Law enforcement quality supervision system based on computer vision and semantic analysis

Abstract

The present invention relates to the technical field of law enforcement supervision, specifically to a law enforcement quality supervision system based on computer vision and semantic analysis, and the law enforcement quality supervision system based on computer vision and semantic analysis includes an image capture module, a feature extraction module, a semantic analysis module and a logic verification module. In the present invention, by integrating computer vision and semantic analysis technology, accurate recognition and analysis of handprints, seals and signatures in law enforcement documents are achieved, and the accuracy of document verification is significantly improved. Boundary detection and pixel-level optimization ensure the fine capture of image details, so that the system can recognize subtle image changes with higher resolution, improve the system's response speed and sensitivity to complex details, and deep learning technology reduces misjudgment and improves the reliability of authenticity determination in the process of image feature recognition and encoding. Through automated semantic analysis and logic verification, the compliance of law enforcement texts can be comprehensively reviewed, and the transparency and fairness of law enforcement can be strengthened.

Description

Technical Field

[0001] The present invention relates to the technical field of law enforcement supervision, and in particular to a law enforcement quality supervision system based on computer vision and semantic analysis. Background Art

[0002] The field of law enforcement supervision technology involves the use of technical tools and methods to ensure the compliance, transparency and efficiency of law enforcement. Key technologies include data analysis, video surveillance, real-time behavior monitoring and post-event analysis. The purpose is to automate the supervision of law enforcement processes through technical means such as artificial intelligence, machine learning, computer vision, etc., reduce human errors and biases, and improve the fairness of law enforcement. Law enforcement supervision technology is also used in public security, prison management, border control and other government regulatory departments to improve response speed and decision-making quality.

[0003] The main purpose of the law enforcement quality supervision system is to automatically monitor and evaluate the behavior in the law enforcement process to ensure that the behavior of law enforcement personnel complies with legal and ethical standards. The system analyzes the behavior patterns of law enforcement personnel and detects possible violations or abnormal behaviors, thereby providing real-time feedback and follow-up training suggestions. It usually integrates video analysis, voice recognition and other sensor data, and uses deep learning models to perform semantic analysis of law enforcement behavior to enhance law enforcement transparency and public trust.

[0004] Existing technologies usually rely on basic data analysis and video surveillance in law enforcement quality monitoring, and lack advanced recognition capabilities for handprints, seals, and signatures in legal documents, which may lead to high misjudgment rates when processing these elements, especially in the case of poor image quality. In addition, existing methods are not fast enough in big data environments to effectively meet the analysis needs of large amounts of real-time data, which may delay critical law enforcement decisions in emergency situations. The semantic analysis of existing technologies is usually insufficient to handle the matching of complex legal terms and conditions, resulting in the inability to fully and automatically assess the compliance of text content. These limitations have led to reduced efficiency and transparency in the law enforcement process, especially in areas such as public security and border control that require a rapid response. Inefficiency and processing delays may lead to major security risks and management errors. Summary of the invention

[0005] The purpose of the present invention is to solve the shortcomings existing in the prior art and to propose a law enforcement quality supervision system based on computer vision and semantic analysis.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme: the law enforcement quality supervision system based on computer vision and semantic analysis includes:

[0007] The image capture module captures images based on fingerprints, seals and signatures in law enforcement documents, adjusts image brightness and contrast to enhance image quality, applies edge detection to determine image boundaries, and performs pixel-level optimization to ensure maximum image details and generate image output data;

[0008] The feature extraction module uses a convolutional neural network to extract image texture and structural information based on the image output data, performs deep learning analysis on key features in handprints, seals and signatures, encodes shapes, sizes and pressure points respectively, forms a unified feature descriptor based on the encoding, and generates a feature data set;

[0009] The semantic analysis module performs vocabulary analysis based on the text content in the law enforcement documents, identifies key legal terms, uses semantic association technology to link the legal terms in the text with the actual application scenarios, evaluates the consistency by comparing the terms in the text with the feature data set, and marks the inconsistent parts to generate semantic consistency analysis results;

[0010] The logic verification module performs condition matching based on the semantic consistency analysis results, verifies the legal procedures and regulations in the law enforcement text one by one, uses logic matching to check the legality of each legal step, summarizes all legal steps that have passed the verification, marks abnormal points for legal steps that have not passed the verification, and generates logic compliance confirmation results.

[0011] As a further solution of the present invention, the step of acquiring the image output data is:

[0012] Capture digital images of fingerprints, seals and signatures in law enforcement documents to obtain an initial image dataset;

[0013] The brightness and contrast of the initial image data set are adjusted using the formula,

[0014]

[0015] in, is the adjusted image, is the brightness adjustment factor, used to enhance the overall brightness. is the contrast offset, used to adjust the contrast, is the normalization factor, is an exponential adjustment parameter used to control the nonlinear degree of image brightness enhancement. is a logarithmic gain factor, which is used to enhance the details of low-brightness areas and generate an adjusted image dataset;

[0016] Based on the adjusted image data set, an edge detection algorithm is applied, using the formula,

[0017]

[0018] in, and is the gradient of the image in the x and y directions, is the gradient normalization coefficient, is the edge contrast enhancement coefficient, and the image data set after edge detection is obtained;

[0019] The image data set after edge detection is optimized at pixel level using the formula

[0020]

[0021] in, is the adjusted image, yes The edge-enhanced image, is the edge enhancement coefficient, is the regulating factor, It is the power of the enhancement operation, which is used to adjust the detail level of image processing to obtain image output data.

[0022] As a further solution of the present invention, the step of extracting the image texture and structure information is:

[0023] The image output data is normalized so that the pixel value ranges from 0 to 1, using the formula,

[0024]

[0025] in, is the image output data, and are the minimum and maximum values ​​of the image output data, respectively, to generate a standardized image data set;

[0026] The standardized image data is input into the convolutional neural network, and the texture and structural information of the image is extracted through multi-layer convolution operations. The feature extraction formula is used.

[0027]

[0028] in, and Represents the position coordinates of the pixel, Indicates location The characteristic value of is the radius of the convolution kernel, is the weight in the convolution kernel, It is the normalized pixel value involved in the calculation in the convolution operation to generate a feature map set;

[0029] Perform a pooling operation on the feature map set to reduce the dimension and highlight the key features. Use the pooling formula:

[0030]

[0031] in, and Represents the output location of the pooling operation, and is the offset within the pooling window, is the eigenvalue after pooling, is the pooling window, is the step size of the window movement to generate a pooled feature map set.

[0032] As a further solution of the present invention, the step of acquiring the feature data set is:

[0033] Using the pooled feature map set, deep learning analysis is performed to identify and distinguish key features in fingerprints, seals and signatures. Through the deep learning analysis formula,

[0034]

[0035] in, is the characteristic activity, is the learning weight, adjusting the sensitivity to the target feature, is the total number of feature maps after pooling, is the pooled feature value in the feature map set, is a bias term used to adjust the activation threshold and generate an active feature set;

[0036] Encode the active feature set, describing the shape, size and pressure point, and apply the encoding formula,

[0037]

[0038] in, is the encoding result, is the encoding weight of the target active feature, represents the total number of active features, is a single feature in the active feature set, generating the coded feature set;

[0039] Using the encoding feature set, a unified processing formula is applied,

[0040]

[0041] in, is a unified feature descriptor, is the normalization coefficient of each feature descriptor, ensuring that the feature descriptors are comparable between difference images, is the encoding feature, is the number of features to generate the feature dataset.

[0042] As a further solution of the present invention, the step of identifying the key legal terms is:

[0043] Preprocess the text in the law enforcement documents, remove irrelevant symbols and numbers, and use formulas to

[0044]

[0045] in, is the original text data, is the coefficient that adjusts the steepness of the curve, is a constant, is the parameter that determines the threshold and generates the cleaned text dataset ;

[0046] According to the cleaned text data set, vocabulary parsing is performed,

[0047]

[0048] in, is the weight coefficient, which adjusts the importance of the difference text. is the offset angle, used to adjust the resolution angle. is the first parts, is the total number of text parts, generating the parsed vocabulary set ;

[0049] According to the parsed vocabulary set , analyze and identify key legal terms, using formulas,

[0050]

[0051] in, is the recognition strength of the word, It is parsed from the text vocabulary, is the total number of words, It is the first in the dictionary of legal terms. The weight of each entry is used to generate a set of key legal terms.

[0052] As a further solution of the present invention, the step of obtaining the semantic consistency analysis result is:

[0053] Based on the set of key legal terms, a semantic association is established with the legal terms in the text, using the formula,

[0054]

[0055] in, is the result of semantic association, is the regulating factor, is the kth term in the set of key legal terms, is the context feature associated with the kth term, is the total number of terms, generating a semantic association result set;

[0056] The semantic association result and the feature data set are used to perform consistency assessment, using the formula:

[0057]

[0058] in, is the result of semantic association, is the feature dataset, is the maximum eigenvalue in the feature data set, generating a consistency assessment result set;

[0059] The inconsistent parts of the consistency evaluation result set are marked using the formula,

[0060]

[0061] in, is the result of the consistency assessment, represents a function, is the threshold for consistency judgment, and the semantic consistency analysis result is obtained .

[0062] As a further solution of the present invention, the steps of verifying the legitimacy are:

[0063] Extract key information from the semantic consistency analysis results, perform condition matching, and use the formula:

[0064]

[0065] in, Each clause The importance weight of is the attenuation coefficient of the term, used to adjust the matching sensitivity, is the legal clause obtained from semantic analysis, It is a pre-condition of law. Indicates the total number of terms, yes and The distance between It is the normalized result, and the condition matching result set is obtained;

[0066] Verify the condition matching result set one by one, using the formula for verification.

[0067]

[0068] in, is the target condition matching result, yes The verification weight of is the expected standard deviation of legitimacy, used to adjust the expected legitimacy level, Indicates the total number of matching results. is the root mean square deviation of the verification results, which measures the legitimacy level of each clause and obtains the verification results;

[0069] Perform a logical matching test on the verification results and use a formula to perform a comprehensive score.

[0070]

[0071] in, is the weight of the logical match, is a small positive number that ensures that the domain of the logarithmic function is positive. represents the average legitimacy score of the logic check, Indicates the total number of verification results. It is the result of logic verification and outputs the legality check result.

[0072] As a further solution of the present invention, the steps of obtaining the logic compliance confirmation result are:

[0073] The legality test results are summarized using the formula:

[0074]

[0075] in, Represents a single result of the legality check, used to determine whether it passes. Indicates the total number of results that passed the legality check. represents the weight, represents the weighted total number of legal steps that have passed verification, measuring the total number of passes, Indicates a check function that returns 1 (passed) or 0 (failed). Each result The passing threshold is obtained to obtain the summary results of the legal steps;

[0076] Mark the results that fail the legality check using the formula,

[0077]

[0078] in, Each result The passing threshold is used to calculate the weight of the failed ones. Indicates the weight of failure, adjusts the degree of attention paid to the failed verification results, Represents a check function that returns 1 (passed) or 0 (failed). is the weighted total number of legal steps that failed, Indicates the total number of results that failed the legality check, and obtains the summary results of the failed legal steps;

[0079] The following formula is used to integrate the results of the above mentioned legal steps and the results of the above mentioned legal steps:

[0080]

[0081] in, It is the confirmation result, indicating that the logical compliance confirmation result is output through the ratio.

[0082] Compared with the prior art, the advantages and positive effects of the present invention are:

[0083] In the present invention, by integrating computer vision and semantic analysis technology, accurate recognition and analysis of fingerprints, seals and signatures in law enforcement documents are achieved, and the accuracy of document verification is significantly improved. Boundary detection and pixel-level optimization ensure the fine capture of image details, enabling the system to identify subtle image changes with higher resolution, and improving the system's response speed and sensitivity to complex details. Deep learning technology reduces misjudgment and improves the reliability of authenticity determination during the recognition and encoding of image features. Through automated semantic analysis and logical verification, the system can comprehensively review the compliance of law enforcement texts, enhance the transparency and fairness of law enforcement, and at the same time reduce the investment in human resources and the possibility of human errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 is a system flow chart of the present invention;

[0085] Figure 2 This is a flowchart for obtaining the image output data of the present invention;

[0086] Figure 3 The flowchart of extracting image texture and structure information of the present invention;

[0087] Figure 4 A flowchart for obtaining a feature data set of the present invention;

[0088] Figure 5 A flow chart for identifying key legal terms of the present invention;

[0089] Figure 6 A flowchart for obtaining the semantic consistency analysis result of the present invention;

[0090] Figure 7 A flowchart for verifying the legality of the present invention;

[0091] Figure 8 The figure is a flow chart for obtaining the logical compliance confirmation result of the present invention. DETAILED DESCRIPTION

[0092] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0093] In the description of the present invention, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a target position, be constructed and operated in the target position, and therefore cannot be understood as a limitation on the present invention. In addition, in the description of the present invention, "multiple" means two or more, unless otherwise clearly and specifically defined. Example

[0094] See also Figure 1 , the law enforcement quality supervision system based on computer vision and semantic analysis includes:

[0095] The image capture module captures images based on fingerprints, seals and signatures in law enforcement documents, adjusts image brightness and contrast to enhance image quality, applies edge detection to determine image boundaries, and performs pixel-level optimization to ensure maximum image details and generate image output data;

[0096] The feature extraction module uses a convolutional neural network to extract image texture and structural information based on image output data, performs deep learning analysis on key features in handprints, seals, and signatures, encodes the shape, size, and pressure point, forms a unified feature descriptor based on the encoding, and generates a feature data set;

[0097] The semantic analysis module performs vocabulary analysis based on the text content in law enforcement documents, identifies key legal terms, uses semantic association technology to link legal clauses in the text with actual application scenarios, evaluates consistency by comparing the terms in the text with feature data sets, and marks inconsistent parts to generate semantic consistency analysis results;

[0098] The logic verification module performs conditional matching based on the results of semantic consistency analysis, verifies the legal procedures and regulations in the law enforcement text one by one, uses logic matching to check the legality of each legal step, summarizes all legal steps that have passed the verification, marks abnormal points for legal steps that have not passed the verification, and generates logic compliance confirmation results.

[0099] The image output data includes detail enhancement results, color correction results and image boundary clarity optimization results; the feature data set includes handprint shape identification, seal size measurement results and signature pressure point analysis results; the semantic consistency analysis results include term matching, scene relevance and consistency marking; the logical compliance confirmation results include legal step verification results, anomaly point marking records and compliance assessment results.

[0100] See also Figure 2 , the steps to obtain image output data are:

[0101] Capture digital images of fingerprints, seals and signatures in law enforcement documents to obtain an initial image dataset;

[0102] The brightness and contrast of the initial image dataset are adjusted using the formula,

[0103]

[0104] in, is the adjusted image, is the brightness adjustment factor, used to enhance the overall brightness. is the contrast offset, used to adjust the contrast, is the normalization factor, is an exponential adjustment parameter used to control the nonlinear degree of image brightness enhancement. is a logarithmic gain factor, which is used to enhance the details of low-brightness areas and generate an adjusted image dataset;

[0105] Based on the adjusted image dataset, an edge detection algorithm is applied using the formula,

[0106]

[0107] in, and is the gradient of the image in the x and y directions, is the gradient normalization coefficient, is the edge contrast enhancement coefficient, and the image data set after edge detection is obtained;

[0108] The image dataset after edge detection is optimized at the pixel level using the formula

[0109]

[0110] in, is the adjusted image, yes The edge-enhanced image, is the edge enhancement coefficient, is the regulating factor, It is the power of the enhancement operation, which is used to adjust the detail level of image processing to obtain image output data.

[0111] Assume that in the process of identifying a law enforcement document,

[0112] : Brightness adjustment coefficient, assumed to be 1.2, used to enhance the overall brightness;

[0113] : Contrast offset, assumed to be 0.05, used to fine-tune the contrast;

[0114] : Normalization factor, assumed to be 255 (common in processing 8-bit images, maximum brightness value);

[0115] : Exponential adjustment parameter, assumed to be 0.8, used to adjust the nonlinear degree of brightness enhancement;

[0116] : Logarithmic gain coefficient, assumed to be 0.01, to enhance the details of low-brightness areas;

[0117] Assuming the original pixel value ,but,

[0118] Normalized brightness value:

[0119]

[0120] Brightness adjustment:

[0121]

[0122] Pixel value after brightness enhancement:

[0123]

[0124] Contrast adjustment section:

[0125]

[0126] Final pixel value:

[0127]

[0128] Final calculation results Represents the result of adjusting the brightness and contrast of the original pixel value.

[0129] Assumptions:

[0130] : Directional gradient, assumed to be 20;

[0131] : Directional gradient, assumed to be 15;

[0132] : Gradient normalization coefficient, assumed to be 10;

[0133] : Edge contrast enhancement coefficient, assumed to be 0.5;

[0134] Compute the normalized x gradient:

[0135]

[0136] Compute the normalized у gradient:

[0137]

[0138] Edge strength calculation:

[0139]

[0140] Final edge strength Indicates the strength of the edge detection result of a certain point in the image. The larger the value, the more obvious the edge feature of the point is, which is helpful for subsequent image analysis and processing.

[0141] If the following data exists,

[0142] : The pixel value obtained by the above calculation is 0.618;

[0143] : The edge enhancement image is assumed to be 0.58;

[0144] : Edge enhancement coefficient, assumed to be 1.5;

[0145] : Adjustment factor, assumed to be 0.1;

[0146] : The power of the enhancement operation, assumed to be 2;

[0147] Edge enhancement difference calculation:

[0148]

[0149] Difference normalization:

[0150]

[0151] Apply enhancement factors and powers:

[0152]

[0153] Final pixel enhancement calculation:

[0154]

[0155] The final pixel value Indicates the result of edge enhancement processing on a pixel in the original image, showing a significant enhancement of the edge part, which helps the details of the image stand out more. In subsequent image analysis tasks, such as feature recognition or image classification, the enhanced edges can help the algorithm better understand the image content.

[0156] See also Figure 3 , the steps for extracting image texture and structure information are:

[0157] Normalize the image output data so that the pixel value range is between 0 and 1, using the formula,

[0158]

[0159] in, is the image output data, and are the minimum and maximum values ​​of the image output data, respectively, to generate a standardized image data set;

[0160] The standardized image data is input into the convolutional neural network, and the texture and structural information of the image is extracted through multi-layer convolution operations. The feature extraction formula is used.

[0161]

[0162] in, and Represents the position coordinates of the pixel, Indicates location The characteristic value of is the radius of the convolution kernel, is the weight in the convolution kernel, the normalized pixel value involved in the calculation in the convolution operation, and generates a feature map set;

[0163] Perform pooling operations on the feature map set to reduce the dimension and highlight the key features. Use the pooling formula.

[0164]

[0165] in, and Represents the output location of the pooling operation, and is the offset within the pooling window, is the eigenvalue after pooling, is the pooling window, is the step size of the window movement to generate the pooled feature map set.

[0166] Assume that the following data exists:

[0167] : A pixel value in the image output data, assuming that the pixel value range is 0 to 255;

[0168] : The minimum pixel value in the image output data, assuming the minimum value is 0;

[0169] : The maximum pixel value in the image output data, assuming the maximum value is 255;

[0170] Assume there is a pixel value :

[0171] Calculate the normalization process:

[0172]

[0173] Normalized value Indicates the relative position of the original grayscale value 128 in the range of 0 to , which can be used for network input so that the input data is at the same level for easy processing.

[0174] Feature extraction:

[0175] : The pixel position of the target, assuming the position ;

[0176] : The radius of the convolution kernel, assuming that The convolution kernel is ;

[0177] : The weight in the convolution kernel, the weight is learned through network training, and is simplified to a uniform weight of 1 here;

[0178] : Normalize the values ​​of surrounding pixels in the image;

[0179] If the location Around The values ​​are as follows:

[0180]

[0181] Computational feature extraction:

[0182]

[0183] Indicates that the position is through The sum of features calculated by the convolution kernel can be used for the next layer of network or activation function processing.

[0184] Feature Map Pooling:

[0185] : Pooling window size, assuming the window is , that is, each window contains 4 pixels;

[0186] : Step size, assuming it is 2, means the window moves two pixels each time;

[0187] , : The value is 0 or 1;

[0188] Assuming feature mapping In one The areas are as follows:

[0189]

[0190] Compute pooling:

[0191] calculate :

[0192] For pooling output position , the window coverage area is ,

[0193] Therefore, the window includes: .

[0194] Find the maximum value:

[0195] calculate :

[0196] For pooling output position , the window coverage area is .

[0197] Therefore, the window includes: .

[0198] Find the maximum value:

[0199] calculate :

[0200] For pooling output position , the window coverage area is .

[0201] Therefore, the window includes: .

[0202] Find the maximum value:

[0203] calculate :

[0204] For pooling output position , the window coverage area is .

[0205] Therefore, the window includes: .

[0206] Find the maximum value:

[0207] : Indicates that the most significant eigenvalue in the upper left corner of the image is 10, which indicates that this area has higher feature activity. ,and Similarly, it represents the most significant eigenvalues ​​of other regions.

[0208] See also Figure 4 , the steps to obtain the feature data set are:

[0209] Using the pooled feature map set, deep learning analysis is performed to identify and distinguish the key features in fingerprints, seals and signatures. Through the deep learning analysis formula,

[0210]

[0211] in, is the characteristic activity, is the learning weight, adjusting the sensitivity to the target feature, is the total number of feature maps after pooling, is the pooled feature value in the feature map set, is a bias term used to adjust the activation threshold and generate an active feature set;

[0212] Encode the active feature set, describing the shape, size and pressure point, and apply the encoding formula,

[0213]

[0214] in, is the encoding result, is the encoding weight of the target active feature, represents the total number of active features, is a single feature in the active feature set, generating the coded feature set;

[0215] Using the coded feature set and applying the unified processing formula,

[0216]

[0217] in, is a unified feature descriptor, is the normalization coefficient of each feature descriptor, ensuring that the feature descriptors are comparable between difference images, is the encoding feature, is the number of features, generating a feature dataset.

[0218] Assume the total number of feature maps is 4,

[0219] Eigenvalues ​​after pooling As follows, and each weight is 0.5, bias is 1:

[0220]

[0221] Computational feature analysis:

[0222]

[0223] Represents the output after ReLU activation, indicating the characteristic response strength after activation, which is used for the next step of encoding and analysis.

[0224] Assumptions:

[0225] : Coding weight, adjust the contribution of each feature in the final coding, assuming that they are all 1;

[0226] : Features from the active feature set of the step, assuming there are four features: ;

[0227] : The total number of active features, assumed to be 4;

[0228] Calculate feature encoding:

[0229]

[0230] It represents the encoding result obtained by integrating various features such as shape, size and pressure points. This encoding value will be used to form a unified feature descriptor.

[0231] Assumptions:

[0232] : Unification coefficient, used to balance the contribution of features, assumed to be 1;

[0233] : The encoded feature set from step 2, assumed to be ;

[0234] : The number of features, assuming it is 4;

[0235] Compute unified feature descriptors:

[0236]

[0237] It represents the summarized and standardized feature description value, providing a comprehensive quantitative indicator, representing the unified feature description of handprints, seals and signatures for further identification or classification.

[0238] See also Figure 5 , the steps for identifying key legal terms are:

[0239] Preprocess the text in the law enforcement documents, remove irrelevant symbols and numbers, and use formulas to

[0240]

[0241] in, is the original text data, is the coefficient that adjusts the steepness of the curve, is a constant, is the parameter that determines the threshold and generates the cleaned text dataset ;

[0242] According to the cleaned text data set, vocabulary parsing is performed.

[0243]

[0244] in, is the weight coefficient, which adjusts the importance of the difference text. is the offset angle, used to adjust the resolution angle. is the first parts, is the total number of text parts, generating the parsed vocabulary set ;

[0245] According to the parsed vocabulary set , analyze and identify key legal terms, using formulas,

[0246]

[0247] in, is the recognition strength of the word, It is parsed from the text vocabulary, is the total number of words, It is the first in the dictionary of legal terms. The weight of each entry is used to generate a set of key legal terms.

[0248] Suppose when processing a law enforcement text,

[0249] : Original text data, assuming the text contains 1000 characters;

[0250] The coefficient for adjusting the steepness of the curve is set to 0.01;

[0251] : The parameter that determines the threshold, set to 500, means that text with more than this number of characters requires stronger cleaning;

[0252] Assumptions The valid number of characters is 1000. is a constant (take 2.71828), then,

[0253]

[0254] Calculation results Indicates that the original text is almost completely preserved after processing (close to 1 indicates a low degree of cleaning) and is suitable for further analysis.

[0255] Assumptions:

[0256] : Weight coefficient, set to 1;

[0257] : Offset angle, set to (45°);

[0258] : The first parts, assuming there are 5 parts, each part is 0.993 after preprocessing;

[0259] : Set to 5, indicating that the text is divided into 5 parts, then,

[0260]

[0261] Calculation results It represents the weighted sum after lexical parsing, which can be used for the next step of legal terminology recognition.

[0262] like:

[0263] : The first words, assuming that the value of each word is ;

[0264] : The first in the dictionary of legal terms The weight of each item, assuming that each weight is 1;

[0265] : Assuming it is 5, it means there are 5 key words;

[0266] but,

[0267]

[0268] Calculation results This shows that the recognition results of key legal terms are of high strength and can be used to evaluate the prominence of legal terms in texts.

[0269] See also Figure 6 , the steps to obtain the semantic consistency analysis results are:

[0270] Based on the key legal term set, semantic associations are established with legal clauses in the text, using formulas,

[0271]

[0272] in, is the result of semantic association, is the regulating factor, is the kth term in the set of key legal terms, is the context feature associated with the kth term, is the total number of terms, generating a semantic association result set;

[0273] The semantic association results and feature data sets are used for consistency assessment, using the formula,

[0274]

[0275] in, is the result of semantic association, is the feature dataset, is the maximum eigenvalue in the feature data set, generating a consistency assessment result set;

[0276] Mark the inconsistent parts of the consistency assessment result set using the formula,

[0277]

[0278] in, is the result of the consistency assessment, represents a function, is the threshold for consistency judgment, and the semantic consistency analysis result is obtained .

[0279] Assume the following data exists:

[0280] : Adjustment factor, assuming that each factor is set to 1;

[0281] : Key Legal Terms Collection terms, assuming that the weight of each term is evaluated as 0.8;

[0282] : With The context features related to the terms, assuming that the matching degree of the scene is 0.9;

[0283] : Assume that 5 terms are processed;

[0284] but,

[0285]

[0286] Calculation results This indicates that the strength of the established semantic association is high, indicating that the legal terminology matches the actual application scenarios well.

[0287] Conduct a consistency assessment:

[0288] : The semantic association result obtained from the above calculation is 1.791;

[0289] : The value of the feature data set, assumed to be 2.0;

[0290] : Assume that the maximum eigenvalue is 2.0;

[0291] but,

[0292]

[0293] Calculation results (close to 1) shows a high degree of consistency, indicating a good match between the term and the feature dataset.

[0294] Inconsistent parts mark:

[0295] : The result of the consistency assessment is 0.8955.

[0296] : The consistency judgment threshold is assumed to be 0.5.

[0297] but,

[0298]

[0299] Use the sign function to process the difference. The definition of the sign function is:

[0300] if ,but

[0301] if ,but

[0302] if ,but

[0303] Because 0.3955 is greater than 0, so:

[0304]

[0305] Right now,

[0306]

[0307] Calculation results This indicates that no significant inconsistencies were found, so nothing needs to be labeled, showing good consistency between the text and the feature dataset.

[0308] See also Figure 7 , the steps for checking the legality are:

[0309] Extract key information from the semantic consistency analysis results, perform conditional matching, and use formulas to

[0310]

[0311] in, Each clause The importance weight of is the attenuation coefficient of the term, used to adjust the matching sensitivity, is the legal clause obtained from semantic analysis, It is a pre-condition of law. Indicates the total number of terms, yes and The distance between It is the normalized result, and the condition matching result set is obtained;

[0312] Verify the condition matching result set one by one, using formulas for verification.

[0313]

[0314] in, is the target condition matching result, yes The verification weight of is the expected standard deviation of legitimacy, used to adjust the expected legitimacy level, Indicates the total number of matching results. is the root mean square deviation of the verification results, which measures the legitimacy level of each clause and obtains the verification results;

[0315] Perform a logical match test on the verification results and use the formula for comprehensive scoring.

[0316]

[0317] in, is the weight of the logical match, is a small positive number that ensures that the domain of the logarithmic function is positive. represents the average legitimacy score of the logic check, Indicates the total number of verification results. It is the result of logic verification and outputs the legality check result.

[0318] Suppose there are three legal clauses, the distance between the clauses They are , importance weight They are , attenuation coefficient are all set to 1, then,

[0319]

[0320] The result 1.589 indicates the average degree of match between the terms and the pre-conditions, with higher values ​​indicating better match.

[0321] Assume there are three clauses, validation result validate They are , weight for , preset deviation is 0.7, then,

[0322]

[0323] The result 0.548 represents the average deviation, and the lower the deviation, the higher the legitimacy.

[0324] Assume there are three logic check results logic_check They are , weight for ,set up (Right now minus 10, approximately equal to 0.00001), then,

[0325]

[0326] result Represents the average legitimacy score of the logic check, where values ​​close to zero indicate average legitimacy and negative values ​​indicate some problems or risks.

[0327] See also Figure 8 , the steps to obtain the logical compliance confirmation results are:

[0328] Summarize the results of the legality check and use the formula:

[0329]

[0330] in, A single result of the legality check, used to determine whether it passes. Indicates the total number of results that passed the validity check represents the weight, represents the weighted total number of legal steps that have passed verification, measuring the total number of passes, Indicates a check function that returns 1 (passed) or 0 (failed). Each result The passing threshold is obtained to obtain the summary results of the legal steps;

[0331] Mark the results that fail the legality check using the formula,

[0332]

[0333] in, Each result The passing threshold is used to calculate the weight of the failed ones. Indicates the weight of failure, adjusts the degree of attention paid to the failed verification results, Represents a check function that returns 1 (passed) or 0 (failed). is the weighted total number of legal steps that failed, Indicates the total number of results that failed the legality check, and obtains the summary results of the failed legal steps;

[0334] To integrate the results of the legal steps and the results of the legal steps, use the following formula:

[0335]

[0336] Assume there are three legitimacy test results They are , assuming that the threshold Set to 0.9, and set the weight h to 1, then,

[0337] (because )

[0338] (because )

[0339] (because )

[0340] It follows that

[0341]

[0342] result It said two legal steps passed the legality test.

[0343] Use the same and ,but,

[0344]

[0345] (For the items that pass, the factor is 0, so it does not affect the result), and then we can conclude that,

[0346]

[0347] The result indicated that one legal step failed the test and was flagged.

[0348] Based on the above calculation results, and , it is concluded that

[0349]

[0350] The result H≈0.667 indicates that about 67% of the legal steps meet the compliance standard.

[0351] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. Law enforcement quality supervision system based on computer vision and semantic analysis, characterized by: The system comprises: The image capture module captures images based on fingerprints, seals and signatures in law enforcement documents, adjusts image brightness and contrast to enhance image quality, applies edge detection to determine image boundaries, and performs pixel-level optimization to ensure maximum image details and generate image output data; The feature extraction module uses a convolutional neural network to extract image texture and structural information based on the image output data, performs deep learning analysis on key features in handprints, seals and signatures, encodes shapes, sizes and pressure points respectively, forms a unified feature descriptor based on the encoding, and generates a feature data set; The semantic analysis module performs vocabulary analysis based on the text content in the law enforcement documents, identifies a set of key legal terms, uses semantic association technology to link the legal terms in the text with the actual application scenarios, evaluates the consistency by comparing the terms in the text with the feature data set, and marks the inconsistent parts to generate semantic consistency analysis results; The logic verification module performs condition matching based on the semantic consistency analysis results, verifies the legal procedures and specifications in the law enforcement text one by one, uses logic matching to verify the legality of each legal step, summarizes all legal steps that have passed the verification, marks abnormal points for legal steps that have not passed the verification, and generates a logic compliance confirmation result; The steps for obtaining the semantic consistency analysis result are: Based on the set of key legal terms, a semantic association is established with the legal terms in the text, using the formula, in, is the result of semantic association, is the regulating factor, is the kth term in the set of key legal terms, is the context feature associated with the kth term, is the total number of terms, generating a semantic association result set; The semantic association result and the feature data set are used to perform consistency assessment, using the formula: in, is the result of semantic association, is the feature dataset, is the maximum eigenvalue in the feature data set, generating a consistency assessment result set; The inconsistent parts of the consistency evaluation result set are marked using the formula, in, is the result of the consistency assessment, represents a function, is the threshold for consistency judgment, and the semantic consistency analysis result is obtained .

2. The law enforcement quality supervision system based on computer vision and semantic analysis according to claim 1 is characterized in that: The steps for obtaining the image output data are as follows: Capture digital images of fingerprints, seals and signatures in law enforcement documents to obtain an initial image dataset; The brightness and contrast of the initial image data set are adjusted using the formula, in, is the adjusted image, is the brightness adjustment factor, used to enhance the overall brightness. is the contrast offset, used to adjust the contrast, is the normalization factor, is an exponential adjustment parameter used to control the nonlinear degree of image brightness enhancement. is a logarithmic gain factor, which is used to enhance the details of low-brightness areas and generate an adjusted image dataset; Based on the adjusted image data set, an edge detection algorithm is applied, using the formula, in, and is the gradient of the image in the x and y directions, is the gradient normalization coefficient, is the edge contrast enhancement coefficient, and the image data set after edge detection is obtained; The image data set after edge detection is optimized at pixel level using the formula in, is the adjusted image, yes The edge-enhanced image, is the edge enhancement coefficient, is the regulating factor, It is the power of the enhancement operation, which is used to adjust the detail level of image processing to obtain image output data.

3. The law enforcement quality supervision system based on computer vision and semantic analysis according to claim 1 is characterized in that: The steps for extracting the image texture and structure information are as follows: The image output data is normalized so that the pixel value ranges from 0 to 1, using the formula, in, is the image output data, and are the minimum and maximum values ​​of the image output data, respectively, to generate a standardized image data set; The standardized image data is input into the convolutional neural network, and the texture and structural information of the image is extracted through multi-layer convolution operations. The feature extraction formula is used. in, and Represents the position coordinates of the pixel, Indicates location The characteristic value of is the weight in the convolution kernel, It is the normalized pixel value involved in the calculation in the convolution operation to generate a feature map set; The feature map set is pooled to reduce the dimension and highlight the key features. The pooling formula is used. in, and Represents the output position of the pooling operation, and is the offset within the pooling window, is the eigenvalue after pooling, is the pooling window, is the step size of the window movement to generate the pooled feature map set.

4. The law enforcement quality supervision system based on computer vision and semantic analysis according to claim 3 is characterized in that: The steps for obtaining the feature data set are: Using the pooled feature map set, deep learning analysis is performed to identify and distinguish key features in fingerprints, seals and signatures. Through the deep learning analysis formula, in, is the characteristic activity, is the learning weight, adjusting the sensitivity to the target feature, is the total number of feature maps after pooling, is the pooled feature value in the feature map set, is a bias term used to adjust the activation threshold and generate an active feature set; Encode the active feature set, describing the shape, size and pressure point, and apply the encoding formula, in, is the encoding weight of the target active feature, represents the total number of active features, is a single feature in the active feature set, generating the coded feature set; Using the encoding feature set, a unified processing formula is applied, in, is a unified feature descriptor, is the normalization coefficient of each feature descriptor, ensuring that the feature descriptors are comparable between difference images, is the encoding feature, is the number of features, generating a feature dataset.

5. The law enforcement quality supervision system based on computer vision and semantic analysis according to claim 1 is characterized in that: The steps for identifying the key legal terms are: Preprocess the text in the law enforcement documents, remove irrelevant symbols and numbers, and use formulas to in, is the original text data, is the coefficient that adjusts the steepness of the curve, is a constant, is the parameter that determines the threshold and generates the cleaned text dataset ; According to the cleaned text data set, vocabulary parsing is performed, in, is the weight coefficient, which adjusts the importance of the difference text. is the offset angle, used to adjust the resolution angle. is the first parts, is the total number of text parts, generating the parsed vocabulary set ; According to the parsed vocabulary set , analyze and identify key legal terms, using formulas, in, is the recognition strength of the word, It is parsed from the text vocabulary, is the total number of words, It is the first in the dictionary of legal terms. The weight of each entry is used to generate a set of key legal terms.

6. The law enforcement quality supervision system based on computer vision and semantic analysis according to claim 1 is characterized in that: The steps for verifying the legality are: Extract key information from the semantic consistency analysis results, perform condition matching, and use the formula: in, Each clause The importance weight of is the attenuation coefficient of the term, used to adjust the matching sensitivity, is the legal clause obtained from semantic analysis, It is a pre-condition of law. Indicates the total number of terms, yes and The distance between It is the normalized result, and the condition matching result set is obtained; Verify the condition matching result set one by one, using the formula for verification. in, is the target condition matching result, yes The verification weight of is the expected standard deviation of legitimacy, used to adjust the expected legitimacy level, Indicates the total number of matching results. is the root mean square deviation of the verification results, which measures the legitimacy level of each clause and obtains the verification results; Perform a logical matching test on the verification results and use a formula to perform a comprehensive score. in, is the weight of the logical match, is a small positive number that ensures that the domain of the logarithmic function is positive. represents the average legitimacy score of the logic check, Indicates the total number of verification results. It is the result of logic verification and outputs the legality check result.

7. The law enforcement quality supervision system based on computer vision and semantic analysis according to claim 6 is characterized in that: The steps for obtaining the logic compliance confirmation result are: The legality test results are summarized using the formula: in, Represents a single result of the legality check, used to determine whether it passes. Indicates the total number of results that passed the legality check. represents the weight, represents the weighted total number of legal steps that have passed verification, measuring the total number of passes, Indicates a check function that returns 1 if passed, or 0 if failed. Each result The passing threshold is obtained to obtain the summary results of the legal steps; Mark the results that fail the legality check using the formula, in, Each result The passing threshold is used to calculate the weight of the failed ones. Indicates the weight of failure, adjusts the degree of attention paid to the failed verification results, Indicates a check function that returns 1 for pass or 0 for fail. is the weighted total number of legal steps that failed, Indicates the total number of results that failed the legality check, and obtains the summary results of the failed legal steps; The following formula is used to integrate the results of the above mentioned legal steps and the results of the above mentioned legal steps: in, It is the confirmation result, indicating that the logical compliance confirmation result is output through the ratio.

Images