A method and system for improving the accuracy of OCR recognition models in recognizing specific characters.

Abstract

This invention provides a method and system for improving the accuracy of OCR recognition models in recognizing specific characters, applied in the field of data processing technology. This application processes font type information in training samples, identifies and removes non-target font samples, introduces target font samples, and generates filtered training samples; it processes character confusion-related data in specific country and region codes based on a specific country and region code analysis model to generate character confusion optimization information; it processes the filtered training samples to generate model training results; it processes the model training results, country and region codes, and character types in the character confusion-related data in specific country and region codes, combined with character confusion analysis, to generate character recognition optimization factors; and it processes the model training results, character recognition optimization factors, and character confusion optimization information based on a target optimization model combined with a targeted optimization strategy to generate character recognition optimization results.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for improving the accuracy of OCR recognition models in recognizing specific characters. Background Technology

[0002] In the field of modern information processing, Optical Character Recognition (OCR) technology is crucial, as it can convert text in images into editable text and is widely used in the automatic recognition of critical documents such as ID cards, bank cards, and passports. However, in complex scenarios such as passport OCR recognition, this technology still faces many challenges.

[0003] From an image quality perspective, passport images often suffer from blurriness, tilt, and noise due to poor shooting environment (lighting, angle) or scanning quality issues, severely impacting recognition performance. Traditional OCR methods (such as Tesseract) fall far short of practical requirements when handling passport images in such complex scenarios. Even with deep learning-based models improving accuracy, generalization performance remains poor when dealing with low-quality images. Regarding multilingual processing, passport information may contain multiple languages, such as English, Chinese, and Arabic, requiring OCR models with strong multilingual recognition capabilities. However, existing models have significant shortcomings in this area, and their ability to recognize mixed multilingual text urgently needs improvement. From a structured information extraction perspective, key information such as name, passport number, and nationality in the passport needs accurate location and recognition. However, existing models have considerable room for improvement in this area. For example, when recognizing passport numbers, errors are prone to occur due to font and character adhesion issues.

[0004] Furthermore, the diversity and complexity of text in different scenarios, such as varying font styles and significant differences in lighting conditions, pose challenges to traditional OCR systems, resulting in low recognition accuracy and poor robustness. Although some researchers have used adaptive deep learning algorithms such as transfer learning and incremental learning to improve performance, problems still exist in areas such as domain adaptability.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this application is to provide a method and system for improving the accuracy of OCR recognition models in recognizing specific characters, at least to some extent overcoming the problems existing in the prior art. This is achieved through data filtering, country code-oriented optimization, and multi-dimensional feature fusion to improve the recognition accuracy of OCR models for specific characters such as O / 0. CNN and SVM are used to filter OCR-B font training samples, reducing confusion error by 30%. LSTM and CNN are used to parse country codes, and random forest and PCA are combined to mine confusion patterns, building a graph structure optimization framework. A rule engine generates optimization nodes such as weight adjustments, and finally, the target optimization model is used to fuse and generate the recognition result, increasing the O / 0 recognition accuracy in the US MRZ code from 75% to 92%. This method achieves targeted optimization through a closed loop of "data filtering - pattern analysis - framework construction - optimization execution," improving robustness by 20% compared to traditional methods, and is applicable to global MRZ code scenarios.

[0007] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0008] According to one aspect of this application, a method for improving the accuracy of an OCR recognition model in recognizing specific characters is provided, comprising: acquiring font type information in training samples, including target fonts and non-target fonts, as well as character confusion-related data in specific country and region codes, wherein the character confusion-related data includes the confusion probability of O and 0, and the target font includes optical character recognition (OCR) font B; processing the font type information in the training samples, identifying and eliminating non-target font samples, and introducing target font samples to generate filtered training samples; processing the character confusion-related data in specific country and region codes based on a specific country and region code analysis model, and combining it with target font recognition analysis to generate character confusion optimization information; processing the filtered training samples, retraining the OCR recognition model, and generating model training results; processing the model training results, country and region codes, and character types in the character confusion-related data in specific country and region codes, combined with character confusion analysis, to generate character recognition optimization factors; and processing the model training results, character recognition optimization factors, and character confusion optimization information based on a target optimization model and a targeted optimization strategy to generate character recognition optimization results.

[0009] Another aspect of this application discloses an apparatus for improving the accuracy of OCR recognition models in recognizing specific characters, characterized by comprising: an acquisition module for acquiring font type information in training samples, including target fonts and non-target fonts, as well as character confusion-related data in specific country and region codes, wherein the character confusion-related data includes the confusion probability between O and 0, and the target font includes optical character recognition (OCR) font B; a processing module for processing the font type information in the training samples, identifying and eliminating non-target font samples, introducing target font samples, and generating filtered training samples; processing the character confusion-related data in specific country and region codes based on a specific country and region code analysis model, and combining it with target font recognition analysis to generate character confusion optimization information; processing the filtered training samples, retraining the OCR recognition model, and generating model training results; processing the model training results, country and region codes, and character types in the character confusion-related data in specific country and region codes, combined with character confusion analysis, to generate character recognition optimization factors; and processing the model training results, character recognition optimization factors, and character confusion optimization information based on a target optimization model and a directional optimization strategy to generate character recognition optimization results.

[0010] According to another aspect of this application, an electronic device is characterized by comprising: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the above-described method for improving the accuracy of OCR recognition model for recognizing specific characters by executing the executable instructions.

[0011] According to another aspect of this application, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a second processor, implements the above-described method for improving the accuracy of OCR recognition models in recognizing specific characters.

[0012] This application provides a method and system for improving the accuracy of OCR recognition models for specific characters. First, it combines CNN and SVM to filter training samples, eliminating non-target fonts (such as Arial) and retaining OCR-B font samples, thus improving data purity and reducing subsequent model training confusion errors by 30%. Second, it uses LSTM and CNN to parse country codes and label confused data (such as "USA-O0"), combining random forest and PCA to mine confusion patterns (such as high confusion probability for the first O / O), and builds a graph structure optimization framework. Then, it integrates code-driven confusion analysis through a rule engine to generate optimized nodes including weight adjustment and image sharpening. Finally, based on the target optimization model, it fuses training results and optimization factors to generate recognition results including accuracy and optimization probability, increasing the O / O recognition accuracy in US MRZ codes from 75% to 92%. This method achieves targeted optimization of easily confused characters in specific country codes through a closed-loop process of "data filtering - pattern analysis - framework construction - optimization execution," improving robustness by 20% compared to traditional methods and is applicable to global MRZ code scenarios.

[0013] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0014] Figure 1 This invention provides a flowchart illustrating a method for improving the accuracy of OCR recognition models in recognizing specific characters, according to an embodiment of this application.

[0015] Figure 2 This illustration shows a schematic diagram of a device for improving the accuracy of OCR recognition models in recognizing specific characters, according to an embodiment of this application. Detailed Implementation

[0016] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0017] The following is combined with Figure 1 This application describes a method for improving the accuracy of OCR recognition models in recognizing specific characters according to exemplary embodiments of the present application. It should be noted that the following application scenarios are shown only to facilitate understanding of the spirit and principles of the present application, and the embodiments of the present application are not limited in any way. Rather, the embodiments of the present application are applicable to any suitable scenario.

[0018] In one embodiment, this application also proposes a method and system for improving the accuracy of OCR recognition models in recognizing specific characters. Figure 1The diagram illustrates a flowchart of a method for improving the accuracy of OCR recognition models in recognizing specific characters according to an embodiment of this application.

[0019] S101, Obtain font type information from the training samples.

[0020] In one implementation, font type data is extracted from document images containing multiple fonts (such as passport and visa MRZ code images) by scanning or importing such images. For example, from a passport MRZ code image of a certain country, fonts are identified as including OCR-B font (target font) and Arial font (non-target font). The target font is explicitly defined as OCR-B font, and the non-target font is other font types. For instance, in the training samples, some samples are identified as using OCR-B font (target font), while others use Times New Roman font (non-target font).

[0021] Collect confusion data on the characters "O" and "0" in different country and region codes. For example, it was found that the confusion probability of "O" and "0" in the ID card numbers of one country is 30%, while the confusion probability in the MRZ code of another country is 25%. In the character confusion-related data of specific country and region codes, this is specifically reflected in the probability value of "O" and "0" being misidentified. For example, in a certain region code, the probability of "O" being misidentified as "0" is 22%, and the probability of "0" being misidentified as "O" is 18%. In the font type information of the training samples, the OCR-B font is identified as the target font. For example, when processing OCR recognition training samples of airport entry and exit documents, the OCR-B font is selected as the primary target font for recognition.

[0022] S102, process the font type information in the training samples, identify and remove non-target font samples, introduce target font samples, and generate the filtered training samples.

[0023] In one implementation, a font type information extraction process is used to perform feature analysis and type labeling on the training sample data. Target font sample selection rules and non-target font removal algorithms are introduced to achieve type purification and target focus of the sample data. By combining a Convolutional Neural Network (CNN) and a Support Vector Machine (SVM), font feature analysis and classification selection of passport MRZ code image training samples are achieved. The core objective is to retain target font (OCR-B) samples and remove non-target font (such as Arial) samples, providing a clean dataset for subsequent OCR model training. A three-layer Convolutional Neural Network (CNN) is used, with the following structure: Input layer: Receives the pixel matrix (e.g., 224×224×3) of the passport MRZ code image, supporting RGB or grayscale image input. Convolutional layer 1: Uses 32 3×3 convolutional kernels with a stride of 1, padding = 1, and ReLU activation function to extract basic features of font edges and contours. Pooling layer 1: 2×2 max pooling to reduce dimensionality and retain key features. Convolutional Layer 2: 64 3×3 convolutional kernels further extract detailed features such as stroke thickness and corner curvature. Pooling Layer 2: 2×2 max pooling to compress feature dimensions. Fully Connected Layer: 128 nodes, mapping convolutional features to font feature vectors.

[0024] The following is an example of feature extraction: For the character "O" in the MRZ code of a passport from a certain country, the CNN uses convolution operations to identify its outline as a circular closed curve (OCR-B font feature), or to discover that the stroke edges are smooth and sans-serif (Arial font feature), and generates corresponding feature vectors (including parameters such as outline curvature and stroke width distribution). Parameter configuration: Learning rate: 0.001, using the Adam optimizer. Number of iterations: 50 rounds, batch size = 32. Training data: Contains 100,000 MRZ code image samples with font types labeled. A linear kernel support vector machine (SVM) is used to map the feature vectors extracted by the CNN to a high-dimensional space, and a hyperplane is used to separate the OCR-B font and Arial font samples. Classification logic and parameters: The input is the 128-dimensional font feature vector output by the CNN. The decision function is f(x) = ω. T x+b, where ω is the weight vector and b is the bias. The regularization parameter, C=10, balances classification accuracy and generalization ability. The kernel function is a linear kernel, suitable for high-dimensional separable features extracted by CNNs.

[0025] The annotation example is as follows: For a character feature vector extracted by CNN, SVM calculates its distance to two types of hyperplanes: If it is closer to the OCR-B hyperplane (e.g., the contour curvature parameter is greater than the threshold of 0.7), it is labeled as "target font". If it is closer to the Arial hyperplane (e.g., the stroke width uniformity parameter is less than the threshold of 0.3), it is labeled as "non-target font".

[0026] Prioritize retaining OCR-B font samples with an SVM classification confidence score ≥ 0.8. For samples with a confidence score between 0.5 and 0.8, initiate secondary verification (such as manual review or cross-validation). Remove all samples labeled as Arial or other non-target fonts. The execution process is as follows: Input 1000 MRZ code image training samples. After CNN feature extraction, the SVM annotation results are: 700 OCR-B images (confidence score ≥ 0.8), 200 OCR-B images (confidence score 0.5-0.8), and 100 Arial images. Retain the 700 high-confidence OCR-B samples. Perform secondary feature comparison on the 200 medium-confidence samples (such as calculating cosine similarity with standard feature vectors in the OCR-B font database), ultimately retaining 150 images. Remove all 100 Arial samples, generating a filtered target font training sample set of 850 images.

[0027] By combining CNN and SVM, font classification accuracy can reach over 95%, an improvement of approximately 15% compared to a single model (SVM alone achieves approximately 80% accuracy). Data purity is ensured: after removing non-target font samples, the confusion error during subsequent OCR model training is reduced by 30%, especially significantly improving the recognition of easily confused characters such as "O" and "0". Generalization ability is enhanced: the filtered samples focus on OCR-B font features, improving the model's robustness in different MRZ code scenarios across various countries (e.g., handling differences in scanning resolution and image noise) by 20%.

[0028] By connecting a font feature database with a confusion probability analysis model, a font type matching matrix is ​​constructed. The sample validity weights are calculated using a target font feature mapping and transformation model, establishing a dynamic font type selection mechanism. The confusion probability analysis model (e.g., logistic regression model) and the target font feature mapping and transformation model (e.g., neural network model) are used. The logistic regression model includes an input layer (character feature vector) and an output layer (confusion probability value); the neural network model includes an input layer (target font feature vector), a hidden layer (1 layer, 50 nodes), and an output layer (validity weight value). The logistic regression regularization coefficient λ = 0.1; the neural network learning rate is 0.01, and the activation function is ReLU.

[0029] By connecting to the OCR-B font feature database and using a logistic regression model to analyze that the confusion probability of "O" and "0" in a certain country's MRZ code is 25%, a font type matching matrix is ​​constructed. The training sample features are compared with the OCR-B font features through a neural network model, and the sample validity weight (e.g., 0.85) is calculated to dynamically select high-weight samples.

[0030] Using font type as the classification dimension, a multi-dimensional font feature matrix is ​​constructed by fusing target font feature data, non-target font feature data, and character confusion probability data to achieve font dimension filtering in the training sample data. The multi-dimensional feature fusion model (such as an autoencoder model) consists of an input layer (a concatenated vector of font features and confusion probability data), two hidden layers (100 and 50 nodes respectively), and an output layer (reconstructing the feature vector). The autoencoder has a learning rate of 0.005 and a sparsity regularization parameter of 0.01.

[0031] Classified by font type (OCR-B / non-OCR-B), a multi-dimensional font feature matrix is ​​constructed by fusing OCR-B font feature vectors, non-OCR-B font feature vectors, and character confusion probability (e.g., 22%). Dimensionality reduction and filtering are performed using an autoencoder to retain samples with high similarity to OCR-B font features.

[0032] Missing font information is processed through a sample feature completion model, and key features are enhanced by a target font feature extraction mechanism. These features are then fused with the font type matching matrix features to generate training sample selection results. The model consists of a sample feature completion model (e.g., Generative Adversarial Network (GAN)) and a target font feature extraction mechanism (e.g., an attention mechanism model). The GAN includes a generator (multi-layer fully connected network) and a discriminator (multi-layer convolutional network); the attention mechanism model adds an attention layer to the CNN. The GAN generator has a learning rate of 0.0002, and the discriminator has a learning rate of 0.0001; the attention mechanism weight coefficient α = 0.5.

[0033] For training samples lacking font information, GAN is used to complete font features (such as missing strokes), and attention mechanism model is used to enhance key features of OCR-B fonts (such as vertical line thickness). These features are then fused with font type matching matrix features to select samples that meet the OCR-B font standard.

[0034] S103 processes character obfuscation-related data in specific country and region codes based on a specific country and region code analysis model, and generates character obfuscation optimization information by combining target font recognition analysis.

[0035] In one implementation, a specific country / region code analysis model and target font recognition processing are combined to perform real-time parsing and extraction of code information in character confusion-related data, generating code-tagged confusion data. Through the collaboration of a Long Short-Term Memory (LSTM) network and a Convolutional Neural Network (CNN), semantic parsing and font recognition of country codes in MRZ codes are achieved, thereby extracting confusion data of specific characters (such as "O" and "0") and generating a code-tagged confusion dataset, providing a foundation for subsequent country / region-specific optimization. The country code parsing process of the LSTM model is as follows: a two-layer LSTM network is used, with the following structure: Input layer: receives the character sequence of the country code (e.g., "USA" converted to an ASCII encoding sequence or word embedding vector), with an input dimension of sequence length × feature dimension (e.g., 3 × 128). LSTM layer 1: 128 memory units, capturing long-distance dependencies between characters (e.g., spelling rules of country codes). LSTM layer 2: 128 memory units, further refining the semantic features of the code (e.g., "USA" corresponds to the country identifier of the United States). Fully connected layer: The output dimension is the number of country code categories, and the code classification is completed through the softmax function.

[0036] The input is the character sequence "USA<" (< being a padding character) from the MRZ code of a passport. The LSTM uses memory units to track the character order, recognizes the combination pattern of "U", "S", and "A", matches it to a pre-trained country code dictionary, and outputs the parsed result "USA", labeled as the country code for the United States. Learning rate: 0.001, using the Adam optimizer. Iterations: 100 epochs, batch size = 64. Training data: Contains MRZ code samples from 200 countries / regions, each code corresponding to a unique label.

[0037] The font recognition process of the CNN model is as follows, employing a three-layer CNN architecture (same as the CNN model used for sample selection mentioned earlier), focusing on extracting the visual features of characters: Input layer: Image slices of single characters in the country code (e.g., a 28×28 pixel matrix for "O"). Convolutional layer: 32 3×3 convolutional kernels are used to extract contour features, and 64 3×3 convolutional kernels are used to extract stroke details. Pooling layer: 2×2 max pooling is used for dimensionality reduction, preserving key features (e.g., the thickness of vertical lines and the roundness of corners in OCR-B fonts). Fully connected layer: The output is the font type probability distribution (OCR-B / non-OCR-B).

[0038] For the character "O" in the country code "USA", the CNN extracts its outline as a circle with specific thickness variations at the edges (OCR-B font features). This is compared with standard feature vectors in the OCR-B font database (e.g., outline curvature 0.8, stroke width standard deviation 0.2), outputting an OCR-B font confidence score of 0.92. Convolution kernel size: 3×3, stride 1, padding = 1. Activation function: ReLU, pooling kernel size: 2×2. The confusion data extraction logic is to extract the corresponding character region (e.g., the position of "O" and adjacent "0") from the MRZ code image based on the country code position parsed by LSTM. This is then connected to a pre-collected country code confusion probability database (e.g., the probability of "O" being misidentified as "0" in the US MRZ code is 25%), combined with the CNN font recognition results, to determine the confusion type (e.g., O→0 or 0→O). The format is {country code}-{confusion character pair}, such as "USA-O0". The data includes confusion probability (e.g., 25%), font type (OCR-B), and positional information (e.g., the 3rd digit of the MRZ code). LSTM parses the country code as "USA," and CNN identifies the character "O" as an OCR-B font, with adjacent "0" characters belonging to the same font. Checking the US MRZ code confusion database, the mutual confusion probabilities of "O" and "0" are found to be: O→0 (25%) and 0→O (18%), respectively. The generated labeled data is {"code":"USA","confusion":"O0","probability":{"O→0":0.25,"0→O":0.18},"font":"OCR-B"}.

[0039] Semantic and visual features are fused, with LSTM ensuring accuracy in country code parsing (accuracy ≥ 98%) and CNN guaranteeing reliability in font recognition (OCR-B recognition rate ≥ 95%). The combination of these two approaches results in an error rate of less than 3% for obfuscated data labeling. Obfuscated data with country codes can accurately pinpoint problem scenarios (such as O0 obfuscation in the US MRZ code), providing a basis for subsequent adjustments to training sample weights and optimization of model parameters based on country / region. The model supports rapid adaptation to new country codes (requiring only dictionary updates and minor sample fine-tuning), making it suitable for MRZ code processing scenarios in different regions globally.

[0040] Combining character obfuscation analysis and feature extraction strategies, the obfuscated data marked with codes is processed to perform obfuscation pattern recognition and generate preliminary obfuscation analysis data. Obfuscation analysis models (such as random forest) and feature extraction models (such as principal component analysis, PCA) are used. The random forest contains multiple decision trees (100 trees); the PCA model is used for dimensionality reduction. The maximum depth of the random forest is 5, and PCA reduces the dimensionality to 10 dimensions. The obfuscated data marked "USA-O0" includes the following basic features: country code (USA), character pair (O0), and obfuscation probability (O→0 = 25%, 0→O = 18%). Environmental features: character position (e.g., the 3rd digit of the MRZ code), font type (OCR-B), and image quality parameters (resolution, noise level). Contextual features: adjacent characters (e.g., "A" on the left, "1" on the right), and the field type of the character (passport number / visa number). The "USA-O0" data is converted into a numerical feature vector: one-hot encoding: country code "USA" → [1,0,0,...] (200-dimensional country dictionary). Positional characteristics: MRZ code 3rd bit → [0,0,1,0,...] (10-dimensional position vector). Confusion probability: O→0=0.25, 0→O=0.18 → concatenated to [0.25,0.18].

[0041] A linear dimensionality reduction model maps high-dimensional features to a low-dimensional subspace through orthogonal transformations, preserving the principal variances. Target dimensionality: 10 dimensions (original feature dimensions approximately 50). Variance preservation threshold: 95% (cumulative variance contribution rate of the first 10 principal components ≥ 95%). Feature standardization (mean 0, variance 1). The covariance matrix is ​​calculated, and eigenvalues ​​and eigenvectors are solved. The first 10 principal components are selected based on eigenvalue sorting.

[0042] The following is an example of dimensionality reduction. For the 50-dimensional feature vector of "USA-O0", PCA identifies the top three principal components as follows: Principal Component 1 (variance contribution rate 30%): linear combination of character position and font features (e.g., the first 5 digits of the MRZ code + OCR-B font weights). Principal Component 2 (variance contribution rate 25%): correlation between confusion probability and adjacent characters (e.g., the association between O0 confusion and the digits on the right). Principal Component 3 (variance contribution rate 20%): coupling features between the country code and image quality (e.g., the generally high scanning resolution of the US MRZ code). The 10-dimensional vector after dimensionality reduction retains 95% of the key information of the original data.

[0043] An ensemble learning model, consisting of 100 decision trees, employs a Bagging strategy to reduce variance. Number of decision trees: 100. Maximum depth: 5 (to limit tree complexity and avoid overfitting). Feature sampling rate: 0.8 (80% of features are randomly selected for splitting in each tree). Node splitting criterion: Gini impurity. For the 10-dimensional features after PCA dimensionality reduction, each tree randomly selects samples and features for splitting. Each tree performs a binary classification on whether it represents a "high-confusion pattern," and the final result is determined by voting.

[0044] For the "USA-O0" data, Random Forest identified the following key confusion patterns: Pattern 1: When O / 0 is at the beginning of an MRZ code number (position ≤ 3), the confusion probability increases by 40% (e.g., when the first digit of a US passport number is O, it is easily misidentified as 0). Pattern 2: In OCR-B fonts, when the width of the vertical line of a character is < 1.5 pixels, the confusion probability of O0 increases by 25% (insufficient scanning resolution causes the strokes to become thinner). Pattern 3: When adjacent characters are letters (e.g., "O" followed by "A"), the confusion probability is 15% lower than when followed by numbers (the visual difference between letters and numbers is greater).

[0045] The generated preliminary analysis data includes the following: Pattern description: Confusion patterns are recorded using natural language and structured fields (e.g., "First digit + vertical line width < 1.5 pixels → Confusion probability increased by 40%"). Feature importance: Feature weights calculated by random forest (e.g., position feature importance 0.35, font feature importance 0.25). Probability distribution: Confusion probability ranges for different patterns (e.g., the probability of O→0 for pattern 1 is 25%–45%).

[0046] Based on country and region code nodes, character obfuscation information, and obfuscation optimization rule generation rules, the preliminary obfuscation analysis data is structured to build an obfuscation optimization framework. The framework model is shown in the graph database model. The graph database contains country code nodes, obfuscation information nodes, rule nodes, and the edges between them. Node attributes include code, probability, and rule, while edge weights represent the association strength. Using "USA" as the country code node, it connects to the obfuscation information node "O0 obfuscation probability 30%", and generates associations between nodes according to optimization rules, thus building a structured obfuscation optimization framework.

[0047] Optimized nodes are automatically generated based on the target font features and the associated / obfuscated logical relationships. Anomalies are flagged and logically adjusted based on a node association detection mechanism. The optimization includes a node generation model (e.g., the Apriori association rule model) and an anomaly detection model (e.g., an isolation forest). The Apriori model mines association rules; the isolation forest is used to detect anomalies. Apriori has a minimum support of 0.2 and a minimum confidence of 0.8; the isolation forest has 100 trees. Based on OCR-B font features, "O0 obfuscation-font features" optimized nodes are generated using Apriori. Anomalies between a country code node and the obfuscated node are detected and adjusted using the isolation forest.

[0048] According to the preset optimization mode generation rules, code-driven obfuscation analysis information is integrated with the obfuscation optimization framework and optimization nodes to generate character obfuscation optimization information containing the obfuscation optimization framework and optimization nodes. The rule engine model automatically associates code-driven obfuscation analysis data (such as the "USA-O0" obfuscation pattern) with the obfuscation optimization framework and optimization nodes to generate structured character obfuscation optimization information, providing rule support for targeted optimization of the OCR model. The rule engine model architecture consists of the following components: Rule base: stores preset optimization mode generation rules (such as "weighting samples with the first 'O0' of the US MRZ code"). Inference engine: performs rule matching, logical reasoning, and conflict resolution. Fact base: stores input obfuscation analysis data (such as country codes, obfuscation patterns, and optimization nodes). Rule trigger threshold: 0.7 (rules are activated when the matching degree is ≥0.7). Inference step size: 1 (rules are processed sequentially).

[0049] The fact base construction process is as follows: Input code drives obfuscated data. Taking "USA-O0" as an example, it includes: the country code "USA", the obfuscated character pair "O0"; obfuscation patterns (such as "obfuscation probability increases by 40% when the first digit is present" and "obfuscation probability increases by 25% when the vertical line width is <1.5 pixels"); and generated optimization nodes (such as "adjust font vertical line width parameter" and "adjust the weight of the first and second sample"). The patterns described in natural language are converted into structured facts (such as "first digit" is converted into "character position ≤ 3").

[0050] The system invokes preset optimization rules, with examples as follows: When the country code is "USA" and a first-digit obfuscation pattern exists, an optimization node is generated that "increases the weight of the first O0 sample by 20%"; when the obfuscation pattern is related to the width of the vertical lines in the font, an optimization rule is generated that "sharpenes the image to enhance the vertical line features". The system is categorized and stored according to country code (e.g., United States, Japan), obfuscation pattern (position, font features), and optimization direction (sample weight, image preprocessing).

[0051] The inference engine scans the "USA-O0" data and matches two rules: "confusion with the first US name" and "insufficient vertical line width," with matching scores of 0.85 and 0.78 respectively (both exceeding the threshold of 0.7). Based on the matching results, corresponding optimization nodes (weight adjustment, image sharpening) are generated. The correlation between optimization nodes is detected (e.g., whether weight adjustment and image sharpening conflict), and adjustments are made according to the principle of "country code rule priority." The optimization framework (a graph structure with "USA" as the node) is integrated with the newly generated nodes to form a complete optimization scheme.

[0052] Using the country code "USA" as the core node, connect the edge to "O0 confusion probability 0.25" and associate it with sub-nodes such as "font thickness" and "position adjustment"; optimize the node set: Node 1: increase the weight of the first O0 sample by 20% (trigger rule: US first confusion); Node 2: image sharpening intensity 1.5 (trigger rule: insufficient vertical line width); rule association record: mark the preset rule corresponding to each optimization strategy for easy traceability.

[0053] S104: Process the selected training samples, retrain the OCR recognition model, and generate the model training results.

[0054] In one implementation, feature extraction processing is performed on the selected training samples to generate font type features, character confusion features, sample validity features, and target font features. The font type features include feature parameters for both target and non-target fonts. The character confusion features include data related to character confusion in specific country and region codes, and the confusion probability between 0 and O. The sample validity features include sample validity weights and dynamic selection mechanism parameters. The target font features include the feature vector of the OCR-B font and key feature parameters. Specifically, the font type features include feature parameters for both target fonts (e.g., OCR-B) and non-target fonts (e.g., Arial). For example, the contour curvature of the OCR-B font is 0.8, and the stroke width standard deviation is 0.2, while the stroke width uniformity parameter of the Arial font is 0.9.

[0055] Character confusion features include data related to character confusion in specific country and region codes, as well as the probability of confusion between O and 0. For example, in the US MRZ code, the probability of misidentifying "O" as "0" is 25%, and the probability of misidentifying "0" as "O" is 18%. Sample validity features cover sample validity weights and dynamic screening mechanism parameters. For example, if a sample validity weight is 0.85, samples with a confidence level ≥ 0.8 are retained in the dynamic screening mechanism. Target font features include the feature vector and key feature parameters of the OCR-B font. For example, the feature vector of the OCR-B font is [0.75, 0.82, 0.68...], and the key feature parameter is the thickness of the vertical line (1.8 pixels).

[0056] Feature extraction is performed on the model training parameters to generate optimization algorithm features, training strategy features, model structure features, and evaluation metric features. The optimization algorithm features include directional optimization strategy parameters and target optimization model parameters; the training strategy features include the number of retraining iterations and the learning rate; the model structure features include the number of network layers and nodes in the OCR recognition model; and the evaluation metric features include recognition accuracy and confusion matrix parameters. The optimization algorithm features include directional optimization strategy parameters and target optimization model parameters. For example, in the directional optimization strategy, the weight of the first O0 sample in the US MRZ code is increased by 20%, and the learning rate parameter of the target optimization model is adjusted to 0.001. The training strategy features include the number of retraining iterations and the learning rate. For example, the number of retraining iterations is 100, and the learning rate is 0.001.

[0057] Model structure features include the number of network layers and nodes in the OCR recognition model. For example, an OCR recognition model may have 5 network layers with 128, 256, 256, 128, and 64 nodes per layer. Evaluation metrics include recognition accuracy and confusion matrix parameters. For instance, a recognition accuracy of 95% means that "O" is misidentified as "0" 10 times and "0" is misidentified as "O" 8 times in the confusion matrix.

[0058] Based on font type features, character confusion features, sample validity features, and target font features, combined with optimization algorithm features, training strategy features, model structure features, and evaluation index features, the model training results are generated. These results characterize the OCR recognition model's accuracy in recognizing specific characters, its performance under different font types, and its correlation with character confusion data, forming an OCR recognition model training outcome that includes training process records, model performance indicators, and optimization effect evaluation. Features such as the feature vector of the OCR-B font, the confusion probability of the US MRZ code (25%), and the sample validity weight (0.85) are combined with parameters in the optimization algorithm, such as increasing the weight of the first US sample by 20%, iterating the training strategy 100 times, and the model's 5-layer network structure, for analysis. The generated model training results show that the OCR recognition model's accuracy in recognizing "O" and "0" in the US MRZ code increased from 75% to 92%. It also performs well under the OCR-B font and has a high degree of correlation with character confusion data. The training results of the OCR recognition model include training process records, model performance indicators (such as recognition accuracy of 92% and confusion matrix), and optimization effect evaluation (such as a 30% reduction in confusion error after optimization).

[0059] S105 processes the model training results, country and region codes, and character types in the character confusion-related data of specific country and region codes, combined with character confusion analysis, to generate character recognition optimization factors.

[0060] In one implementation, semantic analysis is performed on the text content in country and region codes and character types by combining code semantic parsing technology to obtain code semantic keywords and character confusion paragraph information. Semantic analysis is carried out on the character confusion data of the US MRZ code. According to the ICAO9303 specification, the following is an example of a standard machine-readable code: P<USAERIKSSON<<ANNA<MARIA<<<<<<<<<<<<<<<<<<<; O050055586USA640410531041501640410036243242. The country code (the first 2 - 5 digits "USA" in the first line) and the document number (the first 1 - 9 digits "O05005558" in the second line) are parsed from it. Core information such as the country code "USA" and character types "letter O and digit 0" are extracted through code semantic parsing technology, and code semantic keywords such as "USA", "passport machine-readable code", and "confusion between the first digit O / 0 of the document number" are obtained, as well as character confusion paragraph information such as "the first character 'O' of the document number is easily misrecognized as '0'".

[0061] Combined with the multi-dimensional feature fusion strategy, the model training results, code semantic keywords, character confusion paragraph information, font features, and confusion probability features are processed to establish cross-dimensional feature associations and generate multi-dimensional fusion feature data. The model training results (recognition accuracy 92%) of the US MRZ code, code semantic keywords ("USA", "first digit of the document number"), character confusion paragraph information ("the first digit O of the document number is misrecognized as 0"), OCR-B font features (vertical line thickness 1.8 pixels), and confusion probability features (O→0 confusion probability 25%) are deeply fused to construct the association logic of "USA document number first digit O-0 - font vertical line thickness - confusion probability 25% - recognition accuracy 92%", and multi-dimensional fusion feature data including country code, character position, font features, confusion probability, and recognition performance are generated.

[0062] Based on the cross-dimensional association rules and character recognition logical relationships, the multi-dimensional fusion feature data is structurally integrated to complete the construction of the cross-dimensional association network. According to the core association rule of "country code → document number position → font features → confusion probability → recognition performance", with "USA" as the root node, the sub-node "first digit O0 of the document number" is connected, and further leaf nodes such as "vertical line thickness of the font 1.8 pixels", "confusion probability 25%", and "recognition accuracy 92%" are associated to form a structurally clear association network.

[0063] Based on the character recognition optimization generation rules and application requirements, a cross-dimensional association network is verified and abnormal associations are adjusted using an optimization factor integrity detection mechanism to generate a target character recognition optimization framework. To address the need to "improve the recognition accuracy of the first O / 0 of the US MRZ code certificate number," the association strength of "USA-certificate number first font feature" in the network is detected (e.g., an imbalance in the weights of vertical line thickness and confusion probability is found). After adjusting the association logic, an optimization framework is generated with the goal of "optimizing the first O / 0 of the US certificate number," incorporating strategies such as font feature enhancement and sample weight adjustment.

[0064] According to the preset optimization factor generation rules, semantic analysis results, multi-dimensional fusion feature data, and cross-dimensional association networks are integrated to generate character recognition optimization factors that include code semantic features, character confusion features, and model training features. The core logic is to transform the "US ID number first 0 / 0 confusion" scenario features into an executable optimization strategy: when the first character of the US MRZ code ID number is detected to be in OCR-B font and the vertical line thickness is <1.5 pixels, a dual optimization strategy is triggered—increasing the training weight of the scenario sample by 20%, and simultaneously enhancing the vertical line feature to more than 1.5 pixels through image sharpening; the target accuracy is set at 95%, and the current optimization target is to reduce the confusion error by 3%.

[0065] This optimization factor is scalable. For example, when processing the Japanese MRZ code, you only need to replace the country code with "JPN" and adjust the parameters according to its certificate number structure (such as the rule of the first character of the Japanese MRZ code certificate number) and confusion probability data to quickly generate an adaptive optimization strategy.

[0066] S106. Based on the target optimization model and the directional optimization strategy, the model training results, character recognition optimization factors and character confusion optimization information are processed to generate character recognition optimization results.

[0067] In one implementation, based on the target optimization model, the correlation and matching degree between the model training results and the character recognition logic is introduced. The effectiveness of features and the accuracy of parameters in the character recognition optimization factors and character confusion optimization information are identified to generate an optimization anomaly feature set. The target optimization model analyzes the correlation and matching degree between the model training results of the US MRZ code (recognition accuracy 92%) and the character recognition logic (first digit O0 is easily confused). It finds that the feature effectiveness of "vertical line thickness ≥ 1.5 pixels" in the character recognition optimization factors is insufficient (the average thickness of vertical lines in actual scanned images is 1.2 pixels), and the accuracy of parameters is questionable (the model convergence speed did not meet expectations after increasing the sample weight by 20%). Therefore, an optimization anomaly feature set is generated, such as "vertical line thickness parameter not meeting the standard" and "insufficient effect of sample weight adjustment".

[0068] By comparing the model training results with the preset standard recognition model library and optimization parameter library, and verifying them with training examples, the deviation of the model performance from the optimization target is determined, and an optimization deviation evaluation result is generated. The training results of the US MRZ code model (accuracy 92%, target 95%) are compared with the standard accuracy (95%) of the same scenario in the standard model library. Combined with training examples (30 misidentifications of the first digit O0 in 1000 tests), the deviation of the performance from the target is determined to be 3%, and an optimization deviation evaluation result is generated, such as "Current accuracy is 3% below the target, misidentification rate exceeds the standard by 12%".

[0069] The deviation evaluation results are correlated with the set of abnormal features. Combining model characteristics, input parameters, and optimization suggestions from training feedback, the deviation evaluation results are weighted to generate an optimization vector that integrates the recognition logic and optimization process. For example, the deviation of "accuracy difference of 3%" is correlated with the abnormal features "vertical line thickness parameters not meeting standards" and "insufficient sample weight adjustment effect." Combining model characteristics (sensitivity to vertical line features), input parameters (current sharpening intensity 1.5), and training feedback (suggestion to increase sample weight to 25%), the deviation evaluation results are weighted to generate an optimization vector, such as "vertical line thickness weight 0.4, sample weight weight 0.6, overall optimization direction is to enhance sharpening to 2.0 and increase weight to 25%."

[0070] The optimization deviation evaluation results and the effect optimization vector are normalized and comprehensively calculated. Considering the real-time requirements of character recognition, the resulting character recognition optimization results include an optimization probability value, an effect deviation coefficient, and an optimization warning line. After normalizing the "deviation degree of 3%" and the effect optimization vector, and considering real-time requirements (response time ≤ 50ms), the following optimization results are generated: an optimization probability value of 85% (the possibility of improving to 95%), an effect deviation coefficient of 0.15 (the deviation between the current strategy and the optimal strategy), and an optimization warning line set at an accuracy of 93% (below this value triggers emergency optimization).

[0071] This application achieves precise optimization through data screening, country code-specific optimization, and multi-dimensional feature fusion. First, training samples are screened using a combination of CNN and SVM, eliminating non-target fonts (such as Arial) and retaining OCR-B font samples to improve data purity and reduce subsequent model training confusion errors by 30%. Second, LSTM and CNN are used to parse country codes and label confused data (such as "USA-O0"), combined with random forest and PCA to mine confusion patterns (such as high confusion probability for the first O / O), building a graph structure optimization framework. Then, a rule engine integrates code-driven confusion analysis to generate optimized nodes including weight adjustments and image sharpening. Finally, based on the target optimization model, the training results and optimization factors are fused to generate recognition results including accuracy and optimization probability, increasing the O / O recognition accuracy in US MRZ codes from 75% to 92%. This method, through a closed-loop process of "data screening - pattern analysis - framework construction - optimization execution," achieves targeted optimization of easily confused characters in specific country codes, improving robustness by 20% compared to traditional methods and is applicable to global MRZ code scenarios.

[0072] In one implementation, such as Figure 2 As shown, this application also provides an apparatus for improving the accuracy of OCR recognition models in recognizing specific characters, comprising:

[0073] The acquisition module 201 is used to acquire font type information in the training samples, including target fonts and non-target fonts, as well as character confusion related data in specific country and region codes. The character confusion related data includes the confusion probability of O and 0, and the target font includes optical character recognition B font.

[0074] The processing module 202 is used to process the font type information in the training samples, identify and remove non-target font samples, introduce target font samples, and generate filtered training samples; process the character confusion-related data in the specific country and region codes based on the specific country and region code analysis model, and generate character confusion optimization information by combining target font recognition analysis; process the filtered training samples, retrain the OCR recognition model, and generate model training results; process the model training results, country and region codes, and character types in the character confusion-related data in the specific country and region codes, and generate character recognition optimization factors; and process the model training results, character recognition optimization factors, and character confusion optimization information based on the target optimization model and the targeted optimization strategy to generate character recognition optimization results.

[0075] The computer-readable storage medium provided in the above embodiments of this application and the method for improving the accuracy of OCR recognition model for specific characters provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.

[0076] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for evaluating methods, electronic devices, electronic equipment, and readable storage media for improving the accuracy of OCR recognition models for specific characters are basically similar to the embodiments of the methods for improving the accuracy of OCR recognition models for specific characters described above, and therefore are described simply. Relevant parts can be referred to in the descriptions of the embodiments of the methods for improving the accuracy of OCR recognition models for specific characters described above.

Claims

1. A method for improving the accuracy of OCR recognition models in recognizing specific characters, characterized in that, include: Obtain font type information from the training samples, including target fonts and non-target fonts, as well as character confusion-related data in specific country and region codes. Character confusion-related data includes the confusion probability of O and 0. Target fonts include optical character recognition B fonts. The font type information in the training samples is processed to identify and remove non-target font samples, introduce target font samples, and generate filtered training samples. Based on the analysis model of specific country and region codes, the character obfuscation related data in the codes of specific countries and regions are processed, and combined with the target font recognition analysis, character obfuscation optimization information is generated; The selected training samples are processed to retrain the OCR recognition model and generate model training results. This includes feature extraction processing of the selected training samples to generate font type features, character confusion features, sample validity features, and target font features. Font type features include feature parameters of target fonts and non-target fonts. Character confusion features include character confusion-related data in specific country and region codes and the confusion probability of O and 0. Sample validity features include sample validity weights and dynamic screening mechanism parameters; target font features include the feature vector of optical character recognition (OCR) font B and key feature parameters. Feature extraction is performed on the model training parameters to generate optimization algorithm features, training strategy features, model structure features, and evaluation metric features. Optimization algorithm features include directional optimization strategy parameters and target optimization model parameters; training strategy features include the number of retraining iterations and the learning rate; model structure features include the number of network layers and nodes of the OCR recognition model; and evaluation metric features include recognition accuracy and confusion matrix parameters. Based on font type features, character confusion features, sample validity features, and target font features, combined with optimization algorithm features, training strategy features, model structure features, and evaluation metric features, the model training results are generated. The model training results are used to characterize the OCR recognition model's recognition accuracy for specific characters, the model's performance under different font types, and the degree of correlation with character confusion data, forming an OCR recognition model training result that includes training process records, model performance metrics, and optimization effect evaluation. The model training results, country and region codes, and character types in character confusion-related data in specific country and region codes are processed in combination with character confusion analysis to generate character recognition optimization factors. Based on the objective optimization model and the targeted optimization strategy, the model training results, character recognition optimization factors and character confusion optimization information are processed to generate character recognition optimization results.

2. The method as described in claim 1, characterized in that, The font type information in the training samples is processed to identify and remove non-target font samples, introduce target font samples, and generate filtered training samples, including: The training sample data is analyzed and labeled with features by extracting font type information. Target font sample selection rules and non-target font removal algorithms are introduced to achieve type purification and target focus of the sample data. By connecting the font feature database with the confusion probability analysis model, a font type matching matrix is ​​constructed. The sample validity weight is calculated through the target font feature benchmarking and transformation model, and a dynamic font type screening mechanism is established. Using font type as the classification dimension, a multi-dimensional font feature matrix is ​​constructed by integrating target font feature data, non-target font feature data, and character confusion probability data to achieve font dimension filtering of training sample data; Missing font information is processed by a sample feature completion model, key features are enhanced by a target font feature extraction mechanism, and the results are fused with font type matching matrix features to generate training sample selection results.

3. The method as described in claim 1, characterized in that, Based on a country and region code analysis model, character obfuscation-related data in country and region codes are processed. Combined with target font recognition analysis, character obfuscation optimization information is generated, including: By combining code analysis models for specific countries and regions with target font recognition processing, code information in character obfuscation-related data is parsed and extracted in real time to generate obfuscated data with code tags. By combining character obfuscation analysis and feature extraction strategies, the obfuscated data of code tags is processed, obfuscation pattern recognition is performed, and preliminary obfuscation analysis data is generated. Based on country and region code nodes, character obfuscation information, and obfuscation optimization rule generation rules, the preliminary obfuscation analysis data is structured and organized to complete the obfuscation optimization framework construction. Based on the characteristics of the target font, the system automatically generates optimized nodes by associating and obfuscating logical relationships, and marks abnormal associations and makes logical adjustments based on the node association detection mechanism. According to the preset optimization mode generation rules, the code-driven obfuscation analysis information is associated and integrated with the obfuscation optimization framework and optimization nodes to generate character obfuscation optimization information containing the obfuscation optimization framework and optimization nodes.

4. The method as described in claim 1, characterized in that, The model training results, country and region codes, and character types in character confusion-related data from specific country and region codes are processed in conjunction with character confusion analysis to generate character recognition optimization factors, including: Semantic analysis and feature extraction are performed on character obfuscation-related data. Combined with code semantic parsing technology, semantic parsing is performed on the text content in country and region codes and character types to obtain code semantic keywords and character obfuscation paragraph information. By combining multi-dimensional feature fusion strategies, the model training results, code semantic keywords, character confusion paragraph information and font features, and confusion probability features are processed to perform cross-dimensional feature association and generate multi-dimensional fused feature data. Based on the cross-dimensional association rules and character recognition logic, multi-dimensional fused feature data is structurally integrated to complete the construction of a cross-dimensional association network; Based on the character recognition optimization generation rules and application requirements, the cross-dimensional association network is verified and abnormal associations are adjusted based on the optimization factor integrity detection mechanism to generate the target character recognition optimization framework. According to the preset optimization factor generation rules, the semantic analysis results, multi-dimensional fusion feature data and cross-dimensional association network are linked and integrated to generate character recognition optimization factors that include code semantic features, character obfuscation features and model training features.

5. The method as described in claim 4, characterized in that, Based on the objective optimization model and the targeted optimization strategy, the model training results, character recognition optimization factors, and character obfuscation optimization information are processed to generate character recognition optimization results, including: Based on the target optimization model, the correlation and matching degree between the model training results and the character recognition logic is introduced. The effectiveness of the features and the accuracy of the parameters in the character recognition optimization factor and the character confusion optimization information are identified to generate a set of optimization anomaly features. By comparing the model training results with the preset standard recognition model library and optimization parameter library, and verifying them with training examples, the deviation of the model performance from the optimization target is determined, and the optimization deviation evaluation results are generated. The optimization deviation evaluation results are correlated with the optimization anomaly feature set. Combined with model characteristics, input parameters and optimization suggestions in training feedback, the optimization deviation evaluation results are weighted to generate an effect optimization vector that integrates recognition logic and optimization process. The optimization deviation evaluation results and the effect optimization vector are normalized and comprehensively calculated. Combined with the real-time requirements of character recognition, the character recognition optimization results containing optimization probability values, effect deviation coefficients and optimization warning lines are generated.

6. An apparatus for improving the accuracy of OCR recognition models in recognizing specific characters, characterized in that, The apparatus for implementing the method of claim 1 includes: The acquisition module is used to acquire font type information in the training samples, including target fonts and non-target fonts, as well as character confusion-related data in specific country and region codes. The character confusion-related data includes the confusion probability of O and 0, and the target font includes optical character recognition B font. The processing module is used to process the font type information in the training samples, identify and remove non-target font samples, introduce target font samples, and generate filtered training samples; it processes character confusion-related data in specific country and region codes based on a specific country and region code analysis model, and generates character confusion optimization information by combining target font recognition analysis; it processes the filtered training samples, retrains the OCR recognition model, and generates model training results; it processes the model training results, country and region codes, and character types in the character confusion-related data in specific country and region codes, combined with character confusion analysis, to generate character recognition optimization factors; and it processes the model training results, character recognition optimization factors, and character confusion optimization information based on a target optimization model and a targeted optimization strategy to generate character recognition optimization results.

7. An electronic device, characterized in that, include: First processor; and memory for storing executable instructions of the first processor; The first processor is configured to execute the method for improving the accuracy of OCR recognition model for specific characters as described in any one of claims 1 to 5 by executing the executable instructions.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the second processor, it implements the method for improving the accuracy of OCR recognition model for specific characters as described in any one of claims 1 to 5.

Images