Intelligent paper marking system based on image cutting and recognition

Abstract

The invention discloses an intelligent paper marking system based on image cutting and recognition, belongs to the technical field of intelligent paper marking, aims at solving the problem that the answer quality of students cannot be comprehensively and accurately evaluated, and comprises an image acquisition module which is used for acquiring test paper images, automatically adjusting scanning parameters according to the size and color mode of test paper, supporting multi-angle scanning and acquiring the test paper images; the image preprocessing module comprises an image enhancement unit, a tilt correction unit and a binarization processing unit and is used for carrying out brightness and contrast adjustment, deblurring, tilt correction and binarization processing on the acquired image; through the image cutting, character recognition and semantic understanding technologies, the answer content, including complex handwritten fonts, altered characters and the like, on test papers in various formats can be accurately recognized, misjudgment caused by non-standard formats or writing problems is avoided, the scoring result is more objective and accurate, and the real answer level of students can be better reflected.

Description

Technical Field

[0001] This invention belongs to the field of intelligent marking technology, specifically relating to an intelligent marking system based on image segmentation and recognition. Background Technology

[0002] With the rapid development of computers and artificial intelligence, they are gradually being applied to various fields of work and life. In the education and teaching process, examinations are an effective way to evaluate teaching outcomes. After an examination, teachers have to grade a large number of student papers and summarize and analyze the results of all the graded students. Traditional grading and analysis work takes up a lot of teachers' time and even reduces their lesson preparation time. Therefore, computer-based intelligent grading technology has emerged, which can greatly save teachers' grading time.

[0003] Currently common intelligent marking systems rely heavily on answer sheets with specific formats when processing objective questions. They judge answers by scanning and recognizing the filled-in areas. This method has extremely high requirements for the standardization of answer sheets, and misjudgments are easily caused by improper filling by candidates or deviations in the scanning equipment. For subjective question marking, existing technologies often use simple keyword matching or text recognition based on fixed templates. This makes it difficult to accurately understand the complete semantics and logical relationships of students' answers, and thus cannot comprehensively and accurately assess the quality of students' answers. Moreover, their ability to recognize complex handwriting and corrections is limited, and they cannot effectively adapt to diverse answering scenarios.

[0004] Therefore, an intelligent marking system based on image segmentation and recognition is needed to solve the problem that existing technologies cannot comprehensively and accurately assess the quality of students' answers. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent marking system based on image segmentation and recognition to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent marking system based on image segmentation and recognition, comprising:

[0007] The image acquisition module is used to acquire images of the test paper and automatically adjust the scanning parameters according to the size and color mode of the test paper, and supports multi-angle scanning;

[0008] The image preprocessing module includes an image enhancement unit, a tilt correction unit, and a binarization processing unit, which are used to perform brightness and contrast adjustment, deblurring, tilt correction, and binarization processing on the acquired image.

[0009] The image segmentation module includes a region localization unit and an answer region segmentation unit. It locates different functional regions of the test paper based on a deep learning object detection algorithm and segments the answer region using contour detection and connected component analysis techniques.

[0010] The text recognition module includes a multi-text / multi-language recognition unit and an altered / blurred text processing unit. It uses a model based on convolutional neural networks combined with attention mechanisms to recognize multi-font / multi-language text and uses generative adversarial networks to process altered / blurred text.

[0011] The semantic understanding and scoring module includes a semantic analysis unit and a scoring model unit. It uses natural language processing technology to perform semantic analysis on the answers to subjective questions, constructs a comprehensive scoring model to score subjective questions from multiple dimensions, and scores objective questions based on the matching of the recognition results with the standard answers.

[0012] The results output module includes a teacher-side display unit and a student-side display unit. The teacher-side display unit provides teachers with student grade reports and answer analysis, while the student-side display unit shows students their test paper answers, score details, and teacher comments.

[0013] It should be noted in the solution that the region localization unit, based on the target detection model of the YOLO series algorithms, analyzes the preprocessed image to accurately locate different functional areas in the test paper, such as the answer area, question area, and candidate information area. After locating the answer area, the answer area segmentation unit further segments each answer area into individual answer areas according to the question type and answer format. For subjective question answer areas, it intelligently identifies paragraph breaks and line breaks, segmenting continuous text into reasonable sub-regions, providing more accurate data for subsequent text recognition and semantic analysis.

[0014] It is worth further explaining that, during the model training phase, the region localization unit constructs a labeled dataset containing various test paper types and different layout styles. Each sample image is labeled with the bounding box coordinates and category labels of various functional regions. The model uses Darknet as the backbone feature extraction network and extracts multi-scale features of the image through multi-layer convolution operations. Among them, shallow features are used to capture the detailed information of the region, and deep features are used to extract the global semantic information of the region.

[0015] In the feature fusion stage, a feature pyramid network structure is adopted to fuse feature maps at different levels to generate feature maps suitable for detecting regions of different sizes. For the preset anchor boxes on each feature map, the model predicts the offset, confidence and class probability of the bounding box. Redundant bounding boxes are filtered by the non-maximum suppression algorithm, and finally the accurate localization result of each functional region is output.

[0016] Furthermore, it should be noted that the answer area segmentation unit, for objective questions such as multiple-choice and true / false questions, typically contains multiple independent option filling boxes or answer filling grids. First, a sub-image of the objective question answer area is extracted based on the area positioning results. Then, a connected component analysis algorithm is used to identify the outline of each option. By calculating the circumscribed rectangle of the outline, the boundary range of each option is determined. Then, combined with the number of questions and the arrangement pattern of the options, adjacent option areas are separated. For objective questions with a matrix arrangement, the horizontal and vertical spacing between adjacent outlines is calculated to automatically divide the row and column boundaries, ensuring that each option is segmented individually.

[0017] For subjective questions, the answer area is mostly a continuous blank writing area. A segmentation strategy based on a combination of projection analysis and contour detection is adopted: First, the sub-image of the subjective question area is horizontally projected, and the proportion of black pixels in each row is calculated. When the proportion of multiple consecutive rows is lower than a set threshold, it is determined to be a paragraph separator or line break. At the same time, the connected components of handwritten text are identified by contour detection. Based on the vertical distribution characteristics of the connected components, the results of projection analysis are corrected to avoid missegmentation caused by excessive spacing between individual characters. For the answer area of ​​subjective questions containing formulas or charts, the contour boundaries of formulas or charts are identified by edge detection algorithms and segmented as independent sub-regions to ensure the separation of text and non-text content.

[0018] As a preferred embodiment, the method by which the text processing unit effectively identifies altered or blurred text during student answering of questions includes:

[0019] A multi-feature fusion algorithm for identifying tampered areas is adopted, which combines three dimensions—pixel change rate, edge continuity, and color difference—to determine the tampered areas.

[0020] Based on the sharpness assessment index, blurry regions are located, and the second derivative of the image is calculated using the Laplacian operator. Blurry regions lack high-frequency information and have a weak derivative response, thus identifying blurry regions.

[0021] For the identified altered or blurred areas, the system constructs a semantic-aware generative adversarial network, which combines contextual semantic information to achieve text repair and accurate recognition;

[0022] The accuracy of the restored text image is further improved through multi-model fusion recognition.

[0023] Compared with existing technologies, the intelligent marking system based on image segmentation and recognition provided by the present invention has at least the following beneficial effects:

[0024] (1) Through image segmentation, text recognition and semantic understanding technology, it can accurately identify the answer content on various test papers, including complex handwritten fonts, corrections, etc., avoiding misjudgment caused by non-standard format or writing problems, making the scoring results more objective and accurate, and better reflecting the students' true answer level.

[0025] (2) The image acquisition module and regional positioning technology can adapt to test papers of different sizes and formats without the need for specific answer sheets, making it convenient for teachers to flexibly set questions and for students to answer questions; the text recognition module has the ability to recognize multiple fonts and languages, which can meet diverse teaching needs.

[0026] (3) The fully automated marking process greatly shortens the marking time and reduces the workload of teachers, allowing them to devote more time and energy to teaching. At the same time, the results output module provides teachers and students with detailed grade reports and answer analysis, helping teachers to accurately grasp the teaching effect, discover problems in teaching, provide students with targeted learning suggestions, and promote the improvement of teaching quality. Attached Figure Description

[0027] Figure 1 This is a block diagram of the intelligent marking system based on image segmentation and recognition according to the present invention;

[0028] Figure 2 This is a flowchart illustrating the method for the altered / blurred text processing unit of the present invention to identify altered or blurred text. Detailed Implementation

[0029] The present invention will be further described below with reference to embodiments.

[0030] Reference Figure 1 As shown, this invention provides an intelligent marking system based on image segmentation and recognition, comprising:

[0031] The image acquisition module is used to acquire exam paper images and automatically adjust scanning parameters according to the exam paper size, color mode, etc., supporting multi-angle scanning. It employs a high-resolution, high-color-fidelity scanning device, enabling fast and clear acquisition of exam paper images. To adapt to different sizes and types of exam papers, the scanning device has an automatic parameter adjustment function, intelligently adapting to the size and color mode of the exam paper to ensure the acquired image quality meets subsequent processing requirements. Simultaneously, this module also supports multi-angle scanning of the exam paper to obtain more comprehensive image information, avoiding the inability to recognize some content due to factors such as exam paper creases and shadows. Using a high-resolution scanner equipped with automatic paper feeding, the completed exam paper is placed into the scanner's paper feed slot. The scanner automatically detects the exam paper size, adjusts the scanning parameters, and scans the exam paper at a resolution of 300 dpi in color mode, generating a clear exam paper image file, which is then transferred to the image preprocessing module.

[0032] The image preprocessing module includes an image enhancement unit, a tilt correction unit, and a binarization unit. These units are used to perform brightness and contrast adjustments, deblurring, tilt correction, and binarization on the acquired images. The image enhancement unit utilizes histogram equalization and contrast-limited adaptive histogram equalization techniques to adjust the brightness and contrast of the acquired images, highlighting text information and suppressing background noise, making the text in the image clearer and more legible. For blurred images, Gaussian filtering and median filtering algorithms are used for deblurring to improve image clarity. The tilt correction unit detects text line features in the image, uses the Hough transform algorithm to calculate the image's tilt angle, and then performs rotation correction to ensure the text is horizontal, facilitating subsequent segmentation and recognition operations. The binarization unit employs an adaptive threshold segmentation algorithm, automatically determining the threshold based on the grayscale characteristics of local image regions, converting the color image into a black-and-white binary image, making the text black and the background white, further simplifying image information and improving subsequent processing efficiency.

[0033] The image segmentation module includes a region localization unit and an answer region segmentation unit. It uses a deep learning object detection algorithm to locate different functional regions of the test paper and employs contour detection and connected component analysis techniques to segment the answer region. The region localization unit, based on the YOLO (You Only Look Once) algorithm series of object detection models, analyzes the preprocessed image to accurately locate different functional regions in the test paper, such as the answer region, question region, and candidate information region. After locating the answer region, the answer region segmentation unit, according to the question type and answer format, uses contour detection and connected component analysis techniques to further segment each answer region into individual answer regions. For subjective question answer regions, the system can intelligently identify paragraph breaks, line breaks, and other information, segmenting continuous text into reasonable sub-regions, providing more accurate data for subsequent text recognition and semantic analysis.

[0034] The text recognition module includes a multi-text / multi-language recognition unit and an altered / blurred text processing unit. The multi-text / multi-language recognition unit uses a model based on convolutional neural networks (CNN) combined with an attention mechanism to recognize multiple fonts and languages. The altered / blurred text processing unit introduces generative adversarial network (GAN) technology to address text alterations and blurring that occur during students' answering process. The generator network generates possible correct text based on the features of surrounding clear text and contextual information, while the discriminator network judges the authenticity of the generated text. Through adversarial training between the two, the accuracy of the generated text is continuously optimized, thereby effectively recognizing altered or blurred text.

[0035] The semantic understanding and scoring module includes a semantic analysis unit and a scoring model unit. The semantic analysis unit uses natural language processing (NLP) technology to perform semantic analysis on the answers to subjective questions. The scoring model unit scores subjective questions from multiple dimensions by constructing a comprehensive scoring model, and scores objective questions based on the matching of the recognition results with the standard answers.

[0036] The results output module includes teacher-side and student-side display units. It provides teachers with student performance reports and answer analysis, and displays students' test paper answers, score details, and teacher comments. The final grading results are presented to both teachers and students in an intuitive and clear manner. For teachers, the teacher-side display unit provides detailed student performance reports, including scores for each question, total scores, and answer analysis, facilitating teachers' understanding of students' learning progress and teaching effectiveness. For students, they can log in to the system using their personal accounts and view their test paper answers, score details, and teacher comments and suggestions through the student-side display unit, helping students identify their learning problems and make targeted improvements.

[0037] It is understandable that, during the model training phase, the region localization unit first constructs a labeled dataset containing various test paper types and different layout styles. Each sample image is labeled with the bounding box coordinates and category labels of various functional areas. The model uses Darknet as the backbone feature extraction network and extracts multi-scale features of the image through multi-layer convolution operations. Among them, shallow features are used to capture the detailed information of the region, and deep features are used to extract the global semantic information of the region.

[0038] In the feature fusion stage, a Feature Pyramid Network (FPN) structure is used to fuse feature maps from different levels, generating feature maps suitable for detecting regions of different sizes. For each feature map, a pre-defined anchor box is used, and the model predicts the bounding box offset, confidence level, and class probability. Redundant bounding boxes are filtered using the Non-Maximum Suppression (NMS) algorithm, ultimately outputting the accurate localization result for each functional region.

[0039] This positioning method breaks through the strong dependence of traditional template-matching-based regional positioning methods on the test paper format. It can adapt to test papers of different subjects, grades, and layout styles. Even if the test paper has slight deformation, printing deviation, or local contamination, it can still maintain a high positioning accuracy.

[0040] The answer area segmentation unit, for objective questions such as multiple-choice and true / false questions, typically contains multiple independent option filling boxes or answer filling grids. The system first extracts sub-images of the objective question answer area based on the area positioning results, and then uses a connected component analysis algorithm to identify the outline of each option. By calculating the circumscribed rectangle of the outline, the boundary range of each option is determined, and then adjacent option areas are separated by combining the number of questions and the arrangement pattern of the options. For objective questions with a matrix arrangement, the system automatically divides the row and column boundaries by calculating the horizontal and vertical spacing of adjacent outlines to ensure that each option is segmented individually.

[0041] For subjective questions, the answer area is often a continuous blank writing area, and students' answers may contain multiple paragraphs and line breaks of handwritten text. The system adopts a segmentation strategy based on a combination of projection analysis and contour detection: First, the sub-image of the subjective question area is horizontally projected, and the proportion of black pixels in each row is calculated. When the proportion of multiple consecutive rows is lower than a set threshold, it is determined to be a paragraph separator or line break. At the same time, the connected components of the handwritten text are identified through contour detection. Based on the vertical distribution characteristics of the connected components, the results of the projection analysis are corrected to avoid missegmentation caused by excessive spacing between individual characters. For subjective question answer areas containing formulas or charts, the system identifies the contour boundaries of the formulas or charts through edge detection algorithms and segments them as independent sub-regions to ensure the separation of text and non-text content.

[0042] Reference Figure 2As shown, the method by which the text processing unit effectively identifies altered or blurred text during student answering of questions includes:

[0043] A multi-feature fusion algorithm for identifying altered areas is employed, combining three dimensions: pixel change rate, edge continuity, and color difference. Pixel change rate calculates the difference in grayscale values ​​between adjacent pixels within a local area; altered areas, due to multiple writing marks, exhibit significantly higher pixel value fluctuations than normal text areas. Edge continuity is determined using the Canny edge detection algorithm to extract the edges of the affected areas. Normal text edges present continuous contours, while altered areas show a significantly increased number of edge breakpoints due to intersecting strokes. Color difference is assessed for color scanned images, where altered areas may differ from normal text due to the use of different colored pens or multiple erasures, resulting in differences in the mean values ​​of the RGB color channels. This difference is identified by calculating the region's color moments.

[0044] Blurred regions are located based on sharpness assessment metrics. The second derivative of the image is calculated using the Laplacian operator. Blurred regions lack high-frequency information and have a weak derivative response, thus identifying blurred regions. The calculation formula is as follows:

[0045] V = Var(L(x,y))

[0046] Where L(x,y) is the response value of the Laplacian operator to pixel (x,y), and Var() is the variance calculation function. When V < V0 threshold When the sharpness threshold is reached, the region is determined to be a blurry region;

[0047] For the identified altered or blurred areas, the system constructs a Semantic-Aware Generative Adversarial Network (SA-GAN) to achieve text restoration and accurate recognition by combining contextual semantic information. Its core architecture includes a generator, a discriminator, and a semantic constraint module. The generator adopts an improved U-Net++ architecture, taking as input an image patch containing the altered / blurred area and contextual semantic features, and outputting a restored, clear text image. The discriminator uses a PatchGAN structure, taking as input a real, clear text image or the restored image output by the generator, and outputting an N×N probability map, where each pixel represents the probability that the corresponding area is real text. To avoid semantic conflicts between the generator output and the context, the semantic constraint module introduces semantic consistency loss, extracting semantic vectors of the generated text and context using a pre-trained BERT model, and calculating cosine similarity as a constraint.

[0048] The accuracy of the restored text image needs to be further improved through multi-model fusion recognition. This includes inputting the restored image into the CRNN (Convolutional Recurrent Neural Network) model for initial recognition, outputting the text sequence and confidence score; combining the context semantic vector output by the semantic constraint module, the recognition result is optimized by the beam search algorithm to select the text sequence consistent with the context logic; for recognition results that are still ambiguous, the system calls the character-level attention mechanism to match the features of each character in the restored image with the preset character library, calculate the similarity, and take the character corresponding to the maximum value as the final recognition result.

[0049] Understandably, for subjective questions, the semantic understanding and scoring module utilizes Natural Language Processing (NLP) technology to perform semantic understanding on the identified text and construct a comprehensive scoring model. In addition to considering keyword matching, it also combines multiple dimensions such as semantic similarity, logical coherence, and answer completeness to score student answers. The scoring model unit learns and analyzes a large number of standard answers and student answers of different levels (excellent, average, poor) to automatically determine the weight of each scoring dimension, achieving objective and accurate scoring of subjective questions. For objective questions, the score is directly given based on the matching of the recognition result with the standard answer.

[0050] The above-mentioned intelligent marking system based on image segmentation and recognition is used as follows:

[0051] Step 1: Acquire the exam paper image using the image acquisition module;

[0052] Step 2: Use the image preprocessing module to perform image enhancement, tilt correction, and binarization on the acquired images in sequence;

[0053] Step 3: Use the image cropping module to locate the test paper area and crop the answer area;

[0054] Step 4: Use the text recognition module to perform text recognition on the segmented answer area image, including multi-font and multi-language recognition and processing of altered or blurred text;

[0055] Step 5: Use the semantic understanding and scoring module to perform semantic analysis and scoring on the answers to subjective questions, and score the answers to objective questions based on the matching of the recognition results with the standard answers;

[0056] Step Six: Provide teachers and students with marking results and related analysis reports through the results output module.

[0057] In summary, the advantages of this invention are as follows: Through image segmentation, character recognition, and semantic understanding technologies, it can accurately identify answer content on various formats of exam papers, including complex handwritten fonts and corrections, avoiding misjudgments caused by non-standard formats or writing problems. This makes the scoring results more objective and accurate, better reflecting the students' true answering level. The image acquisition module and region positioning technology can adapt to exam papers of different sizes and formats, eliminating the need for specific answer sheets and facilitating flexible question creation by teachers and student responses. The character recognition module has multi-font and multi-language recognition capabilities, meeting diverse teaching needs. The fully automated marking process significantly shortens marking time, reduces teachers' workload, and allows them to devote more time and energy to teaching. Simultaneously, the results output module provides teachers and students with detailed grade reports and answer analysis, helping teachers accurately grasp teaching effectiveness, identify problems in teaching, provide targeted learning suggestions for students, and promote the improvement of teaching quality.

[0058] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. An intelligent marking system based on image segmentation and recognition, characterized in that, include: The image acquisition module is used to acquire images of the test paper and automatically adjust the scanning parameters according to the size and color mode of the test paper, and supports multi-angle scanning; The image preprocessing module includes an image enhancement unit, a tilt correction unit, and a binarization processing unit, which are used to perform brightness and contrast adjustment, deblurring, tilt correction, and binarization processing on the acquired image. The image segmentation module includes a region localization unit and an answer region segmentation unit. It locates different functional regions of the test paper based on a deep learning object detection algorithm and segments the answer region using contour detection and connected component analysis techniques. The text recognition module includes a multi-text / multi-language recognition unit and an altered / blurred text processing unit. It uses a model based on convolutional neural networks combined with attention mechanisms to recognize multi-font / multi-language text and uses generative adversarial networks to process altered / blurred text. The semantic understanding and scoring module includes a semantic analysis unit and a scoring model unit. It uses natural language processing technology to perform semantic analysis on the answers to subjective questions, constructs a comprehensive scoring model to score subjective questions from multiple dimensions, and scores objective questions based on the matching of the recognition results with the standard answers. The results output module includes a teacher-side display unit and a student-side display unit. The teacher-side display unit provides teachers with student grade reports and answer analysis, while the student-side display unit shows students their test paper answers, score details, and teacher comments.

2. The intelligent marking system based on image segmentation and recognition according to claim 1, characterized in that: The region localization unit uses a target detection model based on the YOLO series algorithms to analyze the preprocessed image and accurately locate different functional areas in the test paper, such as the answer area, the question area, and the candidate information area. After locating the answer area, the answer area segmentation unit further segments each answer area into individual answer areas according to the question type and answer format. For subjective question answer areas, it intelligently identifies paragraph breaks and line breaks, dividing continuous text into reasonable sub-regions, providing more accurate data for subsequent text recognition and semantic analysis.

3. The intelligent marking system based on image segmentation and recognition according to claim 1, characterized in that: During the model training phase, the region localization unit constructs a labeled dataset containing various test paper types and different layout styles. Each sample image is labeled with the bounding box coordinates and category labels of various functional regions. The model uses Darknet as the backbone feature extraction network and extracts multi-scale features of the image through multi-layer convolution operations. Shallow features are used to capture the detailed information of the region, while deep features are used to extract the global semantic information of the region. In the feature fusion stage, a feature pyramid network structure is adopted to fuse feature maps at different levels to generate feature maps suitable for detecting regions of different sizes. For the preset anchor boxes on each feature map, the model predicts the offset, confidence and class probability of the bounding box. Redundant bounding boxes are filtered by the non-maximum suppression algorithm, and finally the accurate localization result of each functional region is output.

4. The intelligent marking system based on image segmentation and recognition according to claim 3, characterized in that: The answer area segmentation unit, for objective questions such as multiple-choice and true / false questions, typically contains multiple independent option filling boxes or answer filling grids. First, it extracts sub-images of the objective question answer area based on the area positioning results. Then, it uses a connected component analysis algorithm to identify the outline of each option. By calculating the circumscribed rectangle of the outline, it determines the boundary range of each option. Then, combining the number of questions and the arrangement pattern of the options, it separates adjacent option areas. For objective questions with a matrix arrangement, it automatically divides row and column boundaries by calculating the horizontal and vertical spacing of adjacent outlines, ensuring that each option is segmented individually. For subjective questions, the answer area is mostly a continuous blank writing area. A segmentation strategy based on a combination of projection analysis and contour detection is adopted: First, the sub-image of the subjective question area is horizontally projected, and the proportion of black pixels in each row is calculated. When the proportion of multiple consecutive rows is lower than a set threshold, it is determined to be a paragraph separator or line break. At the same time, the connected components of handwritten text are identified by contour detection. Based on the vertical distribution characteristics of the connected components, the results of projection analysis are corrected to avoid missegmentation caused by excessive spacing between individual characters. For the answer area of ​​subjective questions containing formulas or charts, the contour boundaries of formulas or charts are identified by edge detection algorithms and segmented as independent sub-regions to ensure the separation of text and non-text content.

5. The intelligent marking system based on image segmentation and recognition according to claim 4, characterized in that: The method for the text alteration / blurring processing unit to effectively identify altered or blurred text during student answering includes: A multi-feature fusion algorithm for identifying tampered areas is adopted, which combines three dimensions—pixel change rate, edge continuity, and color difference—to determine the tampered areas. Based on the sharpness assessment index, blurry regions are located, and the second derivative of the image is calculated using the Laplacian operator. Blurry regions lack high-frequency information and have a weak derivative response, thus identifying blurry regions. For the identified altered or blurred areas, the system constructs a semantic-aware generative adversarial network, which combines contextual semantic information to achieve text repair and accurate recognition; The accuracy of the restored text image is further improved through multi-model fusion recognition.

Images