine learning. Background Art

[0002] In the process of contract signing, seals are important identity verification and document signing tools with irreplaceable legal effect. However, traditional seal recognition methods often rely on manual operations, which are not only time-consuming and laborious, but also easily affected by human factors, resulting in low recognition accuracy. Therefore, how to quickly and accurately identify the authenticity of contract seals and determine the legality of the current contract has become one of the current research topics.

[0003] Therefore, the present invention provides a contract seal recognition method based on machine learning. Summary of the invention

[0004] The present invention provides a contract seal recognition method based on machine learning, which is used to obtain target image features by performing image screening, analysis processing and feature extraction on images of contract seals acquired in real time; based on the first seal recognition result obtained by performing pattern recognition on the target image features by introducing a machine learning algorithm, the current contract seal is verified and compared with a reference seal to determine whether the current contract is legal, so that the authenticity and type of the contract seal can be quickly and accurately identified, and the legality of the current contract can be determined.

[0005] The present invention provides a contract seal recognition method based on machine learning, comprising: Step 1: Use the set acquisition tool to obtain the image of the contract seal of the target contract in real time, and perform image screening to obtain the target seal image; Step 2: Analyze and process the target seal image and extract features to obtain target image features; Step 3: Introduce a machine learning algorithm to perform pattern recognition on the target image features to obtain a first seal recognition result; Step 4: Based on the first seal recognition result, the contract seal of the current target contract is verified and compared with the reference seal to determine whether the current target contract is legal.

[0006] Preferably, the image of the contract seal of the target contract is acquired in real time by using a set acquisition tool and image screening is performed to obtain the target seal image, including: Scanning the contract seal of the current target contract using a set scanner to obtain a first scanned image; Using a set camera to take photos of the contract seal of the current target contract for a preset number of times to obtain a first set of photographed images; Performing clarity analysis on all images in the first captured image set based on a set clarity index, and outputting an image with the highest image clarity as a first camera image; Performing usability analysis on the first scanned image and the first camera image to obtain a usability coefficient; The calculation formula of the available coefficient is as follows: ; In the formula, Represented as the available coefficients of the current image; Represents the i-th available evaluation index of the current image, where i=1, 2, , n; n represents the number of available evaluation indicators; It is expressed as the influence weight of the i-th available evaluation index on the evaluation of image usability; The image with a larger available coefficient between the first scanned image and the first camera image is used as a target seal image.

[0007] Preferably, the clarity index is set to include gradient mean, gradient variance and Laplacian variance.

[0008] Preferably, the target seal image is analyzed and processed and feature extracted to obtain target image features, including: Step 11: performing image preprocessing on the target seal image to obtain a first image; Step 12: Calculate the usable coefficient of the first image. If the usable coefficient of the first image is greater than a set usable threshold, the current first image is output as an applicable image. Step 13: Otherwise, re-photograph the contract seal of the current target contract using the set camera to obtain a second photographed image set; Step 14: selecting an image with the highest definition from the second captured image set and performing preprocessing to obtain a second image; Step 15: Calculate the usable coefficient of the second image. If the usable coefficient of the second image is greater than a set usable threshold, output the current second image as a usable image. If the usable coefficient of the second image is not greater than the set usable threshold, repeating steps 13-15 until a suitable image is output; Step 16: Extract features from the applicable image to obtain target image features.

[0009] Preferably, performing feature extraction on the applicable image to obtain target image features includes: Step 21: inputting the applicable image into a pre-established feature extraction model to obtain a first feature set; Step 22: extracting features from the applicable image using a set feature extraction algorithm to obtain a second feature set; Step 23: performing feature comparison of the same feature category on the first feature set and the second feature set to obtain a first feature comparison result; Step 24: If the first feature comparison result is completely consistent, the feature category corresponding to the current first feature comparison result is marked as a valid feature category; Outputting corresponding feature content of the valid feature category in the first feature set as target image features; If the first feature comparison result is different, marking the feature category corresponding to the current first feature comparison result as a suspicious feature category; Step 25: Compare the corresponding feature content of the suspicious feature category in the first feature set with the corresponding feature content in the second feature set, and obtain the same content parts, and mark them as useful feature content; Step 26: Compare the corresponding feature content of the suspicious feature category in the first feature set with the corresponding feature content in the second feature set to obtain different content parts, and mark them as first undetermined feature content; Comparing the corresponding feature content of the suspicious feature category in the second feature set with the corresponding feature content in the first feature set to obtain different content parts, marking them as second undetermined feature content; Step 27: performing credibility analysis on the first undetermined feature content and the second undetermined feature content to obtain usable feature content; Step 28: Supplement the available feature content to the useful feature content, and then output it as the target image feature; Step 29: If the first feature comparison result is completely different, optimize the current feature extraction model and repeat steps 21-28 until the target image feature output is obtained or the maximum number of iterations is reached; When the maximum number of iterations is reached and the target image features still cannot be obtained, a warning signal is sent to the maintenance personnel for manual processing.

[0010] Preferably, the credibility analysis is performed on the first undetermined feature content and the second undetermined feature content to obtain available feature content, including: Taking the first feature content to be determined and the second feature content to be determined as matching conditions, traversing the set available feature database, and respectively obtaining the available frequencies of the first feature content to be determined and the second feature content to be determined; Extracting feature values ​​of the first undetermined feature content and the second undetermined feature content under different preset association conditions from the preset feature data set, respectively, and obtaining a first feature value set and a second feature value set accordingly; Standardizing the first eigenvalue set and the second eigenvalue set respectively to obtain a first set and a second set; Using the eigenvalues ​​in the first set to sequentially construct a first eigenvalue change curve; constructing a second eigenvalue variation curve in sequence using the eigenvalues ​​in the second set; Extracting curve features of the first characteristic value change curve and performing weighted averaging to obtain a characteristic value change coefficient of the first characteristic value change curve under the current preset association condition; Extracting the curve characteristics of the second eigenvalue change curve and performing weighted averaging to obtain the eigenvalue change coefficient of the second eigenvalue change curve under the current preset association condition; By combining the acquired characteristic value variation coefficient under each type of preset association condition with the available frequency, the credibility index of the current first to-be-determined characteristic content and the second to-be-determined characteristic content is calculated; If the credibility indexes of the first undetermined feature content and the second undetermined feature content are both greater than the set credibility threshold, the current first undetermined feature content and the second undetermined feature content are used as available feature content, and the current first undetermined feature content and the second undetermined feature content and the corresponding current feature values ​​are stored in the set available feature database; If the credibility index of the first pending feature content and the second pending feature content are not both greater than the set credibility threshold, the pending feature content with a larger credibility index among the current first pending feature content and the second pending feature content is used as the available feature content, and is stored in the set available feature database in combination with the corresponding current feature value.

[0011] Preferably, the calculation formula of the credibility index is as follows: ; In the formula, It is expressed as the credibility index of the current pending feature content; 1 indicates the available frequency of the currently undetermined feature content; 2 represents the available frequency of another pending feature content that belongs to the same suspicious feature category as the current pending feature content; It is expressed as the influence weight of available frequency on the reliability of analyzing the current feature content; It is expressed as the characteristic value variation coefficient of the current undetermined characteristic content under the j-th preset association condition, where j=1, 2, , m; m represents the total number of types of preset association conditions; It is expressed as the influence weight of the change degree of the feature value under the j-th preset association condition on the credibility of the analysis of the current feature content; 2 represents the influence weight of feature stability on the credibility of analyzing the current feature content; e represents the natural base.

[0012] Preferably, based on the first seal recognition result, the contract seal of the current target contract is verified and compared with the reference seal to determine whether the current target contract is legal, including: Analyze the first seal recognition result, and if the authenticity recognition result in the first seal recognition result is that the seal is forged, determine that the current target contract is illegal; If the first seal recognition result is that the seal is authentic, obtaining a reference seal from a set seal library according to the seal type extracted from the first seal recognition result; Extract the image features of the reference seal and compare them with the target image features of the same feature category. If there is inconsistency in the feature comparison, the current target contract is judged to be illegal. If there is no inconsistency in feature comparison, the similarity between the seal feature of the reference seal and the target image feature is calculated using a set similarity algorithm to obtain a first similarity coefficient; When the first similarity coefficient is greater than a set similarity threshold, the current target contract is judged to be legal; When the first similarity coefficient is not greater than a set similarity threshold, the current target contract is judged to be illegal.

[0013] Compared with the prior art, the present invention has the following beneficial effects: The target image features are obtained by performing image screening, analysis and feature extraction on the image of the contract seal obtained in real time. The first seal recognition result is obtained by performing pattern recognition on the target image features based on the introduction of a machine learning algorithm, and the current contract seal is verified and compared with the reference seal to determine whether the current contract is legal. This can quickly and accurately identify the authenticity and type of the contract seal, and then determine the legality of the current contract.

[0014] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0015] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 The present invention is a flowchart of a method for recognizing contract seals based on machine learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0018] The embodiment of the present invention provides a contract seal recognition method based on machine learning, such as Figure 1 As shown, including: Step 1: Use the set acquisition tool to obtain the image of the contract seal of the target contract in real time, and perform image screening to obtain the target seal image; Step 2: Analyze and process the target seal image and extract features to obtain target image features; Step 3: Introduce a machine learning algorithm to perform pattern recognition on the target image features to obtain a first seal recognition result; Step 4: Based on the first seal recognition result, the contract seal of the current target contract is verified and compared with the reference seal to determine whether the current target contract is legal.

[0019] In this embodiment, the target contract refers to the contract that currently needs to verify the legality; the set acquisition tool includes a set scanner and a set camera, wherein the set scanner is predetermined; the set camera is a predetermined high-resolution camera; the target seal image refers to an image with high availability screened out from the images acquired by using the set scanner and the set camera; the target image features are features obtained after analyzing and processing the target seal image and extracting features, including color features, shape features, texture features, etc.; the first seal recognition result is composed of a seal type and an authenticity recognition result, wherein the authenticity recognition result includes two results: a forged seal and a genuine seal; the reference seal refers to a pre-stored real seal template.

[0020] In this embodiment, a machine learning algorithm is introduced to perform pattern recognition on the target image features to obtain a first seal recognition result, including: Obtain analysis image data of contract seals of all seal types and genuine and fake contract seals under different preset association conditions; After the acquired analysis image data is marked with seal type and authenticity, the analysis image data is divided according to a set data ratio to obtain training image data and test image data; Extracting image features from the training image data to obtain training features and outputting them as training data; In combination with a random forest algorithm, the training data is used to train a model to obtain a first recognition model; Using the test image data to input into the first recognition model, to obtain a test result; By analyzing the test results, obtaining a performance evaluation coefficient of the current first recognition model; When the performance evaluation coefficient is greater than the set evaluation threshold, the current first recognition model is used as the pattern recognition model; When the performance evaluation coefficient is not greater than the set evaluation threshold, the key parameters in the current first recognition model are optimized to generate a pattern recognition model; The target image feature is input into the pattern recognition model to obtain a pattern recognition result and output it as a first seal recognition result.

[0021] Among them, the preset association conditions refer to size, color, background and lighting; the seal type refers to the category of contract seals, such as official seals and financial seals; the analysis image data refers to the contract seal images of different seal types or authentic contract seal images collected under different association conditions; the set data ratio is pre-set, generally 4:1, that is, the training image data accounts for 80% of all the analysis image data, and the test image data accounts for 20% of all the analysis image data; The training features refer to the features extracted from the training image data using the pre-trained feature extraction layer of the convolutional neural network; the first recognition model is obtained by combining the random forest algorithm and using the training data to train the model, wherein the training data refers to the training features; the test result refers to the recognition result output after the test image data is input into the first recognition model, including the seal type and the authenticity of the seal; the performance evaluation coefficient is used to evaluate the model performance of the current first recognition model, and refers to the accuracy obtained by comparing the seal type recognition result and the seal authenticity recognition result of the test image data in the test result with the real seal type and the real seal authenticity of the test image data; the evaluation threshold is set in advance, generally 0.8; the key parameters refer to the depth of the tree and the splitting criterion; the first seal recognition result is the pattern recognition result obtained by inputting the target image features into the pattern recognition model, which is composed of the seal type and authenticity recognition results.

[0022] The beneficial effects of the above technical solution are: by performing image screening, analysis processing and feature extraction on the image of the contract seal obtained in real time, the target image features are obtained; based on the first seal recognition result obtained by performing pattern recognition on the target image features through the introduction of a machine learning algorithm, the current contract seal is verified and compared with the reference seal to determine whether the current contract is legal, thereby realizing rapid and accurate identification of the authenticity and type of the contract seal, and further determining the legality of the current contract.

[0023] The embodiment of the present invention provides a contract seal recognition method based on machine learning, which uses a set acquisition tool to obtain the image of the contract seal of the target contract in real time and performs image screening to obtain the target seal image, including: Scanning the contract seal of the current target contract using a set scanner to obtain a first scanned image; Using a set camera to take photos of the contract seal of the current target contract for a preset number of times to obtain a first set of photographed images; Performing clarity analysis on all images in the first captured image set based on a set clarity index, and outputting an image with the highest image clarity as a first camera image; Performing usability analysis on the first scanned image and the first camera image to obtain a usability coefficient; The calculation formula of the available coefficient is as follows: ; In the formula, Represented as the available coefficients of the current image; Represents the i-th available evaluation index of the current image, where i=1, 2, , n; n represents the number of available evaluation indicators; It is expressed as the influence weight of the i-th available evaluation index on the evaluation of image usability; The image with a larger available coefficient between the first scanned image and the first camera image is used as a target seal image.

[0024] In this embodiment, the setting of the scanner is pre-set; the setting of the camera is pre-set; the target contract refers to the contract that currently needs to verify the legality; the first scanned image is an image obtained by scanning the contract seal of the current target contract using the set scanner; the preset number of times is pre-set, generally 8 times; the first captured image set is an image obtained by using the set camera to take pictures of the contract seal of the current target contract for a preset number of times; the image clarity is obtained by obtaining the set clarity index of the current image and performing weighted averaging, wherein the set clarity index includes the gradient mean, the gradient variance and the Laplacian variance, which are calculated by using the Sobel operator to calculate the gradient of the image in the horizontal and vertical directions and then calculate the gradient amplitude of each pixel point, and then determine the index according to the gradient amplitude of the entire image; the weight assigned to the set clarity index is obtained by solving the matrix constructed by comparing the set clarity indexes pairwise and scoring the relative importance by using the hierarchical analysis method; the available coefficient is used to indicate the usability of the current image; the target seal image refers to the image with a larger available coefficient between the first scanned image and the first camera image, providing data basis for subsequent seal recognition.

[0025] The beneficial effect of the above technical solution is that by combining the scanner and the camera to collect images and perform clarity analysis and usability analysis, the impact of errors or failures that may occur in a single device on subsequent seal recognition can be reduced.

[0026] The embodiment of the present invention provides a contract seal recognition method based on machine learning, which performs analysis and feature extraction on the target seal image to obtain target image features, including: Step 11: performing image preprocessing on the target seal image to obtain a first image; Step 12: Calculate the usable coefficient of the first image. If the usable coefficient of the first image is greater than a set usable threshold, the current first image is output as an applicable image. Step 13: Otherwise, re-photograph the contract seal of the current target contract using the set camera to obtain a second photographed image set; Step 14: selecting an image with the highest definition from the second captured image set and performing preprocessing to obtain a second image; Step 15: Calculate the usable coefficient of the second image. If the usable coefficient of the second image is greater than a set usable threshold, output the current second image as a usable image. If the usable coefficient of the second image is not greater than the set usable threshold, repeating steps 13-15 until a suitable image is output; Step 16: Extract features from the applicable image to obtain target image features.

[0027] In this embodiment, image preprocessing includes denoising, enhancement and binarization, the purpose of which is to improve the quality of the current image; the first image is obtained after image preprocessing of the target seal image; the set usable threshold is predetermined, generally 0.8; the second captured image set refers to an image set obtained by re-shooting the contract seal of the current target contract a preset number of times using a set camera when the usable coefficient of the first image is not greater than the set usable threshold; the second image refers to the image with the highest clarity in the second captured image set, wherein the clarity is obtained by obtaining the set clarity index of the current image and performing weighted averaging, wherein the set clarity index includes the gradient mean , gradient variance and Laplacian variance are indicators determined by calculating the gradient of the image in the horizontal and vertical directions and then calculating the gradient amplitude of each pixel, and then determining it according to the gradient amplitude of the entire image; the weight assigned to the set clarity index is obtained by solving the matrix constructed after pairwise comparison and relative importance scoring of the clarity indexes using the hierarchical analysis method; the applicable image refers to an image with an available coefficient greater than the set available threshold; the feature extraction model is pre-established for extracting features of the current applicable image, including color, texture, etc.; the target image features are obtained by extracting features of the applicable image using the feature extraction model.

[0028] The beneficial effect of the above technical solution is: by performing image preprocessing and image quality analysis on the current target seal image, and automatically triggering the reshooting and screening process when the image quality does not meet the requirements, feature extraction is performed until a suitable image that meets the requirements is found, thereby effectively ensuring image quality and improving the accuracy of feature extraction.

[0029] The embodiment of the present invention provides a contract seal recognition method based on machine learning, which extracts features of the applicable image to obtain target image features, including: Step 21: inputting the applicable image into a pre-established feature extraction model to obtain a first feature set; Step 22: extracting features from the applicable image using a set feature extraction algorithm to obtain a second feature set; Step 23: performing feature comparison of the same feature category on the first feature set and the second feature set to obtain a first feature comparison result; Step 24: If the first feature comparison result is completely consistent, the feature category corresponding to the current first feature comparison result is marked as a valid feature category; Outputting corresponding feature content of the valid feature category in the first feature set as target image features; If the first feature comparison result is different, marking the feature category corresponding to the current first feature comparison result as a suspicious feature category; Step 25: Compare the corresponding feature content of the suspicious feature category in the first feature set with the corresponding feature content in the second feature set, and obtain the same content parts, and mark them as useful feature content; Step 26: Compare the corresponding feature content of the suspicious feature category in the first feature set with the corresponding feature content in the second feature set to obtain different content parts, and mark them as first undetermined feature content; Comparing the corresponding feature content of the suspicious feature category in the second feature set with the corresponding feature content in the first feature set to obtain different content parts, marking them as second undetermined feature content; Step 27: performing credibility analysis on the first undetermined feature content and the second undetermined feature content to obtain usable feature content; Step 28: Supplement the available feature content to the useful feature content, and then output it as the target image feature; Step 29: If the first feature comparison result is completely different, optimize the current feature extraction model and repeat steps 21-28 until the target image feature output is obtained or the maximum number of iterations is reached; When the maximum number of iterations is reached and the target image features still cannot be obtained, a warning signal is sent to the maintenance personnel for manual processing.

[0030] In this embodiment, the feature extraction model is a model obtained by training a neural network using a pre-established training data set, wherein the training data set is composed of pre-processed image data containing various image features and defined feature types, and the feature types include shape, texture, color, etc.; the first feature set is obtained by collecting the features obtained by extracting features from the current applicable image using the feature extraction model; the set feature extraction algorithm is predetermined, including a color segmentation algorithm, a contour extraction and an edge detection algorithm, etc.; the second feature set is obtained by collecting the features obtained by extracting features from the current applicable image using the set feature extraction algorithm; the feature category refers to the image feature type, including shape, color, texture, etc.; the first feature comparison result refers to the comparison result of corresponding features of the same feature category of the first feature set and the second feature set, including three comparison results: completely consistent, different, and completely different.

[0031] In this embodiment, a valid feature category refers to a feature category whose corresponding first feature comparison result is completely consistent; a suspicious feature category refers to a feature category whose corresponding first feature comparison result is different; a first pending feature content refers to a different part of the content obtained by comparing the corresponding feature content of the suspicious feature category in the first feature set with the corresponding feature content in the second feature set; a second pending feature content refers to a different part of the content obtained by comparing the corresponding feature content of the suspicious feature category in the second feature set with the corresponding feature content in the first feature set; a useful feature content refers to the same part of the content obtained by comparing the corresponding feature content of the suspicious feature category in the first feature set with the corresponding feature content in the second feature set; the available feature content is obtained by performing a credibility analysis on the first pending feature content and the second pending feature content; the maximum number of iterations is predetermined; it is used by maintenance personnel to maintain the contract seal recognition process; and the warning signal is used to remind maintenance personnel that there is a problem with the current image feature extraction and requires manual processing.

[0032] The beneficial effects of the above technical solution are: by using a pre-established feature extraction model and a set feature extraction algorithm to perform dual feature extraction, refine suspicious features, iteratively optimize the model, and timely manual intervention, the accuracy of feature extraction and the reliability of the system are significantly improved, which in turn helps to improve the efficiency and automation of image processing.

[0033] The embodiment of the present invention provides a contract seal recognition method based on machine learning, which performs credibility analysis on the first undetermined feature content and the second undetermined feature content to obtain usable feature content, including: Taking the first feature content to be determined and the second feature content to be determined as matching conditions, traversing the set available feature database, and respectively obtaining the available frequencies of the first feature content to be determined and the second feature content to be determined; Extracting feature values ​​of the first undetermined feature content and the second undetermined feature content under different preset association conditions from the preset feature data set, respectively, and obtaining a first feature value set and a second feature value set accordingly; Standardizing the first eigenvalue set and the second eigenvalue set respectively to obtain a first set and a second set; Using the eigenvalues ​​in the first set to sequentially construct a first eigenvalue change curve; constructing a second eigenvalue variation curve in sequence using the eigenvalues ​​in the second set; Extracting curve features of the first characteristic value change curve and performing weighted averaging to obtain a characteristic value change coefficient of the first characteristic value change curve under the current preset association condition; Extracting the curve characteristics of the second eigenvalue change curve and performing weighted averaging to obtain the eigenvalue change coefficient of the second eigenvalue change curve under the current preset association condition; By combining the acquired characteristic value variation coefficient under each type of preset association condition with the available frequency, the credibility index of the current first to-be-determined characteristic content and the second to-be-determined characteristic content is calculated; If the credibility indexes of the first undetermined feature content and the second undetermined feature content are both greater than the set credibility threshold, the current first undetermined feature content and the second undetermined feature content are used as available feature content, and the current first undetermined feature content and the second undetermined feature content and the corresponding current feature values ​​are stored in the set available feature database; If the credibility index of the first pending feature content and the second pending feature content are not both greater than the set credibility threshold, the pending feature content with a larger credibility index among the current first pending feature content and the second pending feature content is used as the available feature content, and is stored in the set available feature database in combination with the corresponding current feature value.

[0034] In this embodiment, the available feature database is set to be a pre-established database for storing feature content with actual value and significance and the corresponding historical usage times; the available frequency refers to the corresponding historical usage times of the first pending feature content or the second pending feature content in the set available feature database within a unit time. When the first pending feature content or the second pending feature content cannot be successfully matched in the set available feature database, the available frequency is 0.

[0035] In this embodiment, the preset feature data set is pre-composed of various feature contents and corresponding feature values ​​obtained under different preset associated conditions, wherein the preset associated conditions include size, color, background and lighting; the purpose of standardizing the first feature value set and the second feature value set respectively is to eliminate the dimension; the first set is a set obtained by standardizing all the feature values ​​in the first feature value set; the second set is a set obtained by standardizing all the feature values ​​in the second feature value set; the first feature value change curve is a curve constructed in sequence using the feature values ​​in the first set; the second feature value change curve is a curve constructed in sequence using the feature values ​​in the second set; the setting of the credible threshold is preset, generally 0.8.

[0036] In this embodiment, the curve features include the slope of the curve, the acceleration of the curve, the curvature, and the sum of the deviations of the extreme points from the average value; the weights assigned to the curve features are obtained by solving a matrix constructed after pairwise comparison and relative importance scoring of the curve features using the hierarchical analysis method; the eigenvalue variation coefficient is used to express the degree of change of the eigenvalues ​​of the first to-be-determined feature content or the second to-be-determined feature content under preset association conditions, that is, the feature stability, which is obtained by weighted averaging the curve features extracted from the corresponding eigenvalue variation curves of the first to-be-determined feature content or the second to-be-determined feature content; the credibility index is used to express the credibility of the current to-be-determined feature content.

[0037] In this embodiment, for example, the curve features of the first characteristic value change curve of the first undetermined characteristic content L1 are the curve slope s1, the curve acceleration s2, the sum of the deviations of the extreme points from the average value s3, and the curvature s4, and the corresponding weights are 0.3, 0.21, 0.32, and 0.17 respectively; At this time, the corresponding characteristic value change coefficient of the current first undetermined characteristic content L1 is .

[0038] The beneficial effects of the above technical solution are: through the comprehensive use of feature matching, feature value extraction and standardization, feature value change curve construction and credibility index calculation, scientific screening and evaluation of feature content are achieved, providing strong support for improving the accuracy and efficiency of data analysis.

[0039] The embodiment of the present invention provides a contract seal recognition method based on machine learning, and the calculation formula of the trust index is as follows: ; In the formula, It is expressed as the credibility index of the current pending feature content; 1 indicates the available frequency of the currently undetermined feature content; 2 represents the available frequency of another pending feature content that belongs to the same suspicious feature category as the current pending feature content; It is expressed as the influence weight of available frequency on the reliability of analyzing the current feature content; It is expressed as the characteristic value variation coefficient of the current undetermined characteristic content under the j-th preset association condition, where j=1, 2, , m; m represents the total number of types of preset association conditions; It is expressed as the influence weight of the change degree of the feature value under the j-th preset association condition on the credibility of the analysis of the current feature content; 2 represents the influence weight of feature stability on the credibility of analyzing the current feature content; e represents the natural base.

[0040] In this embodiment, the weights assigned to the available frequency and feature stability are obtained by solving the matrix constructed after pairwise comparison and relative importance scoring using the hierarchical analysis method; the influence weights assigned to the degree of change of the eigenvalues ​​under preset correlation conditions are obtained by solving the matrix constructed after pairwise comparison and relative importance scoring using the hierarchical analysis method.

[0041] The beneficial effect of the above technical solution is: by calculating the credibility index, it provides an effective data basis for selecting the feature content to be determined, thereby realizing the scientific screening and evaluation of the feature content, and providing strong support for improving the accuracy and efficiency of data analysis.

[0042] The embodiment of the present invention provides a contract seal recognition method based on machine learning, which verifies and compares the contract seal of the current target contract with the reference seal based on the first seal recognition result to determine whether the current target contract is legal, including: Analyze the first seal recognition result, and if the authenticity recognition result in the first seal recognition result is that the seal is forged, determine that the current target contract is illegal; If the first seal recognition result is that the seal is authentic, obtaining a reference seal from a set seal library according to the seal type extracted from the first seal recognition result; Extract the image features of the reference seal and compare them with the target image features of the same feature category. If there is inconsistency in the feature comparison, the current target contract is judged to be illegal. If there is no inconsistency in feature comparison, the similarity between the seal feature of the reference seal and the target image feature is calculated using a set similarity algorithm to obtain a first similarity coefficient; When the first similarity coefficient is greater than a set similarity threshold, the current target contract is judged to be legal; When the first similarity coefficient is not greater than a set similarity threshold, the current target contract is judged to be illegal.

[0043] In this embodiment, the first seal recognition result is composed of the seal type and the authenticity recognition result, wherein the authenticity recognition result includes two results: seal forgery and seal authenticity; the set seal library is composed of corresponding reference seals of different seal types; the reference seal refers to the real seal template pre-stored in the set seal library; the image feature of the reference seal is obtained by pre-feature extraction, including features such as texture, shape and color; the feature category refers to the type of image feature, such as shape and color; the target image feature refers to the feature extracted from the applicable image using a pre-established feature extraction model; the set similarity measurement algorithm refers to the cosine similarity algorithm; the first similarity coefficient refers to the sum of the feature similarities between the image feature of the reference seal and the target image feature calculated using the cosine similarity algorithm; the set similarity threshold is pre-set, and is generally taken as t1, t1 represents the total number of image features; the target contract refers to the contract that currently needs to be verified for legality.

[0044] In this embodiment, for example, there are seal features a1, a2, a3 of the reference seal A1 and target image features z1, z2, z3, among which a1 is consistent with z1 corresponding to the same feature category, and a2 is inconsistent with z2 corresponding to the same feature category. In this case, the current contract is determined to be illegal.

[0045] In this embodiment, for example, the first similarity coefficient between the image feature of the reference seal A2 and the current target image feature is 7.2, the total number of image features t1 is 10, and the currently set similarity threshold is 8; The first similarity coefficient 7.2 between the image feature of the reference seal A2 and the current target image feature is less than 8, and the current contract is determined to be illegal.

[0046] The beneficial effects of the above technical solution are: through automated seal recognition and obtaining reference seals from the set seal library for feature comparison according to the seal type to ensure the consistency of the seal, a more objective and quantitative basis is provided for the judgment of the legality of the contract, thereby achieving rapid judgment of the authenticity and legality of the contract seal, shortening the contract review time and improving work efficiency.

[0047] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A contract seal recognition method based on machine learning, characterized in that: include: Step 1: Use the set acquisition tool to obtain the image of the contract seal of the target contract in real time, and perform image screening to obtain the target seal image; Step 2: Analyze and process the target seal image and extract features to obtain target image features; Step 3: Introduce a machine learning algorithm to perform pattern recognition on the target image features to obtain a first seal recognition result; Step 4: Based on the first seal recognition result, the contract seal of the current target contract is verified and compared with the reference seal to determine whether the current target contract is legal.

2. According to the machine learning-based contract seal recognition method of claim 1, it is characterized in that: The set acquisition tool is used to obtain the image of the contract seal of the target contract in real time and perform image screening to obtain the target seal image, including: Scanning the contract seal of the current target contract using a set scanner to obtain a first scanned image; Using a set camera to take photos of the contract seal of the current target contract for a preset number of times to obtain a first set of photographed images; Performing clarity analysis on all images in the first captured image set based on a set clarity index, and outputting an image with the highest image clarity as a first camera image; Performing usability analysis on the first scanned image and the first camera image to obtain a usability coefficient; The calculation formula of the available coefficient is as follows: ; In the formula, Represented as the available coefficients of the current image; Represents the i-th available evaluation index of the current image, where i=1, 2, , n; n represents the number of available evaluation indicators; It is expressed as the influence weight of the i-th available evaluation index on the evaluation of image usability; The image with a larger available coefficient between the first scanned image and the first camera image is used as a target seal image.

3. A contract seal recognition method based on machine learning according to claim 2, characterized in that: The clarity indicators are set to include gradient mean, gradient variance and Laplacian variance.

4. The method for contract seal recognition based on machine learning according to claim 1, characterized in that: Analyzing and processing the target seal image and extracting features to obtain target image features include: Step 11: performing image preprocessing on the target seal image to obtain a first image; Step 12: Calculate the usable coefficient of the first image. If the usable coefficient of the first image is greater than a set usable threshold, the current first image is output as an applicable image. Step 13: Otherwise, re-photograph the contract seal of the current target contract using the set camera to obtain a second photographed image set; Step 14: selecting an image with the highest definition from the second captured image set and performing preprocessing to obtain a second image; Step 15: Calculate the usable coefficient of the second image. If the usable coefficient of the second image is greater than a set usable threshold, output the current second image as a usable image. If the usable coefficient of the second image is not greater than the set usable threshold, repeating steps 13-15 until a suitable image is output; Step 16: Extract features from the applicable image to obtain target image features.

5. The method for contract seal recognition based on machine learning according to claim 4 is characterized in that: Extracting features from the applicable image to obtain target image features includes: Step 21: inputting the applicable image into a pre-established feature extraction model to obtain a first feature set; Step 22: extracting features from the applicable image using a set feature extraction algorithm to obtain a second feature set; Step 23: performing feature comparison of the same feature category on the first feature set and the second feature set to obtain a first feature comparison result; Step 24: If the first feature comparison result is completely consistent, the feature category corresponding to the current first feature comparison result is marked as a valid feature category; Outputting corresponding feature content of the valid feature category in the first feature set as target image features; If the first feature comparison result is different, marking the feature category corresponding to the current first feature comparison result as a suspicious feature category; Step 25: Compare the corresponding feature content of the suspicious feature category in the first feature set with the corresponding feature content in the second feature set, and obtain the same content parts, and mark them as useful feature content; Step 26: Compare the corresponding feature content of the suspicious feature category in the first feature set with the corresponding feature content in the second feature set to obtain different content parts, and mark them as first undetermined feature content; Comparing the corresponding feature content of the suspicious feature category in the second feature set with the corresponding feature content in the first feature set to obtain different content parts, marking them as second undetermined feature content; Step 27: performing credibility analysis on the first undetermined feature content and the second undetermined feature content to obtain usable feature content; Step 28: Supplement the available feature content to the useful feature content, and then output it as the target image feature; Step 29: If the first feature comparison result is completely different, optimize the current feature extraction model and repeat steps 21-28 until the target image feature output is obtained or the maximum number of iterations is reached; When the maximum number of iterations is reached and the target image features still cannot be obtained, a warning signal is sent to the maintenance personnel for manual processing.

6. A contract seal recognition method based on machine learning according to claim 5, characterized in that: The credibility analysis is performed on the first undetermined feature content and the second undetermined feature content to obtain usable feature content, including: Taking the first feature content to be determined and the second feature content to be determined as matching conditions, traversing the set available feature database, and respectively obtaining the available frequencies of the first feature content to be determined and the second feature content to be determined; Extracting feature values ​​of the first undetermined feature content and the second undetermined feature content under different preset association conditions from the preset feature data set, respectively, and obtaining a first feature value set and a second feature value set accordingly; Standardizing the first eigenvalue set and the second eigenvalue set respectively to obtain a first set and a second set; Using the eigenvalues ​​in the first set to sequentially construct a first eigenvalue change curve; constructing a second eigenvalue variation curve in sequence using the eigenvalues ​​in the second set; Extracting curve features of the first characteristic value change curve and performing weighted averaging to obtain a characteristic value change coefficient of the first characteristic value change curve under the current preset association condition; Extracting the curve characteristics of the second eigenvalue change curve and performing weighted averaging to obtain the eigenvalue change coefficient of the second eigenvalue change curve under the current preset association condition; By combining the acquired characteristic value variation coefficient under each type of preset association condition with the available frequency, the credibility index of the current first to-be-determined characteristic content and the second to-be-determined characteristic content is calculated; If the credibility indexes of the first undetermined feature content and the second undetermined feature content are both greater than the set credibility threshold, the current first undetermined feature content and the second undetermined feature content are used as available feature content, and the current first undetermined feature content and the second undetermined feature content and the corresponding current feature values ​​are stored in the set available feature database; If the credibility index of the first pending feature content and the second pending feature content are not both greater than the set credibility threshold, the pending feature content with a larger credibility index among the current first pending feature content and the second pending feature content is used as the available feature content, and is stored in the set available feature database in combination with the corresponding current feature value.

7. The method for contract seal recognition based on machine learning according to claim 6 is characterized in that: The calculation formula of the credibility index is as follows: ; In the formula, It is expressed as the credibility index of the current pending feature content; 1 indicates the available frequency of the currently undetermined feature content; 2 represents the available frequency of another pending feature content that belongs to the same suspicious feature category as the current pending feature content; It is expressed as the influence weight of available frequency on the reliability of analyzing the current feature content; It is expressed as the characteristic value variation coefficient of the current undetermined characteristic content under the j-th preset association condition, where j=1, 2, , m; m represents the total number of types of preset association conditions; It is expressed as the influence weight of the change degree of the feature value under the j-th preset association condition on the credibility of the analysis of the current feature content; 2 represents the influence weight of feature stability on the credibility of analyzing the current feature content; e represents the natural base.

8. The method for contract seal recognition based on machine learning according to claim 1, characterized in that: Based on the first seal recognition result, the contract seal of the current target contract is verified and compared with the reference seal to determine whether the current target contract is legal, including: Analyze the first seal recognition result, and if the authenticity recognition result in the first seal recognition result is that the seal is forged, determine that the current target contract is illegal; If the first seal recognition result is that the seal is authentic, obtaining a reference seal from a set seal library according to the seal type extracted from the first seal recognition result; Extract the image features of the reference seal and compare them with the target image features of the same feature category. If there is inconsistency in the feature comparison, the current target contract is judged to be illegal. If there is no inconsistency in feature comparison, the similarity between the seal feature of the reference seal and the target image feature is calculated using a set similarity algorithm to obtain a first similarity coefficient; When the first similarity coefficient is greater than a set similarity threshold, the current target contract is judged to be legal; When the first similarity coefficient is not greater than a set similarity threshold, the current target contract is judged to be illegal.

Images