Test paper content fuzzification method based on deep learning
By using deep learning-based text detection and character recognition technologies, the text coordinates in the scanned exam papers were obtained and then pixelated, thus solving the problem of exam paper leakage and improving the security and confidentiality of the exam papers.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-08-02
- Publication Date
- 2026-03-10
AI Technical Summary
There are risks of leakage in the printing, distribution, storage, safekeeping and technical security of postgraduate entrance examination papers, especially the risk of leakage of exam content and leakage of exam papers by personnel checking the questions.
A deep learning-based approach was used to obtain the coordinates of all text in the scanned exam paper using a text detection model, and to recognize the text of the questions using a character recognition model. Finally, the content other than the questions was pixelated.
It effectively solved the problems of leaked exam paper content and leaked exam papers by those checking the questions, reducing the risk of previous leaks.
Smart Images
Figure CN121640478A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of optical character recognition of deep learning, and particularly relates to a test paper content blurring method based on deep learning. BACKGROUND
[0002] Optical Character Recognition (OCR) is a technology that converts printed or handwritten characters into editable and searchable text format. Image preprocessing: Before performing optical character recognition, the input image usually needs to be preprocessed. This includes operations such as image graying, binarization, noise reduction, contrast enhancement, etc., to facilitate character segmentation and recognition; Character segmentation: Character segmentation is the process of separating characters from the background in an image. This process can be based on image pixel values, edge detection, connected region analysis, etc., to ensure that each character is independently recognized; Feature extraction: After character segmentation, features need to be extracted from each individual character image to facilitate subsequent recognition. Common feature extraction methods include shape features, statistical features, projection features, etc., to represent character shape, contour, texture, etc. information; Character classification: Character classification is to compare the extracted character features with the pre-defined character templates to determine the recognition result of each character. This can be achieved by using machine learning algorithms (such as support vector machines, neural networks, etc.) or rule-based methods; Dictionary correction: In some cases, optical character recognition may have errors, such as misrecognized characters or recognition results not in the dictionary. To improve accuracy, dictionary correction algorithms can be used to correct incorrect recognition results based on common vocabulary and language rules; Post-processing: Finally, the recognized text is post-processed, including removing spaces, merging consecutive characters, correcting spelling errors, etc., to provide more accurate and standardized text results.
[0003] These techniques and steps play a key role in optical character recognition, helping to convert characters in images into editable and searchable text. With the development of deep learning and computer vision, modern OCR systems have made significant progress in accuracy and performance.
[0004] In order to maintain the fairness and security of the examination, the leakage of the postgraduate examination paper is a serious problem, and various measures need to be taken to prevent and respond to the leakage. The following factors that make the postgraduate examination paper prone to leakage: printing and distribution process: there is a potential risk of leakage in the printing and distribution process of the postgraduate examination paper. This includes the security of the paper in the printing factory, the logistics process, and the security of the distribution to the examination centers; storage and storage: the paper is also vulnerable to leakage during storage and storage. The paper may be stored in the office, warehouse or archive room of the examination center or other relevant institutions, and strict security measures need to be taken to ensure the safety of the paper; personnel access: only authorized personnel can access and handle the postgraduate examination paper, but if the personnel access is not strictly managed, there is a risk of leakage. It is necessary to ensure that the paper can only be accessed by authorized personnel and to take measures to record the access and handling of the paper; technical security: electronic storage and transmission of the paper also need to pay attention to security. Electronic paper may be stored on servers, computers or storage devices, and measures such as encryption and access control need to be taken to prevent unauthorized access and leakage; human factors: human factors, such as negligence, misconduct or intentional operation of internal personnel or relevant personnel, may cause the paper to be leaked. This includes copying, photographing, transmitting or leaking the paper to unauthorized personnel.
[0005] Therefore, it is necessary to provide a new deep learning-based paper content blurring method to solve the above technical problems. SUMMARY
[0006] The technical problem solved by the present application is to provide a deep learning-based paper content blurring method that can solve the problem of leakage of postgraduate examination paper content, leakage of examination paper by examination paper checking personnel and high risk of leakage in the past.
[0007] To solve the above technical problems, the deep learning-based paper content blurring method provided by the present application comprises the following steps:
[0008] S1. Collect image data of postgraduate examination paper scans and label;
[0009] S2. Automatically obtain the coordinates of all text in the postgraduate examination paper scan picture through a text detection model;
[0010] S3. Identify the text of the questions in the postgraduate examination paper scan through a character recognition model;
[0011] S4. Finally, the content of the paper except the question part is processed with mosaic according to the detected coordinates and the recognized text of the questions.
[0012] As a further scheme of the present application, the labeling includes text box labeling and in-box text labeling, and when collecting image data for labeling, key words also need to be identified, so the following four stages are also included:
[0013] (1). Pretreatment stage: it is helpful to improve the accuracy and efficiency of the subsequent character recognition algorithm, so the image data needs to be preprocessed first, generally image enhancement, noise reduction, image binarization, character positioning, image rotation and correction are carried out;
[0014] (2). Feature extraction stage: some valuable parameters representing image signals are extracted from image signals;
[0015] (3). Text detection training stage: a certain model is trained on the computer, then the effective feature values after training are obtained, and then these feature data are stored in the memory of the system, and these values are used as the basis for the final recognition stage;
[0016] (4). Character recognition training stage: a certain model is trained on the computer, then the effective feature values after training are obtained, and then these feature data are stored in the memory of the system, and these values are used as the basis for the final recognition stage.
[0017] As a further scheme of the application, the S1 further comprises the following:
[0018] A specific camera is used to take pictures of the scanned test paper, noise reduction processing is carried out, and a filter or other noise reduction algorithm is used to remove noise in the image, such as Gaussian blur and median filter, which helps to reduce interference in the subsequent character recognition process.
[0019] As a further scheme of the application, the S2 further comprises the following:
[0020] The text detection model is trained by using an open source text detection data set, and the text detection model is fine-tuned using scanned test paper pictures, the number of open source text detection data sets is more than 100,000 pictures, and the number of scanned test paper pictures is more than 500.
[0021] As a further scheme of the application, the S3 further comprises the following:
[0022] The text recognition model is trained by using an open source text recognition data set, and the text recognition model is fine-tuned using scanned test paper pictures, the number of open source text recognition data sets is more than 100,000 pictures, and the number of scanned test paper pictures is more than 500.
[0023] As a further scheme of the application, the pretreatment stage specifically comprises the following:
[0024] (1). Data preprocessing for character detection and recognition is an important step, which helps to improve the accuracy and efficiency of the subsequent character recognition algorithm;
[0025] (2). Image enhancement: Perform enhancement processing on the input image, such as adjusting contrast, brightness, sharpening, etc., to improve the clarity and readability of the text;
[0026] (3). Noise reduction: Use filters or other noise reduction algorithms to remove noise in the image, such as Gaussian blur, median filtering, etc., which helps reduce interference in the subsequent text recognition process;
[0027] (4). Image binarization: Convert the image to a binary image, making the text area and background area clearly separated, which can be achieved through thresholding, adaptive thresholding, etc.
[0028] (5). Text localization: Use text detection algorithms, such as edge-based, connected components, etc., to locate the text area in the image, which helps reduce the search space for subsequent text recognition;
[0029] (6). Text segmentation: For some special scenarios, such as multiple columns of text in a document, text segmentation is needed to separate different text areas;
[0030] (7). Image rotation and correction: If the text in the image is inclined or rotated, automatic rotation correction can be performed to make the text horizontal or vertical, to improve the accuracy of subsequent text recognition.
[0031] As a further scheme of the present application, the training phase is as follows:
[0032] (1). Data preparation: Collect image data containing text and perform labeling, labeling the position information of the text, which will be used for model training and evaluation;
[0033] (2). Data preprocessing: Preprocess the image data, including image scaling, normalization, cropping, etc., to adapt to the input requirements of the model;
[0034] (3). Model construction: Select a suitable model architecture, such as a convolutional neural network (CNN) based model, for detecting text in images. Existing model architectures such as FasterR-CNN, YOLO, etc. can be used, or custom model structures can be used;
[0035] (4). Model training: Use the prepared data set and corresponding labeling information to train the text detection model, which involves providing input images to the model and using backpropagation algorithms to optimize model parameters to minimize the difference between predicted results and labels;
[0036] (5). Parameter adjustment: According to the performance of the model on the validation set, perform hyperparameter adjustment, learning rate adjustment, etc. to improve the accuracy and generalization ability of the model;
[0037] (6). Model evaluation: Use an independent test set to evaluate the performance of the model. Common evaluation indicators include accuracy, recall, F1 score, etc.
[0038] Compared with the related art, the deep learning-based test paper content blur method provided by the application has the following beneficial effects:
[0039] 1. The application can solve the problem of test paper content leakage;
[0040] 2. The application can solve the problem of test paper leakage by test checking personnel;
[0041] 3. The application can solve the problem of high risk of leakage in the past. BRIEF DESCRIPTION OF DRAWINGS
[0042] In order to facilitate those skilled in the art to understand, the application will be further described below with reference to the drawings.
[0043] Figure 1 is a schematic diagram of the overall method structure in the application;
[0044] Figure 2 is a schematic diagram of the text detection model network in the application;
[0045] Figure 3 is a schematic diagram of the character recognition model network in the application;
[0046] Figure 4 is a schematic diagram of the test paper scan in the application;
[0047] Figure 5 is a schematic diagram of the test paper blur result in the application. DETAILED DESCRIPTION
[0048] Please refer to Figure 1 , Figure 2 , Figure 3 , Figure 4 and Figure 5 , wherein, Figure 1 is a schematic diagram of the overall method structure in the application; Figure 2 is a schematic diagram of the text detection model network in the application; Figure 3 is a schematic diagram of the character recognition model network in the application; Figure 4 is a schematic diagram of the test paper scan in the application; Figure 5 is a schematic diagram of the test paper blur result in the application. The deep learning-based test paper content blur method includes the following steps:
[0049] S1. Collect image data of test paper scans and label;
[0050] S2. Automatically obtain the coordinates of all texts in the scanned picture of the postgraduate entrance examination paper through a text detection model;
[0051] S3. Recognize the text of the question in the scanned picture of the postgraduate entrance examination paper through a character recognition model.
[0052] S4. Finally, the content of the paper except the question part is processed by mosaicking the recognized text of the question combined with the detected coordinates.
[0053] The annotation includes text box annotation and in-box text annotation, and when collecting image data for annotation, it is also necessary to identify key words, so it also includes the following four stages:
[0054] (1). Preprocessing stage: It is helpful to improve the accuracy and efficiency of the subsequent character recognition algorithm, so it is necessary to preprocess the image data first, which generally includes image enhancement, noise reduction, image binarization, text positioning, image rotation and correction, etc.
[0055] (2). Feature extraction stage: Extract some valuable parameters representing image signals from image signals;
[0056] (3). Text detection training stage: Train on the computer using a certain model, then obtain the effective feature values after training, and then store these feature data in the memory of the system, and use these values as the basis for the final recognition stage;
[0057] (4). Character recognition training stage: Train on the computer using a certain model, then obtain the effective feature values after training, and then store these feature data in the memory of the system, and use these values as the basis for the final recognition stage.
[0058] The S1 also includes the following:
[0059] A specific camera is used to take pictures of the scanned picture of the postgraduate entrance examination paper, and noise reduction processing is performed, using a filter or other noise reduction algorithm to remove noise in the image, such as Gaussian blur and median filter, which helps to reduce interference in the subsequent character recognition process.
[0060] The S2 also includes the following:
[0061] The text detection model is trained through an open source text detection dataset, and the scanned picture of the postgraduate entrance examination paper is fine-tuned, the number of open source text detection datasets is more than 100,000 pictures, and the number of scanned pictures of the postgraduate entrance examination paper is more than 500.
[0062] The S3 also includes the following:
[0063] The text recognition model is trained by an open source text recognition dataset, and then fine-tuned by using the scanned pictures of the postgraduate entrance examination papers. The number of open source text recognition datasets is more than 100,000 pictures, and the number of scanned pictures of the postgraduate entrance examination papers is more than 500.
[0064] The preprocessing stage is specifically as follows:
[0065] (1). Data preprocessing for character detection and recognition is an important step, which helps to improve the accuracy and efficiency of subsequent text recognition algorithms;
[0066] (2). Image enhancement: enhance the input image, such as adjusting contrast, brightness, sharpening, etc., to improve the clarity and readability of the text;
[0067] (3). Noise reduction: use filters or other noise reduction algorithms to remove noise in the image, such as Gaussian blur, median filter, etc., which helps to reduce interference in the subsequent text recognition process;
[0068] (4). Image binarization: convert the image to a binary image, so that the text area and background area are clearly separated, which can be achieved by thresholding, adaptive thresholding, etc.;
[0069] (5). Text localization: use text detection algorithms, such as edge-based, connected component-based methods, to locate the text area in the image, which helps to reduce the search space for subsequent text recognition;
[0070] (6). Text segmentation: for some special scenarios, such as multiple columns of text in a document, text segmentation is needed to separate different text areas;
[0071] (7). Image rotation and correction: if the text in the image is inclined or rotated, automatic rotation correction can be performed to make the text horizontal or vertical, which improves the accuracy of subsequent text recognition.
[0072] The specific construction process is as follows:
[0073] 1). Build a text detection model, the goal is to extract the position coordinates of the text in a scanned image of the postgraduate entrance examination by deep learning technology, the input is a single page image of the scanned image of the postgraduate entrance examination, and the output is the coordinates of the top left corner, top right corner, bottom left corner and bottom right corner of the text in the scanned image;
[0074] The speaker segmentation clustering model labeling rules are as follows:
[0075] Use PPOCRLabel software to frame each text in the scanned image of the postgraduate entrance examination, and frame each line of text and each block of text separately.
[0076] Model building: the speaker segmentation clustering model network uses Differentiable, the Binarization network is built, and the algorithm is summarized as follows:
[0077] a). input the input image into the main body of a feature pyramid;
[0078] b). upsample the pyramid features to the same scale and concatenate to generate features F, then use features F to predict probability map (P) and threshold map (T);
[0079] c). Calculate the approximate binary mapping (B^) from P and f, during training, the probability mapping, threshold mapping and approximate binary mapping are supervised, where the probability mapping and the approximate binary mapping share the same supervision, during inference, the bounding box can be easily obtained from the approximate binary mapping or the probability mapping through a box formula module;
[0080] Model training: use open source dataset for training, then fine-tune through labeled postgraduate scanned image small dataset, and iteratively adjust parameters.
[0081] Model evaluation: use the data of the validation set to select the trained model.
[0082] Model deployment: use Python backend Flask framework for API deployment.
[0083] 2). Build a text recognition model, the goal is to learn the text characters of the postgraduate scanned image through deep learning technology;
[0084] The annotation rules of the text recognition model are as follows:
[0085] Use PPOCRLabel software to recognize the text, and then manually modify the results of the recognition errors.
[0086] Model building: the voiceprint recognition model network uses PP-OCR network building, and the network structure is as follows:
[0087] Use CRNN as a text recognizer, which is widely used in text recognition. CRNN integrates feature extraction and sequence modeling, which uses connectionist temporal classification (CTC) loss to avoid inconsistency between prediction and label. In order to improve the model capacity of the text recognizer and reduce the model size, we use the following 9 strategies: light backbone, data augmentation, cosine learning rate decay, feature map parsing-test, regularization parameter, learning rate warm-up, light head, pre-trained model and PACT quantization.
[0088] Model training: use open source dataset for training, then fine-tune through labeled postgraduate scanned image small dataset, and iteratively adjust parameters.
[0089] Model evaluation: Use the data of the validation set to select the trained model.
[0090] Model deployment: Use the Python backend Flask framework for API deployment.
[0091] 3). Use the Python Flask software backend framework to build the identification interface of each model, and combine the detected coordinates with the recognized text to mosaic the content of the test paper except the question part, and automatically generate the blurred test paper after the blur processing through the Python Flask backend technology, to complete the blur encryption of the test paper.
[0092] The training stage is as follows:
[0093] (1). Data preparation: Collect image data containing text and perform labeling, label the position information of the text, these data will be used for model training and evaluation;
[0094] (2). Data preprocessing: Preprocess the image data, including image scaling, normalization, cropping, etc. to adapt to the input requirements of the model;
[0095] (3). Model construction: Select a suitable model architecture, such as a convolutional neural network (CNN) based model, for detecting text in images. Existing model architectures such as FasterR-CNN, YOLO, etc. can be used, or custom model structures can be used;
[0096] (4). Model training: Use the prepared data set and corresponding labeling information to train the text detection model, which involves providing input images to the model and using the backpropagation algorithm to optimize model parameters to minimize the difference between predicted results and labels;
[0097] (5). Parameter adjustment: According to the performance of the model on the validation set, adjust the hyperparameters, learning rate, etc. to improve the accuracy and generalization ability of the model;
[0098] (6). Model evaluation: Use an independent test set to evaluate the performance of the model. Common evaluation indicators include accuracy, recall, F1 score, etc.
[0099] The present application can solve the problems of test paper content leakage, test paper leakage by test paper checking personnel and high risk of previous leakage.
[0100] The above merely describes preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacements or changes according to the technical scheme and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.
Claims
1. A deep learning-based test paper content blurring method, characterized in that, It comprises the following steps: S1. Collecting image data of the scanned copy of the postgraduate entrance examination paper and labeling; S2. Automatically obtaining the coordinates of all texts in the scanned copy of the postgraduate entrance examination paper through a text detection model; S3. Recognizing the text of the questions in the scanned copy of the postgraduate entrance examination paper through a text recognition model; S4. Finally, the content of the paper except the question part is processed by mosaic in combination with the recognized text and the detected coordinates. 2.The deep learning-based test paper content blurring method according to claim 1, wherein: The labeling includes text box labeling and in-box text labeling, and when collecting image data for labeling, key words also need to be recognized, so the following four stages are also included: (1). Preprocessing stage: It helps to improve the accuracy and efficiency of the subsequent text recognition algorithm, so the image data needs to be preprocessed first, which generally includes image enhancement, noise reduction, image binarization, text positioning, image rotation and correction, etc. (2). Feature extraction stage: Extract some valuable parameters representing image signals from image signals; (3). Text detection training stage: Train on the computer using a certain model, then obtain the effective feature values after training, and then store these feature data in the memory of the system, which will be used as the basis for the final recognition stage; (4). Text recognition training stage: Train on the computer using a certain model, then obtain the effective feature values after training, and then store these feature data in the memory of the system, which will be used as the basis for the final recognition stage. 3.The deep learning based test paper content blurring method according to claim 1, wherein: The S1 also includes the following: Use a specific camera to take pictures of the scanned copy of the postgraduate entrance examination paper, perform noise reduction processing, and use a filter or other noise reduction algorithm to remove noise in the image, such as Gaussian blur and median filtering, which helps to reduce interference in the subsequent text recognition process. 4.The deep learning based test paper content blurring method according to claim 1, wherein: The S2 also includes the following: Train the text detection model using an open source text detection dataset, and fine-tune it using scanned copy of the postgraduate entrance examination paper pictures, the number of open source text detection dataset is more than 100,000 pictures, and the number of scanned copy of the postgraduate entrance examination paper pictures is more than 500. 5.The deep learning based test paper content blurring method according to claim 1, wherein: The S3 also includes the following: Train the text recognition model using an open source text recognition dataset, then fine-tune it using scanned copy of the postgraduate entrance examination paper pictures, the number of open source text recognition dataset is more than 100,000 pictures, and the number of scanned copy of the postgraduate entrance examination paper pictures is more than 500. 6.The deep learning based test paper content blurring method according to claim 1, wherein: The preprocessing stage is as follows: (1). Data preprocessing for text detection and recognition is an important step, which helps to improve the accuracy and efficiency of the subsequent text recognition algorithm; (2). Image enhancement: Perform enhancement processing on the input image, such as adjusting contrast, brightness, sharpening, etc., to improve the clarity and readability of the text; (3). Noise reduction: Use a filter or other noise reduction algorithm to remove noise in the image, such as Gaussian blur, median filtering, etc., which helps to reduce interference in the subsequent text recognition process; (4). Image binarization: Convert the image to a binary image to clearly separate the text area from the background area, which can be achieved through thresholding, adaptive thresholding, etc. (5). Text localization: Use text detection algorithms, such as edge-based, connected component-based methods, to locate the text regions in the image, which helps to reduce the search space for subsequent text recognition; (6). Text segmentation: For some special scenarios, such as multi-column text in documents, text segmentation is needed to separate different text regions; (7). Image rotation and correction: If the text in the image is tilted or rotated, automatic rotation correction can be performed to make the text horizontal or vertical, to improve the accuracy of subsequent text recognition. 7.The deep learning based test paper content blurring method according to claim 1, wherein: The training phase is as follows: (1). Data preparation: Collect image data containing text and perform labeling, labeling the position information of the text, which will be used for model training and evaluation; (2). Data preprocessing: Preprocess the image data, including image scaling, normalization, cropping, etc., to adapt to the input requirements of the model; (3). Model construction: Select a suitable model architecture, such as a convolutional neural network (CNN) based model, for detecting text in images. Existing model architectures such as Faster R-CNN, YOLO, etc. can be used, or custom model structures can be defined; (4). Model training: Use the prepared data set and corresponding labeling information to train the text detection model, which involves providing input images to the model and using backpropagation algorithms to optimize model parameters to minimize the difference between predicted results and labels; (5). Parameter adjustment: According to the performance of the model on the validation set, adjust the hyperparameters, learning rate, etc. to improve the accuracy and generalization ability of the model; (6). Model evaluation: Use an independent test set to evaluate the performance of the model. Common evaluation metrics include accuracy, recall, F1 score, etc.