Intelligent homework test paper marking and mistake question induction method, system and electronic device

Abstract

The application relates to the field of data recognition, and particularly discloses a method and system for intelligent homework paper marking and wrong question induction, and electronic equipment. The method comprises the following steps: collecting image data of homework or a paper, pre-processing the collected image data, detecting the circumscribed polygon coordinates of the text in the image through PANet after the pre-processed image, recognizing the text line picture through CRNN, and outputting a text string; for the science category, decomposing and scoring the problem solving steps through a symbol calculation engine, verifying the vector relationship through an image feature recognition engine; for the art category, performing semantic scoring through an NLP deep analysis engine; and generating a wrong question set corresponding to each subject according to the marking results of each subject. Through task cutting, image processing and deep learning, the application reduces the dependence of a large model on the whole process, improves the system response speed and reduces the consumption of hardware resources under the premise of ensuring the marking accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data recognition, in particular to an intelligent homework test paper marking and mistake summary method, system and electronic device. BACKGROUND

[0002] The current homework and test paper marking mainly relies on manual completion, and teachers need to invest a lot of time in checking, scoring and error marking for each question, especially in the large-scale teaching scene, not only the marking efficiency is low, but also the scoring consistency is insufficient due to fatigue, the feedback cycle is long, and it is difficult to meet the needs of students to correct and review in a targeted manner. The current technology in the field of intelligent homework and test paper marking includes systems based on large language models, which are prone to factual errors in actual application, such as in the recognition of academic concepts, generative models often misstate "reduction division" as the main way of cell replication, leading to students' knowledge understanding deviation. High computing cost is another core obstacle. Taking the VLM2Vec-V2 model as an example, its training requires 8xH100 cluster support, and the inference can be compressed to a single A100 chip, but it is difficult for primary and secondary schools to bear the procurement and maintenance costs of such high-end computing equipment. This resource dependence makes it difficult for products to land, especially in areas where educational resources are relatively weak, the popularization of intelligent marking systems is severely limited. The mistake set generation link also has technical pain points, and the mistake sets generated by existing systems are mostly simple lists of questions, lacking in-depth correlation analysis of the root causes of errors. At the same time, the existing system lacks explainability, and the decision-making process of the automatic marking result is not transparent, resulting in low trust of teachers and students in the scoring basis, further hindering the landing of technology. SUMMARY

[0003] The purpose of the present application is to overcome the defects of the prior art, and to provide an intelligent homework test paper marking and mistake summary method, as well as an intelligent homework test paper marking and mistake summary system and an electronic device, to solve the problems raised in the background art.

[0004] The intelligent homework test paper marking and mistake summary method of the present application comprises the following steps:

[0005] Collecting image data of homework or test papers, and pre-processing the collected image data, which includes distortion correction and image enhancement;

[0006] Transmitting the pre-processed image data to PANet, detecting the bounding polygon coordinates of the text in the image through PANet, and then transmitting it to CRNN, recognizing the text line picture through CRNN, and outputting the text string;

[0007] The recognized text is divided into liberal arts category and science category, and is distributed to the corresponding processing module through subject routing. The problem solving steps of the science category are scored through a symbol calculation engine, and the vector relationship is verified through an image feature recognition engine. The liberal arts category is scored through an NLP deep analysis engine.

[0008] According to the review results of each subject, the error causes are analyzed respectively, and the error question sets corresponding to the subjects are generated according to the error cause analysis results.

[0009] Further, the PANet detects the bounding polygon coordinates of the text in the image, specifically including:

[0010] The PANet receives the normalized image after distortion correction and image enhancement, first extracts features from the image to generate feature maps of different levels from low to high, then uses a feature pyramid structure to fuse multi-scale features to form a fused feature map with details and semantics. During the detection process, the PANet classifies the fused feature map at the pixel level to distinguish text and background areas, and simultaneously regresses the bounding box parameters of the text area. For printed questions, the output is a quadrilateral bounding coordinate, and for handwritten answers, the output is a polygon coordinate with 6-10 vertices. After detection is completed, overlapping and redundant detection boxes are removed, and areas with small areas or abnormal aspect ratios are filtered. The final output of the text bounding polygon coordinates is used to accurately crop the text line area from the image.

[0011] Further, the text line picture is recognized by CRNN, specifically including:

[0012] The CRNN first normalizes the text line picture. The convolutional layer extracts local features of the text line by multiple convolution and pooling operations to generate a feature sequence. The recurrent layer performs time series modeling on the feature sequence output by the convolutional layer, and uses a bidirectional propagation mechanism to capture the dependency relationship between the front and back strokes, adapting to the common continuous writing, tilting and irregular writing forms of student handwriting. The transcription layer converts the sequence features output by the recurrent layer into a text string.

[0013] Further, for the science category, the problem solving steps are scored through a symbol calculation engine, specifically including:

[0014] When the symbolic computing engine receives the science text data distributed by the subject routing, the structured reference answer data packet associated with the question is called synchronously. The symbolic computing engine performs structured analysis on the reference answer to extract the benchmark checking framework. The student's answer unit is matched with the core step framework of the reference answer. If the student uses the same solution path as the reference answer, the reference answer is directly checked point by point. If the student uses other solutions, the "logical integrity" core standard extracted based on the reference answer is checked. The scoring node weight preset in the reference answer is embedded into the dynamic weight mechanism to score the student's answer steps. The student's calculation result is compared with the standard result of the reference answer for equivalence. If the deviation exceeds the preset fault tolerance threshold of the reference answer, the deviation source is located in combination with the intermediate step mark in the reference answer. A structured report including the difference mark between the student's steps and the reference answer is generated.

[0015] Further, the vector relationship is verified by the image feature recognition engine, specifically including:

[0016] When the image feature recognition engine receives the science graphical data distributed by the subject routing, the reference answer vector feature data packet associated with the question is called synchronously. The image feature recognition engine first normalizes and extracts features from the student's answer graphical area image to construct a structured vector data set. Then, according to the vector specification library of the reference answer, multi-dimensional comparison is carried out, and at the same time, the numerical annotations in the graph are cross-verified with the step data output by the symbolic computing engine. After verification, a vector verification report including the deviation type, deviation position and reference answer comparison explanation is generated.

[0017] Further, semantic scoring is performed by the NLP deep analysis engine, specifically including:

[0018] When the NLP deep analysis engine receives the arts text data distributed by the subject routing, the arts reference answer data packet associated with the question is called synchronously. The reference answer is analyzed and extracted to form a semantic scoring benchmark framework.

[0019] The student's answer text is first converted into a calculable semantic feature sequence through vector conversion, and then a context semantic vector is generated and compared with the semantic vector of the reference answer. The student's answer text is scored,

[0020] The scoring dimensions include semantic coverage dimension, logical coherence dimension and language specification dimension. After scoring, a structured report including semantic deviation details is generated to mark the differences with the reference answer.

[0021] Further, according to the results of each subject, the error causes are analyzed respectively, specifically including:

[0022] The humanities NLP deep analysis engine is classified by subject through a data aggregation interface, and reports generated by the science symbol calculation engine and the image feature recognition engine are classified and collected, forming a structured data set including topic ID, subject label, score details, error labeling, and original answer data;

[0023] For science subjects, based on the preset error type system, the step errors marked by the symbol calculation engine and the vector verification abnormalities of the image feature recognition engine are cross-compared, associated with the science error knowledge base, and the error root cause is located. For multi-dimensional errors in the same question, the cause and effect relationship is analyzed;

[0024] For humanities subjects, based on the semantic deviation points extracted by the NLP deep analysis engine, combined with the text logical relationship graph and the humanities knowledge base, hierarchical classification is performed. For objective question errors, match the knowledge point errors, and for subjective question errors, decompose from three dimensions of semantic coverage, logical coherence, and language standardization to identify error types;

[0025] During the error analysis process of each subject, the student's historical answer data is associated synchronously, and the repetitive errors are identified through longitudinal comparison and marked as weak links that need to be strengthened. A structured analysis report is generated for each error question, including error location, error type, associated knowledge point, and error level. According to the error cause analysis result, the error question set corresponding to the subject is generated.

[0026] The intelligent homework test paper marking and error question induction system includes:

[0027] Multi-source acquisition unit: used to acquire homework / test paper images through mobile phone camera and error question printer camera;

[0028] Preprocessing center: used for distortion correction and image enhancement;

[0029] Intelligent marking engine: used for text detection and character recognition, where text detection is based on PANet to extract the bounding polygon coordinates of text in the image and crop the text line area; character recognition is through CRNN to convert text line images into text strings, filter invalid characters and check semantic relevance;

[0030] Subject routing distribution module: used to divide text into science and humanities, and distribute to the corresponding marking module;

[0031] The sub-module for science subject review includes a symbol calculation engine and an image feature recognition engine, the symbol calculation engine is used for analyzing the reference answer step logic, comparing the formula application, calculation process and result of the student's answer, and outputting the step score and difference mark; the image feature recognition engine is used for extracting the graphic vector feature and verifying the standard matching with the reference answer; the sub-module for arts subject review includes an NLP deep analysis engine, the NLP deep analysis engine scores the text from the dimensions of semantic coverage, logical coherence and language standardization;

[0032] The error cause analysis module is used for collecting the review results of the text and science subjects, associating the error types with the subject knowledge points, locating the error sources, and generating a structured error cause report;

[0033] The error question generation module is used for integrating the question image, student answer and error cause analysis result, and generating an error question set according to the subject classification.

[0034] An electronic device includes one or more processors, which execute programs stored in a memory to perform the intelligent homework paper review and error question induction method described above.

[0035] A computer-readable storage medium stores a computer program, which is executed by a processor to implement the intelligent homework paper review and error question induction method described above.

[0036] The beneficial effects of the present application are as follows:

[0037] The present application effectively reduces the dependence on the whole process of large models through the architecture design of task cutting and module division, deeply integrates image processing technology, deep learning network and special engine, on the one hand, with the help of distortion correction, image enhancement preprocessing, and accurate text detection and recognition of PANet and CRNN, a high-quality data foundation is laid for subsequent review; on the other hand, the classification processing of the text and science subjects is realized through subject routing, the step score and vector relationship verification are completed by relying on the symbol calculation engine and the image feature recognition engine for the science subject, and the semantic score is performed by the NLP deep analysis engine for the arts subject, which significantly improves the pertinence and accuracy of the review and avoids the limitations of a single model.

[0038] This architecture greatly reduces the hardware resource consumption and computing power cost, breaks through the bottleneck of high-cost deployment of traditional large models, and is more suitable for the landing needs of large-scale teaching scenes such as schools. At the same time, the error cause analysis and subject-specific error question set generation based on multi-dimensional review results can accurately associate knowledge points and error sources, form a complete teaching closed loop from review to review, and help students improve their targetedness. In addition, the structured reports output by each link clearly present the scoring basis and error details, enhancing the explainability of the system, improving the trust of teachers and students in the automatic review results, and effectively promoting the practicalization of intelligent review technology in teaching scenarios. Attached Figure Description

[0039] Figure 1 This is a flowchart of the intelligent homework test paper grading and error summary method of the present invention;

[0040] Figure 2 This is an architecture diagram of the intelligent homework test paper grading and error summary system of the present invention. Detailed Implementation

[0041] This application discloses an intelligent method for grading homework and test papers and summarizing incorrect answers. First, the terms involved in this invention are explained as follows:

[0042] PANet: Path Aggregation Network, used for multi-scale feature fusion and bounding box regression in text detection.

[0043] CRNN: short for Convolutional Recurrent Neural Network, is used for text recognition. It combines convolutional layers, recurrent layers, and transcription layers to convert text lines of images into strings.

[0044] LSTM: Long Short-Term Memory, the core structure of the recurrent layer in CRNN, used to capture the temporal dependencies of text sequences.

[0045] CTC: Short for Connectionist Temporal Classification, it is the loss function of the transcription layer in CRNN, which solves the problem of variable-length alignment between input features and output labels.

[0046] NLP: Natural Language Processing, used for semantic analysis and scoring of humanities texts.

[0047] OCR stands for Optical Character Recognition, which is used to convert text in homework / exam paper images into processable text strings.

[0048] The intelligent homework test paper grading and error summary method of this application, such as Figure 1 The process includes the following steps:

[0049] S1. Collect image data of assignments or test papers, and preprocess the collected image data, including distortion correction and image enhancement.

[0050] S2. The pre-processed image data is transmitted to the PANet, the bounding polygon coordinates of the text in the image are detected through the PANet, and then transmitted to the CRNN, the text line picture is recognized through the CRNN, and the text string is output;

[0051] S3. The recognized text is divided into liberal arts category and science category, and distributed to the corresponding processing module through subject routing;

[0052] S4. For the science category, the problem solving steps are decomposed and scored through the symbol calculation engine, and the vector relationship is verified through the image feature recognition engine. For the liberal arts category, semantic scoring is performed through the NLP deep analysis engine;

[0053] S5. According to the review results of each subject, the error causes are analyzed respectively, and the error question sets of the corresponding subjects are generated according to the error cause analysis results.

[0054] Among them, for S1. Collecting image data of homework or test paper, the collected image data is pre-processed, and the pre-processing includes distortion correction and image enhancement;

[0055] In specific implementation, collecting image data of homework or test paper supports image collection through student personal smart phone camera or high-definition camera deployed in school.

[0056] The collected image data is pre-processed, and the pre-processing includes distortion correction and image enhancement; in specific implementation, for the image collected by the mobile phone, the distortion correction first calls the camera parameter library built-in the mobile phone, and the radial distortion and tangential distortion existing in the mobile phone collected image are corrected by matching the intrinsic matrix and distortion coefficient and applying the distortion correction model formula.

[0057] Specifically, a camera parameter database of different mobile phone models is constructed in advance, and the parameters of different mobile phone cameras are calibrated. By making the mobile phone to be calibrated shoot the checkerboard calibration board image in different poses, the checkerboard corner point coordinates are extracted by using the calibration algorithm, and the intrinsic matrix (including focal length parameters fx, fy and principal point coordinates cx, cy) and distortion coefficient (including radial distortion coefficients k1, k2, k3 and tangential distortion coefficients p1, p2) of the camera are calculated. These parameters are associated with the mobile phone model and stored in the database. When collecting homework images using a mobile phone, the corresponding intrinsic matrix and distortion coefficient are automatically matched by reading the device identification information to correct the distortion of the image.

[0058] In practice, the original coordinates of each pixel in the image are first converted into normalized coordinates (x, y) in the camera coordinate system. Then, the distortion correction model formula is applied to correct radial and tangential distortion. Finally, the corrected coordinates are converted back to the image coordinate system to generate a corrected image that eliminates edge curvature and image distortion, providing basic image data with minimal geometric distortion for subsequent image enhancement and text detection and recognition.

[0059] Among them, the distortion correction model formula includes the radial distortion correction formula and the tangential distortion correction formula;

[0060] The formula for radial distortion correction is: x'=x*(1+k1*r²+k2*r) 4 +k3*r 6 ), y'=y*(1+k1*r²+k2*r 4 +k3*r 6 );

[0061] The formula for tangential distortion correction is:

[0062] x''=x'+2*p1*x*y+p2*(r²+2*x²), y''=y'+p1*(r²+2*y²)+2*p2*x*y;

[0063] In the formula, x and y are the normalized coordinates of the pixel in the camera coordinate system, k1 is the first-order radial distortion coefficient, k2 is the second-order radial distortion coefficient, and k3 is the third-order radial distortion coefficient. k1, k2, and k3 are used to correct the nonlinear offset of the pixel caused by the lens curvature as it gets farther away from the principal point. x' and y' are the normalized coordinates after radial distortion correction. r² is the square of the distance from the pixel to the principal point, calculated as r² = x² + y², used for nonlinear calculation of radial distortion correction. The tangential distortion correction formula is a further calculation based on the radial distortion correction formula. p1 and p2 are tangential distortion coefficients used to correct the offset of the pixel in the tangential direction caused by lens mounting tilt or optical center offset. x'' and y'' are the final normalized coordinates after radial and tangential distortion correction. The corrected normalized coordinates are obtained through x'' and y'', and then transformed back to the image coordinate system to complete the distortion correction.

[0064] Furthermore, in a preferred embodiment, a distortion correction algorithm based on spatial adaptive nonlinear mapping is employed to address radial and tangential distortions present in images captured by the mobile phone, achieving pixel-level precise correction. The specific distortion correction algorithm formula is as follows: In the formula, (x,y) are the original coordinates of the image pixels in the normalized imaging plane, ( , ( ) represents the corrected coordinates. , where r is the Euclidean distance from the pixel to the center of the image;α is the distortion strength factor, used to quantify the overall influence degree of radial distortion, α The larger the value is, the higher the correction strength is. β is the spatial attenuation coefficient, controlling the decay rate of the distortion correction effect with the distance from the center, β The larger the value is, the more significant the edge distortion correction is. gamma is the tangential coupling coefficient, used to correct the coupling effect of tangential distortion and radial distortion caused by lens assembly eccentricity; tanh(·) is the hyperbolic tangent function, which simulates the nonlinear change law of distortion with distance through its nonlinear characteristics; in the formula, α ⋅tanh( β ⋅r)⋅x and α ⋅tanh( β ⋅r)⋅ y terms are used to correct radial distortion, γ⋅(3x 2 y+y 3 ) and γ⋅(x 3 +3xy 2 ) terms are used to correct tangential distortion and the coupling effect of the two.

[0065] By constructing a distortion correction algorithm, pixel-level accurate correction is realized, radial and tangential distortion caused by lens curvature and assembly eccentricity is eliminated, and a geometrically regular image is output, providing high-quality input for subsequent image enhancement, text detection and character recognition, avoiding text positioning or recognition errors caused by distortion, and ensuring the accuracy of intelligent review.

[0066] For the images collected by the error question printer, the light and slight perspective distortion caused by fixed focal length is restored to a regular rectangle based on the four vertex coordinates detected on the paper edge.

[0067] First, the edge of the pre-processed error question printer collected image is detected, the edge contour in the image is extracted, and the edge gap is eliminated. Then, the largest connected region contour is retained through contour area screening, and this contour is the edge contour of the work paper. The screened contour is polygonally approximated to obtain a quadrilateral contour composed of four vertices. The clockwise ordering of the four vertices is determined by calculating the distance and angle relationship between the vertices. The size of the target regular rectangle is set according to the actual size of the work paper, and the size is calculated according to the width-height ratio of the original paper contour to maintain the image proportion. Then, the tilted or trapezoidal distorted paper region is mapped to the target regular rectangle region, and finally a geometrically regular rectangular image is output, eliminating the influence of perspective distortion on subsequent text detection.

[0068] The image enhancement link first processes the uneven illumination problem, decomposes the illumination component and the reflection component of the image, suppresses the overexposure of the strong light area while improving the details of the shadow area; then the color or grayscale image is converted into a black and white text image, highlighting the contrast between the text and the background; for the Gaussian noise and salt and pepper noise generated during acquisition, Gaussian filtering and median filtering are applied respectively to remove them, and at the same time the text edge is sharpened to strengthen the outline clarity of the handwriting strokes, and finally a standardized image that meets the OCR recognition standard is output, providing high-quality input for subsequent text detection and recognition.

[0069] For S2, the pre-processed image data is transmitted to PANet, which detects the bounding polygon coordinates of the text in the image, and then transmitted to CRNN, which recognizes the text line picture and outputs the text string; in specific implementation, the standardized image data output after preprocessing is transmitted to the intelligent marking module through an internal data interface, and the module first performs text detection and OCR, which are completed by PANet and CRNN respectively, to obtain the text string. The intelligent marking module uses PANet to detect the bounding polygon coordinates of the text in the image. Specifically, the PANet of the intelligent marking module receives the standardized image after distortion correction and image enhancement, first extracts features from the image to generate feature maps at different levels from low to high. The low-level features retain local information such as text edges and stroke details, and the high-level features capture global semantic information of the text area. Then, the multi-scale features are fused using a feature pyramid structure, the high-level semantic features are passed down from top to bottom, and the low-level detail features are supplemented through horizontal connection, forming a fused feature map that combines details and semantics, effectively adapting to different sizes of text in the assignment.

[0070] For the common tilting, continuous writing and curved strokes of student handwriting, the model introduces a large amount of data including such samples during the training stage, simulates real writing variations through data enhancement methods such as random rotation and elastic deformation, and enhances the perception of irregular text forms. In the detection process, PANet classifies the fused feature map at the pixel level to distinguish between text and background areas, and also regresses the bounding box parameters of the text area. For regular text such as printed questions, the model outputs the coordinates of the quadrilateral bounding box, and for tilted or curved handwriting answers, it outputs the coordinates of a polygon with 6-10 vertices to accurately enclose the text edge. Even if there are erasing marks or assignment line interference, the model can still accurately define the text range. After detection is complete, overlapping and redundant detection boxes are removed, and areas with abnormally small areas or aspect ratios are filtered out. The final output of the text bounding polygon coordinates is used to accurately crop the text line area from the image, providing a high-quality input area for subsequent CRNN text recognition.

[0071] The PANet adopted by the text detection network of the present application takes ResNet-50 as a backbone, and after four-scale fusion of the output features by FPN, semantic features are extracted by a 3x3 convolution layer (256 channels), and finally a 1x1 convolution is performed to respectively output a 1-channel text probability map and a 6-channel kernel probability map.

[0072] Specifically, the ResNet-50 backbone network of the PANet extracts image features layer by layer through four residual convolution stages to generate multi-level feature maps from C2 to C5, wherein C2 retains the detailed features of 1 / 4 of the original image resolution, and C5 is compressed to 1 / 32 resolution to capture global semantics. In the four-scale fusion process of the FPN, C5 is first reduced to 256 channels by a 1x1 convolution as the initial features, and then sequentially connected with C4, C3 and C2 through top-down bilinear interpolation. After the 3x3 convolution is performed on the fused features of each layer to eliminate the aliasing effect of the upsampling, a fusion feature pyramid from P2 to P5 is formed, and each layer maintains a 256-channel dimension to balance the details and semantics. The 3x3 convolution layer performs deep convolution operation on the fused feature map, enhances the local receptive field through a convolution kernel with a dilation rate of 1, and strengthens the discrimination between the text edge and the background. The output semantic feature map is sent to two parallel 1x1 convolution branches: the text probability map branch adopts a sigmoid activation function to output the pixel-level text existence probability, and the kernel probability map branch outputs 6 parameters corresponding to each pixel through regression, representing the offset, angle and vertex distribution coefficient of the polygon bounding box, respectively, to realize accurate parameterization description of the printed quadrilateral and the multi-vertex polygon of the handwritten text.

[0073] The CRNN is used to recognize the text line picture to output a text string. In specific implementation, the text line picture accurately cropped by the PANet is input into the CRNN for character recognition. First, the text line picture is normalized to uniformly adjust the image size to a fixed height, and the width is scaled in proportion to maintain the text form, and irrelevant color interference is removed through grayscale to ensure the consistency of the input features. The CRNN is composed of a convolution layer, a recurrent layer and a transcription layer. The convolution layer adopts a lightweight convolution structure to extract local features of the text line through multiple convolution and pooling operations to generate a feature sequence, effectively capturing the edge contour, thickness variation of the handwritten strokes and the font features of the printed text. The recurrent layer adopts a bidirectional LSTM to model the time sequence of the feature sequence output by the convolution layer, and uses a bidirectional propagation mechanism to capture the dependency relationship between the front and rear strokes, and adapt to the common continuous writing, inclination and irregular writing forms of student handwriting. The transcription layer converts the sequence features output by the recurrent layer into a text string based on the CTC loss function, solving the problem of indefinite length alignment between input features and output labels in character recognition, especially suitable for scenarios with overlapping strokes and uneven intervals in handwritten text.

[0074] In the training stage, a text line image dataset covering printed subject text and handwritten answer text of students of different grades is used, including Chinese characters, English, numbers, punctuation marks, and special symbols for science, etc. Through data enhancement methods such as random scaling, slight rotation, adding Gaussian noise, and contrast jitter, image variations caused by different shooting lights, paper materials, and writing habits are simulated. During the model training process, the sample weight of high-frequency error characters is increased in combination with the distribution characteristics of discipline characters, so as to improve the recognition accuracy. After the recognition is completed, the output text string is filtered by a preset character dictionary to remove invalid noise characters, and a preliminary check is performed in combination with a discipline field table to ensure the semantic relevance of the output text and the homework content, thereby providing accurate text input for subsequent classification and grading of science and arts.

[0075] The CRNN adopted by the character recognition network of the present application is composed of 5 layers of convolution + maximum pooling in the convolution part, and an output of 8x(W / 16)x512 feature sequence is obtained. After modeling by two layers of bidirectional LSTM with 256 hidden units respectively, a variable-length string is output by CTC decoding.

[0076] Specifically, the 5 layers of convolution in the CRNN convolution part use convolution kernel sizes of 3x3, 3x3, 3x3, 2x2, and 2x2 in sequence, the convolution step size of the first three layers is set to 1x1, and the step size of the last two layers is adjusted to 1x2 to adapt to the compression requirement of the horizontal features of the text line; a batch normalization layer and a ReLU activation function are connected after each layer of convolution to suppress gradient disappearance and enhance nonlinear feature expression. The maximum pooling layer uses a 2x2 pooling kernel with a step size of 2x2, which is followed by the 1st, 2nd, 3rd, and 5th layers of convolution, respectively. The edge contour features of the handwriting strokes are enhanced by selectively retaining the local maximum value. After 5 layers of convolution and pooling operation, the height of the input text line image is gradually compressed from the normalized fixed value to 8, the width is scaled by a ratio of 1 / 16 of the original width, the number of channels is finally fixed to 512 through the number of convolution kernels in each layer, and a feature sequence of 8x(W / 16)x512 is formed. This dimension design not only retains enough horizontal temporal information for the recurrent layer modeling, but also carries rich stroke detail features through a high number of channels. When processing the feature sequence, the forward LSTM captures the stroke order dependence from left to right, and the backward LSTM perceives the context association of connected strokes from right to left. The 256 hidden units in each layer dynamically filter noise features through the gating mechanism, and the hidden state output is spliced to form a 512-dimensional time sequence feature vector.

[0077] After completing text detection and OCR, the intelligent marking module internally calls NLP to cut the questions, distinguishing between question descriptions and student answers. First, a double-layer cutting logic is constructed based on the semantic features and format features of the text sequence. The rule engine is used to extract the question number in the text as the initial cutting anchor point. For texts without explicit question numbers, the spatial layout information reflected by the text bounding polygon coordinates output by PANet is used to analyze the line spacing changes and paragraph spacing features to identify the physical separation boundaries between questions, such as the line spacing between the question description area and the student answer area, which is usually greater than the internal line spacing of the question. This helps to distinguish between different question units. Then, a pre-trained text classification model is used to classify the cut-out text segments semantically to identify the boundaries between "question requirements" and "student answers". The model learns the semantic difference features between "stem description" and "answer content" to separate the question core content and student answer part. For science questions containing graphics and formulas, the graphics region coordinates labeled by PANet and the formula text recognized by CRNN are associated to ensure that the question description and corresponding answer content associated with graphics are classified into the same question unit. After cutting, a structured question unit list containing question numbers, question texts, corresponding answer texts, and associated image region indexes is generated. The internal verification mechanism checks the semantic coherence of adjacent question units to avoid cutting errors and provides structured input for accurate classification during subsequent subject routing distribution.

[0078] For S3, the recognized text is divided into arts and science categories. In specific implementation, the recognized text includes question descriptions and student answer content. First, the subject identification words, question type features, and core content keywords in the question are extracted. Based on the pre-constructed arts and science feature library, the arts and science categories are distinguished by keyword matching and feature vector comparison. The arts categories include Chinese, English, history, and politics, and the science categories include mathematics, physics, chemistry, and biology.

[0079] The text is distributed to the corresponding processing module through subject routing. In specific implementation, based on the arts and science classification results, a preset subject-module mapping table is called to achieve accurate routing. Arts category text is automatically distributed to the NLP deep analysis engine module, and science category text is simultaneously distributed to the symbolic computation engine and image feature recognition engine module. During routing, the text data is structured and packaged, including question ID, subject label, original recognized text, student answer text, and associated image region index, to ensure that the target module can directly obtain all the required information.

[0080] For S4. For the science category, the problem-solving step decomposition score is calculated by the symbolic computation engine. In specific implementation, when the symbolic computation engine receives the science text data distributed by the subject routing, it synchronously calls the structured reference answer data package associated with the question, which includes the core step logic chain of the reference answer, key formula annotation, scoring node weight distribution, and standard expression elements.

[0081] The engine first performs structured analysis on the reference answer to extract the benchmark verification framework: disassembles the core steps of the answer, clearly defines the formula application conditions corresponding to each step (such as "Ohm's law only applies to pure resistance circuit scenarios"), and defines the result tolerance range. These framework elements form a dual verification system with the formula theorem library of the subject.

[0082] When decomposing the student's answer steps, the engine matches the student's answer units with the core step framework of the reference answer through step semantic alignment algorithms. If the student uses the same solution path as the reference answer, the engine directly checks the formula writing, logical derivation order, and result requirements in the reference answer point by point. For example, if the reference answer clearly states "Lorentz force formula F=qvB" is required, and the student incorrectly writes F=qE, the engine directly marks the formula as incorrect. If the student uses other reasonable methods (such as using geometric methods instead of algebraic methods for mathematical problems), the engine checks based on the "logical integrity" core standard extracted from the reference answer to avoid false positives due to step form differences.

[0083] Among them, the "logical integrity" core standard extracted from the reference answer is as follows: This standard is based on the inherent laws of the subject and the core requirements of the question, covering four dimensions: problem modeling rationality, step coherence, core logic closed loop, and conclusion validity. When checking, the engine decomposes the student's solution into several logical nodes and compares them with the core logical elements extracted from the reference answer (such as the basic framework of "problem transformation, principle application, process derivation, and result verification").

[0084] For example, in a math problem, if the student uses a geometric method instead of the algebraic method provided in the reference answer, the engine checks whether the conditions for applying the geometric theorem match the given conditions in the problem, whether the addition of auxiliary lines effectively establishes the relationship between the known and unknown quantities, whether the derivation of the angle and length relationship has any logical gaps, and whether the final conclusion is consistent with the constraints in the problem statement. In a physics problem, if the student uses the law of conservation of momentum instead of Newton's law of motion to solve the problem, the engine checks whether the prerequisites for the conservation of momentum (such as the system having zero external force) are consistent with the problem, whether the analysis of momentum changes during the process fully covers all participating objects, whether the unit conversion of physical quantities in the derivation process is standardized, and whether the result is consistent with the physical actual situation (such as the direction of velocity and the trend of energy conservation). At the same time, the engine calls the discipline-specific principle library for auxiliary verification, such as the axiomatic system in mathematics, the application range of conservation laws in physics, and the reaction rules in chemistry, to ensure that the logical basis of the student's solution does not deviate from the core principles of the discipline. If the student's solution has gaps (such as skipping key condition verification), gaps (such as no causal relationship between derivation steps), or contradictions (such as the conclusion conflicting with the premise), it is marked as logically incomplete. If all core logical nodes meet the discipline rules and form a closed loop, even if the step form is different from the reference answer, it is still considered a valid solution, ensuring compatibility with diverse and reasonable solutions.

[0085] In the scoring stage, the engine embeds the scoring node weights preset in the reference answer into a dynamic weight mechanism; for example, if the reference answer indicates that "formula selection is a core scoring point," the corresponding step weight is automatically increased, and if the student does not correctly use the formula, the corresponding weight value is directly deducted. For additional derivation details that are not required by the reference answer but are reasonably supplemented by the student (such as additional explanations of unit conversion in a physics problem), additional points are added through positive reinforcement rules. In the calculation result verification stage, the engine compares the student's calculation result with the standard result in the reference answer for equivalence, allowing for minor deviations due to rounding. If the deviation exceeds the tolerance threshold preset in the reference answer, the engine locates the source of the deviation (such as "step 4 data entry value error" or "formula transformation coefficient omission") based on the intermediate step annotations in the reference answer. After scoring is complete, the engine generates a structured report that includes the differences between the student's steps and the reference answer (such as "missing formula derivation process in reference answer step 3" or "calculation result consistent with reference answer but step logic has gaps"). This provides accurate references for subsequent error analysis, ensuring that the scoring criteria align with teaching standards while accommodating students' diverse and reasonable problem-solving expressions.

[0086] The vector relationship is verified by an image feature recognition engine. In specific implementation, when the image feature recognition engine receives the science and technology graph data distributed by the subject routing, the reference answer vector feature data packet associated with the question is synchronously called, the data packet includes the standard definition of the vector elements in the reference answer (such as the geometry question “△ABC, AB = AC and ∠A = 60°”, the mechanics question “the friction force f is opposite to the direction of the pulling force F1”), the space relationship specification (such as “the magnetic induction line direction is perpendicular to the paper surface inward” and “the voltage table and the resistance are connected in parallel in the circuit diagram”) and the labeling format requirement (such as “the length unit needs to be labeled on the right side of the numerical value” and “the vector arrow length of the force is proportional to the size”). The engine first performs special processing on the graph area image of the student's answer, defines the effective area of the graph based on the polygon coordinate output by PANet, eliminates the interference such as the work line and the erasing trace, strengthens the edge contour of the vector features such as the line segment, the angle, the arrow and the symbol, and then normalizes the graph to the size and the coordinate system matched with the reference answer vector specification, so as to ensure the consistency of the comparison benchmark. In the feature extraction stage, the basic vector elements such as the straight line and the curve are recognized, the vertex coordinates are located, the numerical labels (such as “3 cm” and “45°”) and the symbol identifiers (such as “⊥”, “∠” and “,”) in the graph are recognized, and the structured vector data set including the element type, the space parameter (length, angle and direction vector) and the labeling information is constructed.

[0087] During the verification process, the image feature recognition engine compares the student's answer with the reference answer based on the vector specification library. For geometry problems, the engine verifies the length ratio of line segments (e.g., if the reference answer is "AB:BC=2:1", the engine checks if the pixel distance ratio of the corresponding line segments in the student's answer is within ±5% tolerance), the angle value (e.g., if the reference answer is "∠B is a right angle", the engine checks if the pixel angle of the corresponding angle in the student's answer is within 85°-95°), and the position relationship (e.g., if the reference answer is "CD⊥AB", the engine verifies the perpendicular relationship by calculating the slope). For mechanics problems, the engine focuses on the consistency of the direction of the vector (e.g., if the reference answer is "gravity G is vertically downward", the engine checks if the deviation of the direction vector of the student's arrow is more than 10°), and the size of the vector (e.g., if the reference answer is "F1=3F2", the engine checks if the ratio of the two force values in the student's answer is matched). For circuit diagrams, the engine checks the connection topology of the elements (e.g., if the reference answer is "slide rheostat is connected in series with the bulb", the engine checks if the connection of the elements in the student's answer forms a series circuit). At the same time, the engine cross-verifies the numerical value in the student's answer with the step data output by the symbolic calculation engine to ensure that the numerical value in the student's answer is consistent with the numerical value in the step (e.g., if the step calculation is "DE=5cm" and the numerical value in the student's answer is "DE=5cm", the engine checks if they are consistent). After the verification is completed, the engine generates a vector verification report including the type of deviation (e.g., "angle deviation 15°" "force direction error" "line segment ratio inconsistency"), the location of the deviation (precise to the pixel coordinate range of the graph), and the reference answer comparison explanation (e.g., "not consistent with the reference answer 'EF∥GH'"). The report provides accurate evidence for error analysis in the graph dimension. Through the combination of traditional image processing technology and reference answer specification, the engine realizes collaborative verification with the symbolic calculation engine, and improves the comprehensiveness and accuracy of science grading.

[0088] For liberal arts, the NLP deep analysis engine is used for semantic scoring. In specific implementation, when the NLP deep analysis engine receives the liberal arts text data distributed by the subject routing, it synchronously calls the associated liberal arts reference answer data package, which includes core scoring points (e.g., key information for reading comprehension, argument elements for essay writing), semantic coverage range (e.g., three core viewpoints for discussion questions), logical level requirements (e.g., structure specification of "argument-point-evidence-conclusion" for argumentative essays), and language specification standards (e.g., interpretation of virtual words in classical Chinese, correct forms of English grammar).

[0089] The NLP deep analysis engine first analyzes the structure of the reference answer, extracts the semantic scoring benchmark framework, and breaks down the necessary core semantic units (e.g., "three historical significance of the Xinhai Revolution" for history questions), determines the weight proportion of each unit (e.g., 40% for argument points and 30% for evidence in essay writing), and defines the semantic equivalence range (e.g., allowing synonymous expressions but including core keywords). These framework elements form a double-checking system with the liberal arts knowledge base (e.g., a classical Chinese dictionary, a historical event timeline).

[0090] For student answer text, the engine first converts the text into a computable semantic feature sequence through word segmentation, part-of-speech tagging, and semantic vector conversion, generates a context semantic vector, and compares it with the semantic vector of the reference answer.

[0091] The scoring process is carried out from multiple dimensions, including semantic coverage, which calculates the semantic overlap between the student's answer and the reference answer core score points, such as reading comprehension questions that include 5 key information points, and the student covers 3 points, which is scored according to 60% weight; logical coherence dimension, based on the logical relationship graph of the reference answer (such as "reason-result" "comparison-analysis"), through dependency syntax analysis to detect whether the logical connection of the student's text meets the hierarchical requirements, such as the "total-part-total" structure of the argumentative essay, deduct the corresponding weight; language standardization dimension, according to the language standard of the reference answer (such as the use of "zhi" in classical Chinese, correct form of English tenses), through grammar check, word accuracy analysis to mark language errors.

[0092] For the diversity of subjective questions, the engine sets up a semantic fault-tolerant mechanism, if the student's answer uses different expressions but is semantically equivalent (such as "people's livelihood problem" and "people's livelihood security"), through synonym library and semantic vector cosine similarity calculation to confirm the effective score points; if it exceeds the core semantic range of the reference answer but the expression is reasonable, combined with the teaching standards of the subject to add 1-5% innovation points. After scoring, the engine generates a structured report including semantic deviation details, marking the differences with the reference answer (such as "missing the second core point of the reference answer" "the logical hierarchy does not meet the requirements of the reference answer" "the "yan" word in classical Chinese is interpreted differently from the reference answer standard"), providing accurate semantic basis for subsequent error analysis, anchoring the scoring standard through the reference answer, and compatible with the diversified expression of liberal arts subjective questions, achieving the balance between standardization and flexibility.

[0093] S5. According to the review results of each subject, the error causes are analyzed respectively. In specific implementation, first, the liberal arts NLP deep analysis engine is classified and collected according to the subject by the data aggregation interface, and the reports generated by the science symbol calculation engine and the image feature recognition engine are classified and collected according to the subject, forming a structured data set including topic ID, subject label, score details, error label (such as "formula application error", "semantic understanding deviation", "geometric angle label error") and original answer data. For science subjects, based on the preset error type system (covering three dimensions of concept class, calculation class and operation class), the step errors (such as "misuse of Newton's second law formula" and "mistake in formula of quadratic equation root") marked by the symbol calculation engine and the vector verification anomaly (such as "line segment length measurement deviation exceeds 5%" and "force direction angle label error") of the image feature recognition engine are cross-compared, and are associated to the science error knowledge base (storing typical error patterns such as "trigonometric function formula confusion" and "improper addition of geometric auxiliary line") through rule matching, and the error root cause (such as "knowledge point memory not firm", "insufficient spatial imagination ability" and "careless calculation") is located. For multi-dimensional errors (such as formula error leading to calculation result deviation, and then affecting image vector matching) in the same question, the cause and effect relationship is analyzed through error association algorithm to avoid isolated judgment.

[0094] For liberal arts subjects, focus on the semantic deviation points (such as "argument omission", "emotion analysis error" and "error in use of virtual words in classical Chinese") extracted by the NLP engine, and combine the text logical relationship graph and the liberal arts knowledge base (including error types such as "historical event timeline confusion", "logical break of argumentative essay" and "deviation in understanding of ancient poetry images"), and perform hierarchical classification: objective question errors preferentially match knowledge point errors (such as "historical event occurrence time error"), and subjective question errors are decomposed from three dimensions of semantic coverage (whether core points are omitted), logical coherence (whether the connection between arguments is reasonable), and language standardization (whether grammar and words are accurate), to identify error types (such as "incomplete reading comprehension", "disordered argument structure" and "error in interpretation of classical Chinese characters and words"). For special writing styles such as classical Chinese and ancient poetry, additional field-specific rule base (such as "error in judgment of word class use" and "deviation in identification of rhetorical devices") is used to refine error description.

[0095] In the analysis process, the student's historical answer data is synchronously associated, and repeated errors (such as "English tense use error" and "incomplete factorization") are identified through longitudinal comparison and marked as weak links that need to be emphasized. Finally, a structured analysis report including error position (such as "step 2" and "second paragraph, third sentence"), error type (such as "concept error" and "logical error"), associated knowledge points (such as "mechanical energy conservation law" and "historical significance of the Xinhai Revolution") and error level (key error / secondary error) is generated for each error question, providing accurate attribution basis for subsequent subject-specific error question set generation.

[0096] According to the error cause analysis result, the corresponding subject error question set is generated, in the specific implementation, the error question set generation process is started based on the error cause analysis report of the subject, first, the error question core information is extracted from the structured data set, including the question original image segment (the question region cropped by PANet), the text content recognized by CRNN, the error type label (such as 'concept confusion', 'calculation error','semantic deviation'), the associated knowledge points and the error level. For the science error question set, it is classified and collected according to subjects such as mathematics, physics and chemistry, each error question includes question image, student original answer scan, error step label (formula misuse or calculation error position marked by symbolic calculation engine), vector relationship verification exception explanation (such as angle / length deviation detected by image feature recognition engine), and automatically supplemented correct problem solving step disassembly (including key formula prompt, unit conversion specification) and similar question type expansion link. For geometric error questions, standardized geometric figure sketches are embedded, correct auxiliary line addition positions and angle / length parameters are labeled, and the figure deviation in student answers is compared.

[0097] For the liberal arts error question set, it is classified and arranged according to subjects such as Chinese, English and history, each error question includes question text (such as reading comprehension original text segment, classical Chinese original text), student answer text, semantic deviation points marked by NLP engine (such as'missing argument''misuse of emotional words'), and supplemented knowledge point analysis (such as classical Chinese virtual word interpretation, historical event timeline analysis, English grammar rule explanation). For subjective questions, the logic framework diagram (such as argumentative thesis argument progression relationship) and semantic optimization suggestions (such as 'this place can supplement the melancholy connotation of the setting sun image to strengthen emotional expression') of the standard answer are added.

[0098] During the generation process, the system automatically adds personalized labels to the error question set, marks high-frequency error knowledge points (such as 'quadratic function maximum value solving' 'ancient poetry uses scenery to express emotion') based on historical error analysis, and sorts them according to error level (key errors are placed at the top). Format adaptation to meet scene requirements, using lightweight layout, supporting A4 and small size paper printing of error question printer, clear division of questions and analysis, error points marked with color (such as red box error steps, yellow highlighted semantic deviation words), and edge reserved for note space. At the same time, a unique two-dimensional code is generated for each error question, which is associated with the complete analysis video of the original question and the same type of exercise question bank, and students can trace the detailed explanation by scanning the code with their mobile phones. After generation is completed, the error question set data is synchronized to local storage and cloud backup, supporting students to view electronic version through mobile terminal or one-key printing through school deployed error question printer, realizing the closed-loop optimization of error question collection, analysis and review.

[0099] The application also provides an intelligent homework paper marking and error question induction system, which comprises Figure 2As shown, the system is based on a basic support platform, driven by an intelligent marking engine core process, through discipline routing distribution to realize the processing of subject and discipline, and finally complete the error cause analysis and error question generation, covering the whole process of closed loop from homework collection to error question output.

[0100] The basic support platform provides image preprocessing, data storage, computing power acceleration and security capability, including the following functional modules:

[0101] Multi-source acquisition unit: used for collecting homework / paper images through mobile phone camera and error question printer camera, real-time checking image clarity, integrity and lighting conditions, guiding users to adjust the shooting angle and light compensation to ensure that the image meets the subsequent processing standards;

[0102] Preprocessing center: used for distortion correction and image enhancement, wherein the distortion correction is to call the device parameter library to correct the radial and tangential distortion of the mobile phone image; through paper edge detection, the perspective distortion of the error question printer image is corrected, and the geometrically regular image is output; image enhancement is to decompose the light and reflection components, balance the light and dark contrast; adaptive binary is used to highlight the text, combined with Gaussian filter and median filter to remove noise, sharpen the text edge, and improve the OCR recognition accuracy;

[0103] Security encryption module: encrypts homework images and marking data to ensure data transmission and storage security.

[0104] Distributed storage module: stores homework images, text recognition results, marking data and error question sets, supports efficient reading and writing and persistence of large-scale data.

[0105] GPU acceleration cluster: provides computing power support for image detection, text recognition, engine inference and other computationally intensive tasks, and improves processing efficiency.

[0106] It also includes an intelligent marking engine, which is the core processing center of the system, receives image data processed by the data preprocessing center, coordinates the storage and computing power resources of the basic support platform, and drives the text processing, discipline classification and discipline marking process. The intelligent marking engine is used for text detection and character recognition. The text detection is based on PANet to extract the bounding polygon coordinates of the text in the image and crop the text line area; the character recognition is to convert the text line image into a text string through CRNN, filter invalid characters and check semantic relevance.

[0107] Discipline routing distribution module: used to extract the subject identification and question type characteristics in the text, match the keywords and compare the characteristics library, divide the text into science and arts, and distribute it to the corresponding marking module;

[0108] The sub-discipline reviewing module includes a science reviewing sub-module and an art reviewing sub-module, wherein the science reviewing sub-module includes a symbol calculation engine and an image feature recognition engine, the symbol calculation engine is used for analyzing reference answer step logic, comparing student answers of formula application, calculation process and results, and outputting step scores and difference annotations; the image feature recognition engine is used for extracting graphic vector features and verifying standard matching with reference answers; the art reviewing sub-module includes an NLP deep analysis engine, and the NLP deep analysis engine scores texts from the dimensions of semantic coverage, logical coherence and language standardization.

[0109] The error cause analysis module collects science and art reviewing results, associates error types (such as formula misuse and semantic omission) and subject knowledge points (such as mechanical energy conservation and essay structure), locates error sources (such as concept confusion and careless calculation), and generates a structured report including error positions, types and associated knowledge points.

[0110] The error question generation module integrates question images, student answer records and error cause analysis results, and generates error question sets according to subjects.

[0111] Through hierarchical architecture design, the system realizes the full-process automation of "multi-source collection, basic preprocessing, intelligent reviewing, sub-discipline analysis, error cause analysis and error question generation", reduces hardware resource consumption while ensuring reviewing accuracy, and adapts to large-scale deployment in campus scenarios.

[0112] The application also provides an electronic device embodiment. The electronic device is in the form of a general-purpose computing device. The components of the electronic device can include but are not limited to one or more processors or processing units, memory, and a bus connecting different components including memory and processing units.

[0113] The bus represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any of a variety of bus structures. For example, these architectures include but are not limited to Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MCA) bus, Enhanced ISA bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnect (PCI) bus.

[0114] Electronic devices typically include a variety of computer-readable media that store instructions that are executed by a processor or computing component. Computer-readable media can include, at least, two types of media, namely computer storage media and communications media. Computer storage media, which is tangible, includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. The system memory 130, the removable storage device 139 and the non-removable storage device 140 are all computer storage media examples (i.e., memory storage.) Communications media typically embodies computer-readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism and includes any information delivery media. Although communications media is not a computer storage media, it can carry computer readable instructions for a program module and can be a source of computer readable instructions for the device 100. For example, instructions for a program module can be received from a network.

[0115] The memory can include computer-readable media in the form of volatile memory, such as random-access memory (RAM) and / or cache memory. The electronic device can further include other removable / non-removable, volatile / non-volatile computer storage media. By way of example only, storage system 130 can be provided for reading from and writing to a non-removable, non-volatile magnetic media (e.g., a hard disk drive).

[0116] The electronic device can also communicate with one or more external devices such as a keyboard or pointing device, a camera, etc. via an I / O interface. The electronic device can also include a display for displaying graphics and images. The electronic device can also be in communication with one or more devices that enable a user to interact with the electronic device. Further, the electronic device can communicate with one or more networks (such as a Local Area Network (LAN), a Wide Area Network (WAN), and / or the Internet) via a network adapter. The network adapter can be any of a variety of modems, including cable modems, telephone modem, and broadband over power line modems. The network adapter interfaces with the electronic device's bus via the input / output (I / O) interface. The processor runs a variety of applications and stores data that are used in connection with the applications and the processes of the present application, e.g., at least one step of the intelligent homework paper grading and mistake summary method provided by the embodiments of the present application, which includes: collecting image data of homework or test paper, preprocessing the collected image data, detecting the outer polygon coordinates of the text in the image through PANet after preprocessing the image, recognizing the text line picture through CRNN, and outputting the text string; for science category, decomposing and scoring the problem solving steps through a symbol calculation engine, verifying the vector relationship through an image feature recognition engine; for arts category, performing semantic scoring through an NLP deep analysis engine; and generating a mistake set corresponding to each subject according to the grading results of each subject.

[0117] The present application also provides a computer-readable storage medium having stored thereon a computer program, which, when executed by a processor, implements at least one step of the intelligent homework paper grading and mistake summary method provided by the embodiments of the present application.

[0118] The computer storage medium of the embodiments of the present application can adopt any combination of one or more computer readable media. The computer readable medium can be a computer readable signal medium or a computer readable storage medium. The computer readable storage medium may, for example, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, system or device, or any combination thereof. More specific examples (non-exhaustive list) of the computer readable storage medium include: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In this document, the computer readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system or device.

[0119] The computer readable signal medium can include a data signal propagated in a baseband or as a part of a carrier wave, in which a computer readable program code is borne. Such propagated data signal can take various forms, including but not limited to an electromagnetic signal, an optical signal or any suitable combination thereof. The computer readable signal medium can also be any computer readable medium that is not a computer readable storage medium and that can transmit, propagate or transport a program for use by or in connection with an instruction execution system, system or device.

[0120] The program code contained on the computer readable medium can be transmitted by any suitable medium, including but not limited to wireless, wire, cable, optical fiber, RF, etc. or any suitable combination thereof.

[0121] Finally, it should be noted that: the above only describes the preferred embodiments of the present application, and is not used to limit the present application, although the present application has been described in detail with reference to the foregoing embodiments, for those skilled in the art, the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced by equivalent, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application, should be included in the protection scope of the present application.

[0122] The above are the preferred embodiments of the present application, not to limit the protection scope of the present application, therefore: any equivalent change, equivalent replacement, improvement, etc. made according to the structure, shape, principle of the present application, should be included in the protection scope of the present application.

Claims

1. A method for intelligent homework paper marking and mistake summary, characterized in that, The method comprises the steps of: collecting image data of the assignment or test paper, and preprocessing the collected image data, the preprocessing including distortion correction and image enhancement; transferring the preprocessed image data to a PANet, detecting the bounding polygon coordinates of the text in the image through the PANet, and then transferring the bounding polygon coordinates to a CRNN, recognizing the text line picture through the CRNN, and outputting a text string; dividing the recognized text into arts and science categories, distributing the text to corresponding processing modules through a discipline routing, decomposing and scoring the problem solving steps through a symbolic calculation engine for the science category, verifying the vector relationship through an image feature recognition engine, and performing semantic scoring through an NLP deep analysis engine for the arts category; analyzing the error causes according to the marking results of each subject, and generating a mistake set corresponding to each subject according to the error cause analysis results; decomposing and scoring the problem solving steps through a symbolic calculation engine for the science category, specifically including: when the symbolic calculation engine receives the science text data distributed through the discipline routing, the structured reference answer data package associated with the question is called synchronously, the symbolic calculation engine performs structured analysis on the reference answer, extracts a benchmark checking framework, matches the student's answer unit with the core step framework of the reference answer, and directly checks the reference answer point by point if the student adopts the same solution path as the reference answer; if the student adopts other solution, the "logical integrity" core standard extracted based on the reference answer is checked; the scoring node weight preset in the reference answer is embedded into a dynamic weight mechanism to score the student's solution steps; the student's calculation result is compared with the standard result of the reference answer for equivalence, and if the deviation exceeds the fault tolerance threshold preset in the reference answer, the deviation source is located in combination with the intermediate step label in the reference answer; a structured report including the difference between the student's steps and the reference answer is generated; verifying the vector relationship through an image feature recognition engine, specifically including: when the image feature recognition engine receives the science graphic data distributed through the discipline routing, the reference answer vector feature data package associated with the question is called synchronously, the image feature recognition engine first normalizes and extracts features of the graphic area image of the student's answer, and constructs a structured vector data set; then, according to the vector specification library of the reference answer, multidimensional comparison is carried out, and the value label in the figure is cross-verified with the step data output by the symbolic calculation engine, and after the verification is completed, a vector verification report including the deviation type, deviation position and reference answer comparison explanation is generated.

2. The intelligent homework test paper grading and error summary method according to claim 1, characterized in that, The PANet detects the bounding polygon coordinates of the text in the image, specifically including: The PANet receives the normalized image after distortion correction and image enhancement, first extracts features from the image to generate feature maps of different levels from low to high, then fuses multi-scale features using a feature pyramid structure to form a fused feature map with details and semantics. In the detection process, PANet classifies the fused feature map at the pixel level to distinguish between text and background areas, and simultaneously regresses the bounding box parameters of the text area. For printed titles, the output is a quadrilateral bounding coordinate, and for handwritten answers, the output is a polygonal coordinate with more than 6 vertices. After detection, overlapping and redundant detection boxes are removed, and areas with small areas or abnormal aspect ratios are filtered out. Finally, the output text bounding polygon coordinates are used to accurately crop the text line area from the image.

3. The intelligent homework test paper grading and error summary method according to claim 1, characterized in that, The text line picture is recognized by the CRNN, specifically including: The CRNN first normalizes the text line picture. The convolutional layer extracts local features of the text line by multiple convolution and pooling operations to generate a feature sequence. The recurrent layer models the feature sequence output by the convolutional layer in time series and uses a bidirectional propagation mechanism to capture the dependency between the previous and subsequent strokes, adapting to common student handwriting forms such as connected writing, tilting, and irregular writing. The transcription layer converts the sequence features output by the recurrent layer into a text string.

4. The intelligent homework test paper grading and error summary method according to claim 1, characterized in that, The semantic score is obtained by the NLP deep analysis engine, specifically including: When the NLP deep analysis engine receives the liberal arts text data distributed by the subject routing, it synchronously calls the liberal arts reference answer data package associated with the question, structurally analyzes the reference answer, and extracts a semantic scoring benchmark framework; For student answer text, first convert the text to a computable semantic feature sequence through vector conversion, then generate a context semantic vector, and compare it with the semantic vector of the reference answer; score the student answer text, The scoring dimensions include semantic coverage, logical coherence, and language standardization. After scoring, a structured report including semantic deviation details is generated, marking the differences with the reference answer.

5. The intelligent homework test paper grading and error summary method according to claim 1, characterized in that, According to the results of each subject review, the error causes are analyzed, specifically including: Through the data aggregation interface, the reports generated by the liberal arts NLP deep analysis engine, the science and engineering symbol calculation engine, and the image feature recognition engine are classified by subject to form a structured data set including question ID, subject label, score details, error annotations, and original answer data; For science and engineering subjects, based on the preset error type system, cross-compare the step errors marked by the symbol calculation engine and the vector verification abnormalities of the image feature recognition engine, associate them with the science and engineering error knowledge base, locate the error sources, and analyze the causal relationships for multi-dimensional errors in the same question; For liberal arts subjects, based on the semantic deviation points extracted by the NLP deep analysis engine, combined with the text logical relationship graph and the liberal arts knowledge base, the errors are classified hierarchically. For objective question errors, match the knowledge point errors, and for subjective question errors, analyze from the three dimensions of semantic coverage, logical coherence, and language standardization. During the error analysis process of each subject, the student's historical answer data is synchronized and associated to identify repetitive errors through longitudinal comparison, mark the weak links that need to be focused on, generate a structured analysis report for each error question including error location, error type, associated knowledge points, and error level, and generate a corresponding error question set based on the error cause analysis results.

6. The intelligent test paper marking and error summary system using the method according to any one of claims 1 to 5, characterized in that, Comprise: Multi-source acquisition unit: used to acquire homework / paper images through mobile phone camera and error question printer camera; Pre-processing center: used for distortion correction and image enhancement; Intelligent grading engine: used for text detection and character recognition, where text detection is based on PANet to extract the bounding polygon coordinates of the text in the image and crop the text line area; Character recognition is through CRNN to convert text line images into text strings, filter invalid characters and check semantic relevance; Subject routing distribution module: used to divide the text into science and arts, and distribute it to the corresponding grading module; Subject grading module: including science grading sub-module and arts grading sub-module, where the science grading sub-module includes symbol calculation engine and image feature recognition engine, the symbol calculation engine is used to analyze the reference answer step logic, compare the formula application, calculation process and result of the student's answer, and output the step score and difference mark; The image feature recognition engine is used to extract the graph vector feature and verify the standard matching with the reference answer; the arts grading sub-module includes NLP deep analysis engine, which scores the text from the dimensions of semantic coverage, logical coherence and language standardization; Error cause analysis module: used to collect the results of science and arts grading, associate error types with subject knowledge points, locate error sources, and generate a structured error cause report; Error question generation module: used to integrate question images, student answers, error cause analysis results, and generate error question sets by subject classification.

7. An electronic device, comprising one or more processors, wherein the one or more processors are configured to: The processor executes the program stored in the memory, thereby executing the intelligent homework and paper grading and error question induction method of any one of claims 1-5.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the intelligent homework and paper grading and error question induction method of any one of claims 1-5.

Images