Generation method and system of MRZ synthetic sample

Abstract

The invention provides an MRZ synthetic sample generation method and system, and is applied to the technical field of certificate data processing. The method comprises the following steps: combining a preprocessed machine readable area corpus and a preset synthesis parameter, simulating quality change, information truncation and physical defects of a machine readable area image, and generating a parameter combination constraint condition based on an international standard and real MRZ image data; processing is carried out on the basis of synthesis parameter combination, corpus preprocessing, sample proportion control and parameter combination constraint conditions, and a machine readable area synthesis sample is generated; processing the generated machine readable area synthesis sample to generate a sample related evaluation result; processing the evaluation result and the parameter setting basis in the synthesis parameters in combination with international standards, and generating a machine readable area synthesis sample generation basis; and processing in combination with parameter combination, sample proportion control, a sample related evaluation result and a parameter compliance feature vector to generate a machine readable region synthetic sample generation optimization result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of certificate data processing, and in particular to a method and system for generating MRZ synthetic samples. BACKGROUND

[0002] In the field of data processing, the training of a machine-readable zone (MRZ) recognition model faces many challenges. From the perspective of data acquisition, due to factors such as data privacy protection, it is difficult to obtain real samples, resulting in a lack of training data. This makes it difficult for the model to fully cover a variety of recognition scenarios, such as when facing complex situations such as character confusion and special font deformation, the recognition effect is poor.

[0003] In terms of training methods, the traditional approach often uses a fixed proportion of real samples and synthetic samples. This approach cannot be dynamically adjusted according to the training process, making it difficult to balance data expansion and precision optimization, and thus leading to the problem of insufficient generalization ability of the model.

[0004] When fine-tuning a pre-trained model, the conventional learning rate strategy also has obvious defects. For example, the fixed learning rate strategy cannot meet the needs of different stages of the model training process, and the simple linear decay strategy can easily cause gradient shock phenomenon, both of which can lead to slow model convergence speed and limited final accuracy, which cannot meet the requirements of high accuracy in practical applications.

[0005] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0006] The purpose of the present application is to provide a method and system for generating MRZ synthetic samples, which at least partially overcomes the problems existing in the prior art. By obtaining preprocessed corpus, synthetic parameters containing ambiguity, and international standards, the image quality is simulated after combination, the parameter association rules are established according to the standards, the conditions containing hard constraints are generated, and then the samples are generated and the compliance is verified. Then extract features, evaluate quality deviation, calculate parameter compliance, and finally analyze risks based on intelligent models, combine historical data and standards to update and generate optimization instructions. The scheme makes the sample comply with international standards, improves the recognition rate by 7%, saves 30% of computing resources, and is suitable for certificate recognition.

[0007] Other characteristics and advantages of the present application will become apparent from the following detailed description, or will be learned by practice of the present application.

[0008] According to an aspect of the present application, a method for generating a machine-readable zone (MRZ) synthetic sample is provided, including: obtaining a MRZ synthetic sample and an international standard, the MRZ synthetic sample including using preprocessed MRZ corpus, specific synthetic parameters, and controlling a proportion of the synthetic sample, wherein the synthetic parameters include blurriness, font size, inter-character spacing, font thickness, and tilt angle, and the corpus preprocessing includes randomly cutting the MRZ corpus in a length range of [5, 14]; the international standard is used to regulate a format, a character set, and a printing quality of a machine-readable area of a travel document; combining the preprocessed MRZ corpus and the preset synthetic parameters, simulating quality variation, information truncation, and physical defects of a MRZ image, establishing a correlation rule between parameters based on the international standard and real MRZ image data, and generating parameter combination constraint conditions; processing based on the synthetic parameter combination, the corpus preprocessing, the sample proportion control, and the parameter combination constraint conditions to generate the MRZ synthetic sample; processing the generated MRZ synthetic sample to generate sample-related evaluation results; processing the evaluation results and parameter settings in the synthetic parameters based on the international standard to generate MRZ synthetic sample generation bases; and processing based on an intelligent model in combination with the parameter combination, the sample proportion control, the sample-related evaluation results, and a parameter compliance feature vector to generate a MRZ synthetic sample generation optimization result.

[0009] According to another aspect of the present application, a device for generating a machine-readable zone (MRZ) synthetic sample is provided, including: an obtaining module configured to obtain a MRZ synthetic sample and an international standard, the MRZ synthetic sample including using preprocessed MRZ corpus, specific synthetic parameters, and controlling a proportion of the synthetic sample, wherein the synthetic parameters include blurriness, font size, inter-character spacing, font thickness, and tilt angle, and the corpus preprocessing includes randomly cutting the MRZ corpus in a length range of [5, 14]; the international standard is used to regulate a format, a character set, and a printing quality of a machine-readable area of a travel document; a processing module configured to combine the preprocessed MRZ corpus and the preset synthetic parameters, simulate quality variation, information truncation, and physical defects of a MRZ image, establish a correlation rule between parameters based on the international standard and real MRZ image data, and generate parameter combination constraint conditions; process based on the synthetic parameter combination, the corpus preprocessing, the sample proportion control, and the parameter combination constraint conditions to generate the MRZ synthetic sample; process the generated MRZ synthetic sample to generate sample-related evaluation results; process the evaluation results and parameter settings in the synthetic parameters based on the international standard to generate MRZ synthetic sample generation bases; and process based on an intelligent model in combination with the parameter combination, the sample proportion control, the sample-related evaluation results, and a parameter compliance feature vector to generate a MRZ synthetic sample generation optimization result.

[0010] According to still another aspect of the present application, an electronic device includes a first processor and a memory storing executable instructions of the first processor, wherein the first processor is configured to perform the above-mentioned method for generating MRZ synthetic samples by executing the executable instructions.

[0011] The method and system for generating MRZ synthetic samples provided by the present application obtain pre-processed corpus (cut 5-14 length), synthetic parameters containing ambiguity, and international standards, combine them to simulate image quality changes, build parameter association rules according to the standards, generate conditions containing hard constraints, and then generate samples and verify compliance. Then, features are extracted, quality deviation is evaluated, parameter compliance is calculated, and finally, risks are analyzed based on an intelligent model, and optimization instructions are generated based on historical data and standards. The scheme makes the samples comply with international standards, improves the recognition rate by 7%, saves 30% of computing resources, and is suitable for certificate identification.

[0012] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and are not limiting to the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 A flowchart of a method for generating MRZ synthetic samples according to an embodiment of the present application is shown.

[0014] Figure 2 A structural schematic diagram of a device for generating MRZ synthetic samples according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0015] The preferred embodiments of the present application are described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application.

[0016] The method for generating MRZ synthetic samples according to the exemplary embodiments of the present application is described below in conjunction with Figure 1 It should be noted that the following application scenarios are only shown for the purpose of facilitating understanding of the spirit and principles of the present application, and the embodiments of the present application are not limited in this respect. On the contrary, the embodiments of the present application are applicable to any applicable scenarios.

[0017] In one embodiment, the present application further provides a method and system for generating MRZ synthetic samples. Figure 1 A flowchart of a method for generating MRZ synthetic samples according to an embodiment of the present application is shown. As shown in Figure 1 the method is applied to a server and includes:

[0018] S101, obtain machine-readable area synthetic samples and international standards, the machine-readable area synthetic samples include using pre-processed machine-readable area corpus, specific synthetic parameters, and controlling the proportion of synthetic samples.

[0019] In an implementation, the machine-readable area corpus is randomly cut into a length between 5 and 14, while the field integrity is preserved through semantic unit identification technology. For example, the original corpus is P<CHEN<<HAO<<<<<<1234567>8901234 (including the name CHEN<<HAO and the ID number 12345678901234), and the field core position is avoided during cutting, such as being cut into P<CHEN<<H (length 7) and AO<<<<<<12345 (length 10), to ensure that the cut string still retains part of the valid semantics, simulating the information truncation scenario caused by real certificate wear.

[0020] The setting of the synthetic parameters (based on the ICAO Doc9303 standard) is as follows: blur: the value range is a Gaussian blur kernel ≤ 5px, for example, the blur is set to 3px, which is used to simulate the image quality degradation caused by the diffusion of printing ink. Font size: it needs to be in the range of 2.0-4.0mm, for example, 2.5mm is selected as the standard font size, which corresponds to the normal printing scenario. Character spacing: the specification is 0.3-0.8mm, and the conventional value is 0.5mm, which is used to simulate the typesetting deviation. Font thickness: it is divided into regular and bold, for example, it is set to bold to simulate the thickening of character edges caused by uneven printing pressure. Angle of inclination: it allows a deviation within ±15°, for example, 8° is set to simulate the inclination of the certificate during scanning.

[0021] Sample proportion control (based on real scene statistics): standard parameter combination (no defects): the proportion is 50%, which is set according to the statistical result that normal certificates account for about 50%-60% in real scenes, and is used as the basic data for model training. Single defect sample (such as blur = 3px): the proportion is 30%, which corresponds to the proportion of single defect certificates (such as slight wear) in reality, and is used to improve the recognition robustness of the model to single defects. Compound defect sample (blur + inclination): the proportion is 20%, which simulates complex defect scenarios such as folding and stains, accounts for about 20% of real certificate defects, and is used to train the comprehensive recognition ability of the model to multi-dimensional defects.

[0022] International standard application (taking ICAO Doc9303 as an example): font size compliance verification: if the parameter is set to 1.8mm (lower than the lower limit of 2.0mm), the system automatically warns and adjusts to 2.0mm to ensure compliance with international standards. Angle of inclination constraint: when the parameter is set to 20° (exceeding the ±15° range), the system forcibly corrects it to 15°, and records the deviation value for subsequent evaluation, to ensure the compliance of the sample generation.

[0023] Parameter association rule (based on printing physical law), font weight and character spacing association: when the font weight is set to bold, the character spacing is automatically increased by 10% (e.g. from 0.5mm to 0.55mm), avoiding character sticking due to bolding, in line with printing process logic. Fuzziness and resolution association: for every 1 level increase in fuzziness (e.g. from 2px to 3px), the resolution is reduced by 5dpi (from 300dpi to 295dpi), simulating the negative correlation between print quality and fuzziness, ensuring that the parameter combination conforms to the physical law.

[0024] The optimized scheme ensures sample authenticity through semantic corpus segmentation, sets parameter baseline based on ICAO international standards, establishes parameter association rules based on printing physical laws, and sets sample proportion based on real scene statistics, forming a complete sample generation system. This system not only meets the requirements of industry standards, but also simulates various defects of real ID cards through multi-dimensional parameter combination, providing scientific and effective data set support for the training of machine readable area (MRZ) recognition model, ensuring that the generated samples have high authenticity and reliability in actual application.

[0025] S102, combine the pre-processed machine-readable area corpus and the pre-set synthesis parameters to simulate the quality change, information truncation and physical defects of the machine-readable area image, establish the association rules between the parameters based on international standards and real MRZ image data, and generate parameter combination constraint conditions.

[0026] In one embodiment, the pre-processed machine-readable area corpus and the pre-set synthesis parameters are combined and analyzed for association degree, generating simulation association degree analysis results of quality change, information truncation and physical defects, wherein the machine-readable area corpus includes randomly segmented data based on semantic unit recognition, and the specific synthesis parameters include fuzziness, font size, character spacing, font weight and inclination angle parameter data simulating image quality change. The pre-set parameters are associated with the semantic segmented corpus to generate simulation association degree results of quality change, information truncation and physical defects. The string P<CHEN<<H (name segment segmentation) and AO<<<<<<12345 (ID number segment segmentation) are segmented by semantic recognition. The pre-set parameters are fuzziness = 3px, font size = 2.5mm, character spacing = 0.5mm, font weight = bold, and inclination angle = 8°. The association between fuzziness and font size: when fuzziness ≥ 3px, font size needs to be ≥ 2.5mm to ensure readability (association degree coefficient 0.78); the association between inclination angle and character spacing: when inclination angle > 5°, character spacing needs to be increased by 5% (0.5mm→0.525mm) to reduce character overlap (association degree coefficient 0.65); the association between corpus segmentation position and physical defects: when the segmentation point is located at the beginning of the field, the probability of simulating ID card edge wear increases by 40%.

[0027] The correlation analysis results are filtered to generate a target simulation variable list, which includes ambiguity parameters, font size parameters, character spacing parameters, font weight parameters, tilt angle parameters, and corpus random segmentation parameters. Key variables are selected from the correlation analysis results to form the target simulation variable list. Variables with a correlation coefficient > 0.6 are included in the list. The target simulation variable list includes: ambiguity parameters (0-5px, step 1px); font size parameters (2.0-4.0mm, step 0.5mm); character spacing parameters (0.3-0.8mm, step 0.1mm); font weight parameters (regular / bold); tilt angle parameters (0-15°, step 1°); and corpus random segmentation parameters (segmentation position, field integrity markers).

[0028] The results of the correlation analysis are filtered to generate parameter combination constraints, including hard constraints, collaborative constraints, and semantic constraints that conform to international standards. Based on the correlation analysis and international standards, hard constraints, collaborative constraints, and semantic constraints are generated. Hard constraints (international standards) are: font size ∈ [2.0, 4.0] mm (ICAODoc9303); tilt angle ≤ 15°. Collaborative constraints (physical laws): font weight = bold → character spacing = base value × 1.1 (e.g., base character spacing 0.5mm → 0.55mm); for every 1px increase in ambiguity → resolution decreases by 5dpi (300dpi → 295dpi). Semantic constraints (corpus processing) are: the segmentation point must not destroy the core semantics of the field (e.g., the first letter of the name cannot be truncated); the segmented string must contain at least one valid character segment (e.g., the first 3 digits of the ID number).

[0029] Based on parameter combination constraints, multi-dimensional simulation input variables are generated. These input variables characterize key features associated with changes in image quality, information truncation, and physical defects in the machine-readable region. Input variables containing features of quality changes, information truncation, and physical defects are generated based on the constraints. Input variable 1 (single defect):

[0030] {"blur":3px,"font_size":2.5mm,"letter_spacing":0.5mm,"font_weight":"bold","slant_angle":0°,"cut_position":"Middle of field","field_integrity":"Name field incomplete"} (Simulation: blurry printing + bold font + worn document with middle segmentation).

[0031] Input variable 2 (compound defect):

[0032] {"blur":2px,"font_size":3.0mm,"letter_spacing":0.55mm,"font_weight":"bold","slant_angle":10°,"cut_position":"Field start","field_integrity":"ID number segment missing"} (Simulation: A composite defect of slight blur + bold font + 10° slant + start segmentation).

[0033] Through a four-layer logic of correlation analysis, variable selection, constraint generation, and variable combination, a systematic design of parameter combinations is achieved: quantitative correlations are established between parameters based on real MRZ data and international standards, avoiding physical contradictions caused by independent parameter settings; hard constraints directly reference ICAO standards to ensure samples conform to industry norms; semantic constraints and collaborative constraints are combined to make the simulated quality changes and information truncation more closely resemble real document defect scenarios. The multi-dimensional input variables generated by this process can effectively cover defect simulations in both single and composite scenarios, providing high-quality synthetic samples for MRZ recognition model training.

[0034] S103 generates machine-readable synthetic samples by processing based on the combination of synthesis parameters, corpus preprocessing, sample ratio control, and parameter combination constraints.

[0035] In one implementation, feature extraction is performed on the combination of synthesis parameters, corpus preprocessing, and sample ratio control to generate ambiguity features, font size features, character spacing features, font thickness features, tilt angle features, corpus random segmentation features, and sample ratio control features. The synthesis parameter combination features include specific combination parameter data conforming to international standards for ambiguity, font size, character spacing, font thickness, and tilt angle. The sample ratio control features include sample ratio setting data for different parameter combinations. Features from parameter combinations, corpus preprocessing, and sample ratios are extracted to generate structured feature data. For the synthesis parameter combination features, the input parameter combinations are: ambiguity = 3px, font size = 2.5mm, character spacing = 0.55mm (bold text automatically increases by 10%), and tilt angle = 8°.

[0036] The feature extraction results are as follows:

[0037] {"blur_feature":3,"font_size_feature":2.5,"letter_spacing_feature":0.55,"font_weight_feature":"bold","slant_angle_feature":8,"is_standard_compliant":true / / Comply with ICAO standards}.

[0038] The corpus preprocessing features include the original corpus: P<CHEN<<HAO<<<<<<1234567>8901234; the segmented corpus: P<CHEN<<H (length 7, segmented in the middle of the name segment, field integrity = partial loss).

[0039] The feature extraction results are as follows:

[0040] {"cut_length":7,"cut_position":"middle","field_integrity":"partial_loss","semantic_unit":"na me"}. The sample ratio control features include the preset ratios: standard sample 50%, single defect 30%, compound defect 20%; the current parameter combination belongs to "single defect (blur degree 3px)", corresponding ratio allocation: 30%.

[0041] Feature extraction processing is performed on the generation process of generating machine-readable area synthetic samples to generate parameter combination simulation features, corpus preprocessing simulation features, and sample ratio control features. Among them, the parameter combination simulation features include the simulation parameter combination data of image quality change and physical defects; the corpus preprocessing simulation features include the randomly segmented data of the truncated corpus; the sample ratio control features include the ratio allocation data of samples with different parameter combinations. Extract the parameter combination simulation, corpus preprocessing simulation, and ratio control features in the generation process. Specifically, the parameter combination simulation feature is the simulated defect: print blur + font boldening;

[0042] The feature extraction results are as follows:

[0043] {"quality_change":"blur","physical_defect":"bold_enlargement","parameter_combination":"blur=3+font_weight=bold"}.

[0044] The corpus preprocessing simulation feature is the simulated information truncation: segmented in the middle of the name segment, simulating character loss caused by document wear.

[0045] The feature extraction results are as follows:

[0046] {"truncation_type":"mechanical_wear","truncation_position":"name_field_middle","characte r_loss_count":4}.

[0047] The sample proportion control feature is used to generate the actual sample type: single defect sample, and the feature extraction result is:

[0048] {"current_ratio_type":"single_defect","remaining_ratio":{"standard":50%,"compound_defect":20%}}.

[0049] Based on the parameter combination constraint condition, combined with the synthetic parameter combination feature, the corpus preprocessing feature, the sample proportion control feature, and the parameter combination simulation feature, the corpus preprocessing simulation feature, and the sample proportion control feature, the parameter combination is verified whether it meets the international standard when generating the sample, and the machine readable area synthetic sample is generated, wherein the machine readable area synthetic sample is used to represent the quality change, information truncation and physical defect simulation sample feature of the MRZ image, and the MRZ synthetic sample generation result containing parameter combination, corpus preprocessing and sample proportion control is formed. Combined with the feature extraction result, the parameter combination compliance is verified and the sample is generated. The parameter combination verification is as follows, the input parameters: tilt angle = 12°, blur = 5px; constraint condition verification: tilt angle 12°≤15°(comply with hard constraint); blur 5px≤5px(comply with hard constraint); collaborative constraint: when blur is 5px, the resolution is automatically adjusted to 275dpi(300dpi-5×5dpi). The verification result is compliant, and the generation is allowed. The sample generation result is as follows, the synthetic sample feature is {"defect_types":["blur=5","slant=12°"],"compliance_grade":"A", / / comply with ICAO standard"sample_type":"compound_defect","visualization":"MRZ image with blur and tilt, name segment is cut in the middle"}

[0050] Through the three-layer processing of feature extraction-process simulation-compliance verification, the systematic generation of MRZ synthetic sample is realized. The parameter combination, corpus segmentation and proportion control are converted into structured features, ensuring that the generation process is traceable, the parameter combination is verified based on ICAO standard and physical law, and the generation of non-compliant or physically contradictory samples is avoided; combined with semantic segmentation and parameter association rules, the quality change and information truncation generated are closer to the real certificate defect scene. The sample generated by this process can not only meet the requirements of international standards, but also cover single scene and compound scene defect simulation, providing training data for MRZ recognition model with compliance and authenticity.

[0051] S104, processing the generated machine readable area synthetic sample to generate sample related evaluation results.

[0052] In one implementation, based on the combination of synthesis parameters, corpus preprocessing features, and sample proportion control features, the quality variation, information truncation, and physical defect simulation features of the generated machine-readable area synthesis samples are extracted to generate a sample feature vector set. Based on the combination of parameters, corpus preprocessing, and proportion control features, the quality variation, information truncation, and physical defect features of the samples are extracted. The input sample parameters are blur = 3px, font size = 2.5mm (standard range), letter spacing = 0.55mm (bold + 10%), slant angle = 8°, corpus segmentation = middle segmentation (length 7) in the name segment, sample type = single defect (blur). The feature vector extraction results are as follows:

[0053] [{"blur":3,"font_size":2.5,"letter_spacing":0.55,"slant_angle":8,"cut_position":"middle","defect_type":"single_blur","compliance":"compliant"}]。

[0054] By comparing the sample feature vector sets under different parameter combinations, the law of the quality of machine-readable area synthesis samples changing with parameters is determined, and sample quality change evaluation information is generated. By comparing the feature vectors of different parameter combinations, the law of the quality changing with parameters is determined. The comparison scenarios are as follows, scenario 1: blur = 2px, font size = 2.0mm → OCR recognition rate 92%; scenario 2: blur = 3px, font size = 2.5mm → OCR recognition rate 85%; scenario 3: blur = 5px, font size = 3.0mm → OCR recognition rate 70%.

[0055] For every 1px increase in blur, the font size needs to increase by at least 0.5mm to maintain a recognition rate decrease of ≤5%, and the quality change evaluation information is generated: {"blur_impact":"for every 1px increase, the recognition rate decreases by an average of 7%","font_size_compensation":"needs to be adjusted by a ratio of 0.5mm / 1px"}.

[0056] The sample quality change evaluation information is compared with a preset normal sample mode library and an abnormal sample mode library containing international standard samples, and a sample quality deviation evaluation result is generated. The evaluation information is compared with the preset mode library, and the deviation is calculated. Normal sample mode library standard: fuzziness ≤ 2px, font size 2.0-3.0mm, recognition rate ≥ 90%; current sample parameters: fuzziness = 3px, recognition rate = 85%; deviation calculation: fuzziness deviation: (3-2) / 2 = 50%; recognition rate deviation: (90-85) / 90 ≈ 5.56%; comprehensive deviation: (50%+5.56%) / 2 = 27.78%. The evaluation result is as follows:

[0057] {"deviation_level":"moderate","standard_reference":"fuzziness exceeds 50%, recognition rate is close to threshold"}.

[0058] The sample quality deviation evaluation result is weighted processed in combination with the historical evaluation data of the machine readable area synthetic sample, and the rationality of the parameter combination, to generate a sample evaluation risk feature vector. The deviation result is weighted processed in combination with the historical data and the parameter rationality. The historical data weight is the fuzzy degree exceeding standard sample historical bad rate: 30% (weight 40%); the current parameter combination rationality score: 70 points (weight 60%). The weight distribution is based on the influence degree analysis of the MRZ sample quality risk, and the specific is as follows: historical data weight (40%): reference historical evaluation data, the same parameter deviation (such as font size, fuzzy degree exceeding standard) leads to the frequency of recognition rate decrease (such as in the past 3 times of evaluation, the risk proportion caused by historical deviation is 40%), which reflects the influence of historical data on current risk. Parameter rationality weight (60%): the rationality of current parameter combination directly determines the compliance and effectiveness of sample generation, and has greater real-time influence on risk, so it is given higher weight. In addition, the weight can be dynamically adjusted according to actual scene (such as international standard update or sample defect type change, the weight proportion can be re-distributed).

[0059] The risk vector is calculated as follows, and the formula for calculating the risk value is: risk value = (deviation degree x historical data weight) + ((1-rationality coefficient) x parameter rationality weight). By substituting the above values into the formula, the risk value = 0.2778 x 40% + (1-70 / 100) x 60% = 0.111 + 0.18 = 0.2911. The percentage form of the deviation degree is converted to a decimal in the [0, 1] interval (such as 27.78% → 0.2778); the rationality score of 100 points is converted to a rationality coefficient in the [0, 1] interval (such as 70 points → 70 / 100 = 0.7), ensuring that the values of the two dimensions can be directly operated in the same dimension. The risk feature vector is {"risk_level":"medium","historical_correlation":0.3,"parameter_rationality":0.7}.

[0060] The sample quality deviation degree evaluation result and the sample evaluation risk feature vector are normalized and integrated to generate a sample evaluation feature parameter set containing a quality deviation coefficient, a sample validity index, and a parameter compliance index. The sample evaluation feature parameter set is composed of a sample feature vector set, sample quality change evaluation information, sample quality deviation degree evaluation result, sample evaluation risk feature vector, quality deviation coefficient, and sample validity index. After normalization processing, the comprehensive evaluation parameters are generated, and the normalization operation is as follows: quality deviation coefficient: 27.78% → 0.28 (normalized to [0, 1]); sample validity index: 1-29.11% = 0.71; parameter compliance index: 1 (complying with font size, tilt angle, etc.). Comprehensive parameter set: {"feature_vectors": [first step feature vector], "quality_change": [second step evaluation information], "deviation_result": [third step deviation degree], "risk_vector": [fourth step risk vector], "quality_bias": 0.28, "validity_index": 0.71, "compliance_index": 1.0}.

[0061] The five-layer process of feature extraction, rule analysis, deviation evaluation, risk weighting and comprehensive operation is used to realize the systematic evaluation of the MRZ synthetic sample. The non-structured information such as quality change and defect type is converted into a calculable feature vector to ensure the objectivity of the evaluation. The sample compliance and quality deviation degree are quantified by comparing with the international standard sample mode library. The reliability of risk evaluation is improved by combining with the weighted processing of historical evaluation data. The multi-dimensional evaluation matrix formed by the quality deviation coefficient and effectiveness index provides a clear direction for sample optimization. The feature parameter set generated by the evaluation process can not only reflect the degree of conformity of the sample to the international standard, but also quantify the influence of defects on the identification effect, providing a scientific basis for subsequent intelligent optimization.

[0062] S105, the evaluation results and parameter settings in the synthetic parameters are combined with the international standards to generate machine-readable zone synthetic sample generation basis.

[0063] In an embodiment, based on the parameter evaluation results and the international standard alignment method, the deviation value of the actual value of the blur degree, font size, character spacing, font thickness and tilt angle in the synthetic parameters from the international standard is calculated to generate parameter deviation data. The parameter deviation calculation is based on the absolute deviation method, that is, the deviation degree is quantified by the difference between the actual parameter value and the standard threshold value, and the formula is: deviation value = actual value - standard threshold value. For interval type standards (such as font size 2.0-4.0mm), the lower limit is 2.0mm and the upper limit is 4.0mm; for single direction type standards (such as tilt angle ≤15°), the threshold value is 15°.

[0064] The international standard detail analysis (ICAODoc9303) is as follows: the standard range of font size is 2.0-4.0mm, which is used to ensure the character recognition accuracy of OCR equipment; too small font size (<2.0mm) may cause character edge blur during scanning, and too large font size (>4.0mm) exceeds the layout specification of the machine-readable area of the certificate. Tilt angle, standard threshold: ≤15°, allows normal deviation of the certificate during scanning or shooting; tilt exceeding 15° will significantly reduce the OCR recognition rate, which needs to be controlled by parameter deviation. Character spacing, standard range: 0.3-0.8mm, to avoid character sticking or sparseness; character spacing <0.3mm may cause character overlap, and >0.8mm affects the compactness of the layout.

[0065] Actual parameter value and deviation calculation process as follows, font size deviation, actual value: 1.8mm; Standard lower limit: 2.0mm; Deviation calculation is 1.8mm-2.0mm=-0.2mm. Deviation property is negative deviation (lower than the standard lower limit), which may lead to increased difficulty in character recognition. Inclination angle deviation, actual value: 12°; Standard threshold: 15°; Deviation calculation is 12°-15°=-3° (since the standard is a one-way threshold, the actual deviation is taken as 0°) compliance judgment: 12°≤15°, the deviation value is recorded as 0°, which meets the standard. Character spacing deviation, actual value: 0.2mm; Standard lower limit: 0.3mm; Deviation calculation is 0.2mm-0.3mm=-0.1mm, the deviation property is negative deviation (lower than the standard lower limit), which may lead to character sticking.

[0066] The impact of font size-0.2mm, experimental data is that the OCR recognition rate decreases by an average of 3% for every 0.1mm decrease in font size; In this example, -0.2mm may cause the recognition rate to decrease by about 6%, which needs to be focused on in sample evaluation. The impact of character spacing-0.1mm, character spacing 0.2mm is close to 40% of the character width (about 0.5mm), which may cause adjacent character edge overlap; Simulate the ink diffusion or typesetting error when printing the certificate.

[0067] Through the standardized deviation calculation method, the parameter compliance can be converted into a quantitative index, providing data support for subsequent parameter optimization. In this example, the negative deviation of font size and character spacing needs to be adjusted or corrected by parameter (such as automatically increasing the font size to 2.0mm and the character spacing to 0.3mm) during sample generation to ensure that the synthesized sample meets the international standard and improves the effectiveness of the OCR model training.

[0068] Filter the parameter deviation data for effectiveness. Deviation data filtering is based on the dual standards of business impact and deviation significance: evaluate whether the deviation has a substantial impact on MRZ recognition effectiveness or international standard compliance; Determine whether it belongs to noise or measurement error by the absolute value of the deviation.

[0069] The original deviation data is analyzed as follows:

[0070]

[0071]

[0072] The threshold for absolute value of deviation is set as follows, core parameters (font size, blur, character spacing): threshold 0.05mm / 0.1px, that is, deviation absolute value≥0.05mm / 0.1px is considered effective; Non-critical parameters (font thickness, inclination angle, etc.): threshold can be relaxed to 0.1mm, but needs to be combined with business impact.

[0073] Business impact priority, specifically, high priority parameters include font size, blur, and character spacing: directly affect character readability, deviations need to be 100% included in evaluation; low priority parameters include the following, font thickness: only affects character edge contour, minor deviations (such as 0.03mm) can be filtered; tilt angle: in this case, 12°≤15°, deviation is 0°, no need to filter.

[0074] The filtering process and result analysis are as follows: font size deviation (-0.2mm), deviation absolute value 0.2mm>0.05mm, and belongs to high priority parameter, retained. Blur deviation (+0.5px), deviation absolute value 0.5px>0.1px (blur parameter threshold), and exceeds the upper limit of the standard, retained. Character spacing deviation (-0.1mm), deviation absolute value 0.1mm>0.05mm, and affects character spacing, retained. Font thickness deviation (+0.03mm), deviation absolute value 0.03mm<0.05mm, and belongs to non-critical parameter, filtered.

[0075] The retained deviations are all parameters that have a significant impact on MRZ recognition (font size, blur, and character spacing), avoiding non-critical deviations from interfering with the evaluation focus. After filtering out noise data, the parameter compliance calculation amount is reduced by about 30%, while ensuring the evaluation accuracy. For the character spacing -0.1mm deviation, the character spacing can be automatically adjusted to 0.3mm during sample generation, and the linkage effect of font size and blur is verified simultaneously.

[0076] The effectiveness filtering realizes the noise reduction processing of deviation data by combining the quantification standard and business logic, ensuring that the subsequent compliance evaluation focuses on the parameter abnormalities that have a substantial impact on the quality of MRZ samples. This mechanism avoids the evaluation distortion caused by "deviation overload", provides more accurate input data for intelligent optimization, and effectively improves the standard compliance of synthetic samples and the training efficiency of recognition models.

[0077] Based on the filtered parameter deviation data, the degree to which the synthetic parameters comply with international standards is calculated to generate parameter compliance data. The compliance calculation uses the relative deviation ratio method, and the core logic is to convert the parameter absolute deviation into the ratio relative to the standard allowed range, the formula is: compliance = 1-(|deviation value| / standard allowed range upper limit). This model quantifies compliance as a value in the [0,1] interval, 1 represents complete compliance with the standard, and 0 represents complete non-compliance; the standard allowed range upper limit is dynamically determined according to the parameter type, interval type parameters (such as font size): the range upper limit is the difference between the standard maximum value and the minimum value (4.0mm-2.0mm=2.0mm); single-direction type parameters (such as blur): the range upper limit is the standard threshold (5px).

[0078] The individual compliance calculation process is as follows: font size compliance, deviation value: |-0.2mm| = 0.2mm (absolute value of negative deviation); standard allowed range upper limit: 4.0mm-2.0mm = 2.0mm. The calculation process is as follows: compliance = 1-(0.2mm / 2.0mm) = 0.9 (90%). The font size deviation accounts for 10% of the standard range, which is still in the acceptable interval, but attention should be paid to the impact on recognition rate.

[0079] Blur compliance, deviation value: |+0.5px| = 0.5px (absolute value of positive deviation); standard allowed range upper limit: 5px (one-way standard threshold). The calculation process is as follows: compliance = 1-(0.5px / 5.0px) = 0.9 (90%); the blur exceeds the standard by 10%, which may cause the image to be slightly blurred, but has not yet broken the threshold of seriously affecting recognition.

[0080] Letter spacing compliance, deviation value: |-0.1mm| = 0.1mm (absolute value of negative deviation); standard allowed range upper limit: 0.8mm-0.3mm = 0.5mm; the calculation process is as follows: compliance = 1-(0.1mm / 0.5mm) = 0.8 (80%); the letter spacing deviation accounts for 20% of the standard range, there is a risk of character sticking, which needs to be adjusted first when generating samples.

[0081] The comprehensive compliance data structure is parsed as follows: {"individual_compliance": {"font_size": 0.9, / / font size compliance 90% "blur": 0.9, / / blur compliance 90% "letter_spacing": 0.8 / / letter spacing compliance 80%}, "average_compliance": 0.87 / / average compliance 87%}. The individual compliance reflects the independent compliance status of each parameter, which is convenient for positioning specific problems; the average compliance provides an overall compliance overview through arithmetic mean ((0.9+0.9+0.8) / 3 = 0.87), which is suitable for batch evaluation of samples.

[0082] Set thresholds (such as average compliance ≥ 0.85) to automatically filter low-compliance samples and ensure that the training data set meets international standards. In the case of low letter spacing compliance (0.8) compared to font size and blur, the letter spacing parameter is adjusted first to improve overall compliance. When the compliance of a certain parameter is <0.8 (such as letter spacing 0.8), a warning is triggered, prompting that the parameter combination may cause the recognition rate to decrease.

[0083] The parameter deviation can be converted into a quantitative compliance index through a standardized compliance calculation model, providing data support for quality evaluation and optimization of the MRZ synthetic sample. The model takes into account both computational efficiency and business logic, enabling quick identification of parameter abnormalities and providing standardized input for intelligent models. After optimization of parameter weights and nonlinear mapping, the accuracy of compliance evaluation can be further improved, ensuring that the generated sample not only meets international standards but also effectively simulates various defects in real-world scenarios.

[0084] The parameter compliance data is weighted based on historical evaluation data of the synthesis parameters and the update frequency of international standards to generate a parameter compliance feature vector. The historical evaluation data and the standard update frequency are combined to generate the feature vector. The historical data weight is the font size historical deviation influence coefficient: 0.4 (in the last 3 evaluations, the font size deviation caused a 40% decrease in recognition rate) The international standard update frequency: ICAO standard has not been updated in the past 5 years, with a weight coefficient of 0.9 (stable) After weighted calculation, the compliance feature vector = comprehensive compliance x historical influence weight x standard stability weight; Font size: 0.9 x 0.4 x 0.9 = 0.324; Blur: 0.9 x 0.2 x 0.9 = 0.162; Character spacing: 0.8 x 0.3 x 0.9 = 0.216; The final feature vector is [0.324, 0.162, 0.216].

[0085] Through the four-layer process of deviation calculation, effectiveness filtering, compliance scoring, and weighting processing, the quantitative evaluation of parameter compliance is achieved: the parameter deviation is converted into a calculable numerical index, ensuring the objectivity of compliance evaluation; noise deviation is excluded, focusing on parameter abnormalities that have a substantial impact on recognition effectiveness; historical evaluation data and standard update frequency are combined to make the compliance feature more relevant to actual application scenarios; multi-dimensional compliance data is integrated into a unified vector to provide standardized input for intelligent optimization. The parameter compliance feature vector generated by this process can not only reflect the standard compliance degree of the current parameter combination, but also dynamically adapt to historical data patterns and changes in industry standards, providing a scientific compliance basis for subsequent intelligent model optimization.

[0086] S106, based on the intelligent model, combining parameter combination, sample proportion control, sample correlation evaluation results, and parameter compliance feature vector for processing, generating machine-readable zone synthetic sample generation optimization results.

[0087] In one implementation, the parameter combination, sample proportion control and sample correlation evaluation results are risk level analysis processing, and the parameter combination risk level evaluation results are generated. Based on the parameter combination, sample proportion and evaluation results, the risk level is analyzed. The input data is the parameter combination: blur = 5px (standard ≤ 5px), font size = 1.8mm (standard 2.0-4.0mm); sample proportion: high blur sample proportion 30%; evaluation results: quality deviation coefficient = 0.28 (threshold 0.2), parameter compliance index = 0.87.

[0088] The risk level is calculated as follows: the negative deviation of font size (-0.2mm) causes the recognition rate to drop by 6%, with a risk weight of 40%; the high blur sample proportion exceeds the historical average (20%), with a risk weight of 30%; the quality deviation coefficient exceeds the threshold, with a risk weight of 30%; the comprehensive risk level is medium risk (65%).

[0089] Based on the quality risk characteristics in the sample correlation evaluation results and the preset quality threshold, the quality abnormality determination result is generated. The quality abnormality determination is based on a two-dimensional threshold comparison method, which realizes the automatic identification of abnormal state by quantitatively comparing the actual characteristics of the sample with the preset quality threshold. The core logic framework is as follows: threshold system construction: establish threshold standards for core indicators such as recognition rate and compliance index; feature data acquisition: extract key quality characteristics from sample evaluation results; threshold comparison operation: determine whether the characteristics exceed the allowed range through numerical comparison; abnormal type classification: generate specific abnormal type and reason description according to the comparison result.

[0090] The preset threshold system is analyzed, and the details are shown in the following table:

[0091]

[0092] The influence analysis of the recognition rate of 82% is as follows, the data source is the actual recognition rate of the sample by the OCR model; the deviation reason is that the font size = 1.8mm (lower than the lower limit of the standard 2.0mm), which causes the character edge to be blurred; the blur = 5px (close to the upper limit of the standard 5px), which further reduces the character clarity; the recognition rate below 85% may cause the misrecognition rate in the actual certificate recognition scene to rise to more than 5% (normal scene misrecognition rate ≤2%).

[0093] The composition analysis of the compliance index of 0.87 is as follows: font size compliance: 0.9 (deviation -0.2 mm); blurriness compliance: 0.9 (deviation +0.5 px); character spacing compliance: 0.8 (deviation -0.1 mm); weighted calculation: (0.9+0.9+0.8) / 3=0.87. The character spacing compliance (0.8) has approached the lower threshold, which may cause the character sticking defect to exceed the range allowed by the international standard. The abnormality judgment of the recognition rate is as follows: comparison formula: 82%<85%→result true; the abnormality type is classified as recognition effect abnormality; the severity is moderate abnormality (deviation 3%, not breaking the emergency threshold 80%). The abnormality judgment of the compliance index is as follows: comparison formula: 0.87<0.9→result true; the abnormality type is classified as parameter compliance abnormality; the severity is critical abnormality (deviation 0.03, close to but not completely over-standard). The comprehensive abnormality judgment result is as follows: abnormality type: compound abnormality (recognition rate not up to standard + compliance critical); risk level: medium risk; which may cause the model to be too adaptable to low-quality samples during training, reducing the recognition robustness in actual scenarios.

[0094] Automatically mark abnormal samples to avoid including samples with a recognition rate <85% or a compliance index <0.9 in the core training set; add the [parameter adjustment needed] label to the current sample to prevent it from entering the model fine-tuning stage. Parameter optimization guidance: the recognition rate abnormality points to font size and blurriness parameters that need to be adjusted; the compliance abnormality points to the character spacing parameter that needs to be corrected first (such as adjusting from 0.2 mm to 0.3 mm). Trigger parameter combination risk warning and prompt that the combination of "font size + blurriness + character spacing" has a synergistic risk; correlate historical data and retrieve the processing scheme of similar abnormal samples in the past 3 months (such as increasing the font size by 0.2 mm to improve the recognition rate to 88%).

[0095] Threshold self-adaptive adjustment: recalibrate the threshold based on quarterly historical data: if the average recognition rate of qualified samples in a quarter is improved to 88%, the threshold can be increased to 86%; if the compliance requirement is improved due to the update of international standards, the compliance index threshold is also adjusted to 0.95. Introduce a weight mechanism to set different weights for different thresholds: recognition rate threshold weight 60% (directly affecting model effect); compliance index weight 40% (ensuring industry standard compliance); comprehensive abnormality score = recognition rate deviation x 60% + compliance deviation x 40% = 3% x 60% + 3.33% x 40% = 3.13%.

[0096] Abnormal grading response, mild abnormality (e.g. recognition rate 84%): only record non-intervention, observe subsequent sample trend; moderate abnormality (e.g. current sample): trigger parameter fine-tuning suggestion; severe abnormality (recognition rate < 80%): suspend sample generation and start full-process diagnosis. The quality abnormality judgment realizes the quantitative identification and classification of the quality problems of the MRZ synthetic samples by constructing a standardized threshold comparison system. This mechanism can accurately locate the potential risks of recognition effect and compliance, and also provide clear basis for parameter optimization and sample screening. In actual application, combined with dynamic threshold adjustment and weight mechanism, the accuracy and adaptability of abnormality judgment can be further improved to ensure that the generated samples always meet the requirements of international standards and model training needs.

[0097] The parameter combination risk level evaluation result and the quality abnormality judgment result are processed to generate sample generation risk warning information. Specifically, the warning content is that the parameter combination (blur = 5px + font size = 1.8mm) has moderate risk: the recognition rate is lower than the threshold value 3%, and the compliance index is close to the critical value; the proportion of high blur samples is 10% higher than the historical average, which may cause model overfitting.

[0098] The historical generation information of the machine-readable area synthetic sample, the intelligent model configuration information and the international standard update record are obtained. The historical data is that in the past 3 months, the sample with blur > 4px and font size < 2.0mm, the average OCR recognition rate decreases by 8%. The international standard is updated, and ICAO recently issued a supplementary standard, which requires that the lower limit of font size is adjusted to 2.1mm (originally 2.0mm).

[0099] Based on the historical generation information of the synthetic sample, the sample generation risk warning information and the international standard update record, the intelligent model configuration information is processed to generate the synthetic sample generation optimization instruction of the machine-readable area. The optimization instruction generation is based on a three-dimensional driving model, which integrates historical generation data, real-time risk warning and international standard update to form a systematic parameter adjustment and model configuration optimization scheme. Its logic framework is as follows: data integration layer: collect historical sample effect data, current risk warning information and standard update clauses; rule reasoning layer: generate candidate solutions through preset optimization rules (such as standard threshold triggering parameter adjustment) and historical experience (such as a certain parameter combination has improved recognition rate); instruction generation layer: convert the reasoning result into executable parameter adjustment instruction and model configuration rule.

[0100] The font size optimization logic is adjusted according to the international standard update: the latest supplementary standard of ICAO increases the lower limit of font size from 2.0mm to 2.1mm; historical data verification: the data in the past 6 months shows that the recognition rate of samples with font size ≥ 2.1mm is 7% higher than that of samples with font size < 2.1mm on average; risk warning correlation: the current sample font size = 1.8mm triggers compliance critical abnormality (compliance index 0.87 < 0.9).

[0101] The adjustment basis of ambiguity control is risk warning: the proportion of high ambiguity samples (ambiguity ≥ 5px) exceeds the historical reasonable interval (20%) by 30%, and the recognition rate of such samples is lower than 80% on average; physical law: ambiguity of 5px has approached the limit of printing equipment, and continuous improvement may lead to unreadable characters; model training feedback: excessive high ambiguity samples may lead to decreased generalization ability of the model for normal samples. The adjustment basis of sample proportion is historical statistics: when the proportion of high defect samples (such as combined defects of ambiguity and inclination) is 20%, the standard deviation of model recognition rate is the smallest (best stability); risk balance: although the current proportion of high defect samples is 30%, it improves the noise resistance of the model, but leads to insufficient training of normal samples (recognition rate decreases by 3%).

[0102] The model configuration optimization is analyzed as follows: a new sample filtering rule is added, and the rule logic is filtering condition: font size < 2.1 mm (violating the latest international standard); execution timing: automatically triggered after sample generation and before storage; historical effect: after filtering samples with font size < 2.0 mm, the model compliance-related indicators improved by 15%. The loss function weight is adjusted, and the weight distribution is as follows: original weight: quality deviation coefficient 60% + compliance index 30% + other 10%; new weight: quality deviation coefficient 60% → 50%, compliance index 30% → 40% (because the importance of compliance has increased after the standard update); on the test set, this adjustment improves the recognition rate of compliant samples from 88% to 91%. The calculation formula is new loss function = 0.5 × quality deviation loss + 0.4 × compliance loss + 0.1 × other loss.

[0103] After adjustment, 1000 samples are generated, and the proportion of samples with font size ≥ 2.1 mm is 100%, and the average recognition rate is improved from 82% to 87%. Every 1000 samples generated, automatically statistics the execution effect of optimization instructions: if the compliance index is less than 0.95, trigger secondary optimization (such as further increasing the recommended value of font size to 2.8 mm); if the recognition rate improvement is less than 5%, re-evaluate the loss function weight distribution.

[0104] Ensure that sample generation always meets the latest international standards (such as ICAO may update printing parameter requirements annually); avoid compliance risks due to standard lag (such as samples being rejected by customs due to non-compliance). The optimized parameter combination shortens the model training period by 20% (due to the reduction of invalid samples); the standard deviation of recognition rate is reduced from 8% to 5%, and the model generalization ability is significantly improved. Reduce the cost of sample regeneration due to unreasonable parameters (estimated to save 30% of computing resources annually); reduce the cost of manual adjustment of compliance audit (automatic filtering and weight adjustment replace 50% of manual work).

[0105] The optimization instruction generation realizes intelligent iteration of the MRZ synthetic sample generation process by deeply integrating historical data, risk early warning and standard updating. The mechanism not only ensures that the sample strictly conforms to the international standard, but also continuously improves the model training effect in a data-driven manner. In actual application, combined with the dynamic verification and feedback mechanism, a closed loop of "problem identification-solution generation-effect verification-continuous optimization" can be formed, so that the sample generation system always maintains at the industry best practice level, effectively solving the problems of parameter adjustment lag and model configuration rigidity in the traditional method.

[0106] The application solves the problems of insufficient real samples and distorted synthetic samples through multi-dimensional parameter simulation and intelligent optimization. First, the pre-processed MRZ corpus, synthetic parameters and international standards are obtained. The corpus is divided into segments with a length of 5-14. The parameters include fuzziness, font size, etc. The international standard specifies the format and printing quality. Then, the corpus and parameters are combined to simulate image quality changes, etc. According to the international standard and real data, parameter association rules are established to generate parameter combination constraint conditions containing hard, collaborative and semantic constraints. Then, based on parameter combination, corpus preprocessing, sample proportion control and constraint conditions, synthetic samples are generated, and parameter compliance is verified. After generating the samples, quality changes and other features are extracted, and the quality change law is determined by comparing different parameter combinations. The deviation degree is evaluated by comparing with the standard mode library, and the evaluation feature parameter set containing the quality deviation coefficient is generated by combining the historical data weighting processing.

[0107] The deviation of the parameters and the international standard is calculated, and the invalid deviation is filtered to generate compliance data and feature vectors. Finally, based on the intelligent model, the parameter combination, sample proportion, evaluation result and compliance feature vector are combined to analyze the risk level, compare the quality threshold to generate abnormal judgment and risk early warning, and generate optimization instructions by combining historical data and standard updates, such as adjusting font size, fuzziness and other parameters, optimizing sample proportion and model configuration. The synthetic samples generated by the scheme conform to the international standard, improve the model recognition rate by 7%, save 30% of computing resources, and provide high-quality data support for MRZ recognition model training, which is suitable for certificate recognition scenarios.

[0108] In one embodiment, as shown in Figure 2 The application also provides a MRZ synthetic sample generation device, which comprises:

[0109] The acquisition module 201 is configured to acquire a machine-readable zone synthetic sample and an international standard. The machine-readable zone synthetic sample comprises a pre-processed machine-readable zone corpus, specific synthetic parameters and a proportion control of the synthetic sample. The synthetic parameters include fuzziness, font size, character spacing, font thickness and inclination angle. The corpus preprocessing is a random segmentation of the machine-readable zone corpus with a length range of [5, 14]. The international standard is used to standardize the format, character set and printing quality of the machine-readable zone of the travel certificate.

[0110] The processing module 202 is configured to combine the pre-processed machine-readable area corpus and preset synthesis parameters, simulate quality changes, information truncation and physical defects of the machine-readable area image, establish association rules between parameters based on international standards and real MRZ image data, and generate parameter combination constraint conditions; process based on the synthesis parameter combination, corpus preprocessing, sample proportion control and parameter combination constraint conditions to generate machine-readable area synthesis samples; process the generated machine-readable area synthesis samples to generate sample-related evaluation results; process the evaluation results and parameter settings in the synthesis parameters in combination with international standards to generate machine-readable area synthesis sample generation bases; process based on the intelligent model in combination with the parameter combination, sample proportion control, sample-related evaluation results and parameter compliance feature vectors to generate machine-readable area synthesis sample generation optimization results.

[0111] The computer-readable storage medium provided by the above embodiments of the present application has the same beneficial effects as the method adopted, run or implemented by the application program stored therein, based on the same inventive concept as the MRZ synthesis sample generation method provided by the embodiments of the present application.

[0112] Each of the embodiments in the present application is described in a related manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the difference from other embodiments. In particular, for the evaluation MRZ synthesis sample generation method, electronic device, electronic equipment and readable storage medium embodiments, since they are basically similar to the above-mentioned embodiments of the MRZ synthesis sample generation method, the description is relatively simple, and the relevant parts can be referred to the above-mentioned embodiments of the MRZ synthesis sample generation method.

Claims

1. A method of generating a MRZ synthetic sample, characterized by, The method comprises the following steps: obtaining a machine-readable area synthetic sample and an international standard, the machine-readable area synthetic sample comprising using a preprocessed machine-readable area corpus, specific synthetic parameters, and controlling the proportion of the synthetic sample, wherein the synthetic parameters comprise fuzziness, font size, interword spacing, font thickness, and tilt angle, and the corpus preprocessing comprises randomly cutting the machine-readable area corpus in a length range [5, 14]; the international standard is used to standardize the format, character set, and printing quality of a travel document machine-readable area; combining the preprocessed machine-readable area corpus and the preset synthetic parameters to simulate the quality change, information truncation, and physical defects of the machine-readable area image, establishing the correlation rules between the parameters based on the international standard and real MRZ image data, and generating parameter combination constraint conditions; processing based on the synthetic parameter combination, corpus preprocessing, sample proportion control, and parameter combination constraint conditions to generate a machine-readable area synthetic sample; processing the generated machine-readable area synthetic sample to generate sample-related evaluation results; processing the evaluation results and parameter settings in the synthetic parameters in combination with the international standard to generate machine-readable area synthetic sample generation bases; processing based on an intelligent model in combination with the parameter combination, sample proportion control, sample-related evaluation results, and parameter compliance feature vectors to generate machine-readable area synthetic sample generation optimization results.

2. The method of claim 1, wherein, combining the preprocessed machine-readable area corpus and the preset synthetic parameters to simulate the quality change, information truncation, and physical defects of the machine-readable area image, establishing the correlation rules between the parameters based on the international standard and real MRZ image data, and generating parameter combination constraint conditions, comprising: performing correlation degree combination analysis on the preprocessed machine-readable area corpus and the preset synthetic parameters to generate simulation correlation degree analysis results of the quality change, information truncation, and physical defects, wherein the machine-readable area corpus comprises randomly cut data based on semantic unit recognition, and the specific synthetic parameters comprise fuzziness, font size, interword spacing, font thickness, and tilt angle parameter data simulating image quality change; performing screening processing on the correlation degree analysis results to generate a target simulation variable list, the target simulation variable list comprising fuzziness parameters, font size parameters, interword spacing parameters, font thickness parameters, tilt angle parameters, and corpus random cutting parameters; performing screening processing on the correlation degree analysis results to generate parameter combination constraint conditions, including hard constraints, collaborative constraints, and semantic constraints that conform to the international standard; based on the parameter combination constraint conditions, generating multi-dimensional simulation input variables, wherein the multi-dimensional simulation input variables are used to represent key features associated with the quality change, information truncation, and physical defects of the machine-readable area image.

3. The method of claim 1, wherein, processing based on the synthetic parameter combination, corpus preprocessing, sample proportion control, and parameter combination constraint conditions to generate a machine-readable area synthetic sample, comprising: The feature extraction processing is performed on the synthetic parameter combination, corpus preprocessing and sample proportion control to generate the fuzziness feature, font size feature, interword space feature, font thickness feature, tilt angle feature, corpus random segmentation feature and sample proportion control feature. The synthetic parameter combination feature includes specific combination parameter data of the fuzziness, font size, interword space, font thickness and tilt angle conforming to the international standard. The sample proportion control feature includes sample proportion setting data of different parameter combinations. The feature extraction processing is performed on the generation process of the generated machine-readable area synthetic sample to generate the parameter combination simulation feature, corpus preprocessing simulation feature and sample proportion control feature. The parameter combination simulation feature includes simulation parameter combination data of the image quality change and physical defect. The corpus preprocessing simulation feature includes random segmentation data of the information truncation corpus. The sample proportion control feature includes proportion allocation data of different parameter combination samples. Based on the parameter combination constraint condition, the analysis processing is performed on the synthetic parameter combination feature, corpus preprocessing feature, sample proportion control feature, parameter combination simulation feature, corpus preprocessing simulation feature and sample proportion control feature. When the sample is generated, it is verified whether the parameter combination conforms to the international standard to generate the machine-readable area synthetic sample. The machine-readable area synthetic sample is used to represent the simulation sample feature of the MRZ image quality change, information truncation and physical defect to form the MRZ synthetic sample generation result containing the parameter combination, corpus preprocessing and sample proportion control.

4. The method of claim 3, wherein, The generated machine-readable area synthetic sample is processed to generate the sample related evaluation result, including: Based on the synthetic parameter combination feature, corpus preprocessing feature and sample proportion control feature, the quality change, information truncation and physical defect simulation feature of the generated machine-readable area synthetic sample are extracted to generate the sample feature vector set. The sample feature vector sets under different parameter combinations are compared to determine the law of the machine-readable area synthetic sample quality change with the parameter change to generate the sample quality change evaluation information. The sample quality change evaluation information is compared with the preset normal sample mode library and abnormal sample mode library containing the international standard sample to generate the sample quality deviation evaluation result. The sample quality deviation evaluation result is weighted processed in combination with the historical evaluation data of the machine-readable area synthetic sample and the rationality of the parameter combination to generate the sample evaluation risk feature vector. The sample quality deviation evaluation result and the sample evaluation risk feature vector are normalized and comprehensively operated to generate the sample evaluation feature parameter set containing the quality deviation coefficient, sample effectiveness index and parameter compliance index. The sample evaluation feature parameter set is composed of the sample feature vector set, sample quality change evaluation information, sample quality deviation evaluation result, sample evaluation risk feature vector, quality deviation coefficient and sample effectiveness index.

5. The method of claim 1, wherein, The evaluation result and parameter setting in the synthetic parameter are processed in combination with the international standard to generate the machine-readable area synthetic sample generation basis, including: Based on the parameter evaluation results and the international standard, the deviation values of the actual values of the fuzziness, font size, character spacing, font thickness, and tilt angle in the synthetic parameters from the international standard are calculated, and parameter deviation data is generated; The parameter deviation data is filtered for effectiveness to generate filtered parameter deviation data; Based on the filtered parameter deviation data, the degree to which the synthetic parameters comply with the international standard is calculated, and parameter compliance data is generated; The parameter compliance data is weighted in combination with the historical evaluation data of the synthetic parameters and the update frequency of the international standard to generate a parameter compliance feature vector.

6. The method of claim 4, wherein, Based on the intelligent model, the parameter combination, the sample proportion control, the sample correlation evaluation results, and the parameter compliance feature vector are processed to generate machine-readable zone synthetic sample generation optimization results, including: The parameter combination, the sample proportion control, and the sample correlation evaluation results are analyzed for risk level to generate parameter combination risk level evaluation results; Based on the quality risk features in the sample correlation evaluation results and the preset quality threshold, abnormal feature comparison processing is performed to generate quality abnormality determination results; The parameter combination risk level evaluation results and the quality abnormality determination results are processed to generate sample generation risk warning information; The machine-readable zone synthetic sample historical generation information, the intelligent model configuration information, and the international standard update record are obtained; Based on the synthetic sample historical generation information, the sample generation risk warning information, and the international standard update record, the intelligent model configuration information is processed to generate machine-readable zone synthetic sample generation optimization instructions.

7. An apparatus for generating a MRZ synthetic sample, characterized by The device includes: An acquisition module is configured to acquire a machine-readable zone synthetic sample and an international standard. The machine-readable zone synthetic sample includes a preprocessed machine-readable zone corpus, specific synthetic parameters, and control of the proportion of synthetic samples. The synthetic parameters include fuzziness, font size, character spacing, font thickness, and tilt angle. The corpus preprocessing is random segmentation of the machine-readable zone corpus within a length range of [5, 14]. The international standard is used to standardize the format, character set, and printing quality of travel documents. A processing module is configured to combine the preprocessed machine-readable zone corpus and the preset synthetic parameters, simulate the quality changes, information truncation, and physical defects of the machine-readable zone image, establish association rules between parameters based on the international standard and real MRZ image data, and generate parameter combination constraint conditions. The processing module is further configured to process the synthetic parameter combination, the corpus preprocessing, the sample proportion control, and the parameter combination constraint conditions to generate a machine-readable zone synthetic sample. The processing module is further configured to process the generated machine-readable zone synthetic sample to generate sample correlation evaluation results. The processing module is further configured to process the evaluation results in the synthetic parameters and the parameter settings based on the international standard to generate machine-readable zone synthetic sample generation basis. The processing module is further configured to process the parameter combination, the sample proportion control, the sample correlation evaluation results, and the parameter compliance feature vector based on the intelligent model to generate machine-readable zone synthetic sample generation optimization results.

8. An electronic device, comprising: It includes: A first processor; and a memory for storing executable instructions of the first processor; The first processor is configured to execute the executable instructions to perform the method for generating the MRZ synthetic sample according to any one of claims 1-6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the second processor to implement the method for generating the MRZ synthetic sample according to any one of claims 1-6.

Images