Work analysis and error diagnosis method, system, computer device and medium

Abstract

The present application relates to a kind of job analysis and error diagnosis method, system, computer equipment and medium, it is related to teaching auxiliary management technical field, including the following steps, original job image is scanned to real object job paper, standard job image is generated after edge enhancement and distortion correction;The image is blocked and positioned to get the answer area map, and then content separation gets the answer element set;Element set feature extraction gets element feature group, and structure analysis gets solution logic chain;If potential doubt is detected, then based on potential doubt map, error association analysis is carried out on standard job image, error association is obtained, and its propagation tracking is carried out, and finally implicit error report is generated. It solves the technical problems that most systems are still limited to answer matching or simple rule judgment, lack of modeling ability for problem solving logic chain, and cannot effectively identify implicit errors of students in reasoning process.

Description

Technical Field

[0001] This invention relates to the field of teaching support management technology, and in particular to methods, systems, computer equipment and media for homework analysis and error diagnosis. Background Technology

[0002] With the deep integration of educational informatization and artificial intelligence technologies, intelligent homework grading and learning diagnosis systems are gradually becoming important tools for improving teaching efficiency and personalized tutoring. Traditional homework analysis mainly relies on teachers' manual grading, which is not only time-consuming and labor-intensive, but also makes it difficult to achieve a structured understanding of students' problem-solving processes and in-depth error attribution. In recent years, although automatic grading technology based on image processing and natural language processing has made some progress, most systems are still limited to answer matching or simple rule judgment, lacking the ability to model the logical chain of problem-solving and failing to effectively identify implicit errors in students' reasoning process. Summary of the Invention

[0003] The purpose of this invention is to at least partially solve one of the technical problems existing in the prior art.

[0004] To achieve the above objectives, the present invention provides a job analysis and error diagnosis method, comprising the following steps: The physical work test paper is scanned and collected to obtain the original work image, and the edge enhancement and distortion correction of the original work image are performed to obtain the standard work image; The standard operation image is segmented and located to obtain the answer area map, and the answer area map is separated to obtain the answer element set; Feature extraction is performed on the set of answer elements to obtain feature groups, and structural analysis is performed on the feature groups to obtain the solution logic chain; If there are potential issues in the solution logic chain, then an error correlation analysis is performed on the standard operation image based on the potential issue graph to obtain an error correlation network. The error correlation network is then propagated and tracked to obtain a hidden error report.

[0005] Furthermore, edge enhancement and distortion correction are performed on the original work image to obtain a standard work image, including: The original work image is subjected to grayscale noise reduction processing to obtain a noise-reduced work image; Edge features are extracted from the noise-reduced image to obtain an edge-enhanced image, and local sharpening is performed on the edge-enhanced image to obtain a sharpened image. The sharpened image is subjected to distortion region detection to obtain distortion region markers, and geometric correction processing is performed based on the distortion region markers to obtain a standard image.

[0006] Furthermore, the standard work image is segmented and located to obtain an answer area map, including: The standard work image is subjected to line detection by Hough transform to obtain a reference line set, and the reference line set is adaptively projected to obtain a projection density map. Density clustering is performed on the projected density map using maximum suppression to obtain a clustered region map, and the clustered region map is further subdivided into equal-interval subdivisions to obtain a raster matrix map; Based on the raster matrix diagram, the clustering region diagram is subjected to morphological processing to obtain a region segmentation diagram, and the boundary of the region segmentation diagram is traced to obtain a contour chain code diagram. The contour chain code map is fused into regions to obtain a merged region map, and the merged region map is constrained and optimized to obtain the answer region map.

[0007] Furthermore, the answer area map is subjected to content separation to obtain a set of answer elements, including: Texture features are extracted from the answer area map to obtain a texture feature map, and the printed and handwritten areas are segmented based on the texture feature map to obtain a printed area map and a handwritten area map; The printed area map is locally contrast-enhanced to obtain an enhanced printed area map, and the enhanced printed area map is then subjected to connected component segmentation and semantic clustering to obtain a title element matrix. Handwriting feature analysis is performed on the handwritten area map to obtain a handwriting feature vector set, and text and formula area classification is performed based on the handwriting feature vector set to obtain a text area map and a formula area map. The text region map and formula region map are subjected to structured analysis to obtain the analysis element group, and the analysis element group is semantically associated and integrated with the question element matrix to obtain the answer element set.

[0008] Furthermore, feature extraction is performed on the set of answer elements to obtain element feature groups, and structural parsing is performed on the element feature groups to obtain the solution logic chain, including: Semantic hierarchical analysis is performed on the set of answer elements to obtain a hierarchical feature map, and deep semantic parsing is performed on the hierarchical feature map to obtain a semantic element matrix; Based on the semantic element matrix, knowledge points are mapped onto the answer element set to obtain a knowledge point association graph, and the knowledge point association graph is recursively expanded to obtain element feature groups. Temporal correlation analysis is performed on the feature group of the elements to obtain a reasoning time sequence diagram, and the reasoning time sequence diagram is divided into logical units to obtain a sequence of logical steps; Based on the logical step sequence, dependency relationships are constructed for the feature group to obtain a derivation dependency tree, and the path of the derivation dependency tree is restored to obtain the solution logic chain.

[0009] Furthermore, based on the logical step sequence, dependency relationships are constructed on the feature groups to obtain a derivation dependency tree, including: Semantic role labeling is performed on each logical step in the logical step sequence to obtain a step predicate-argument structure set. Cross-modal semantic alignment is then performed between the step predicate-argument structure set and the feature group to obtain a step-feature binding mapping table. A directed acyclic graph is constructed on the logical step sequence based on the step-feature binding mapping table to obtain an initial derivation graph. Dominant node identification and subtree merging operations are then performed on the initial derivation graph to obtain the derivation dependency tree.

[0010] Furthermore, based on the potential problem map, error correlation analysis is performed on the standard operation image to obtain an error correlation network, including: Spatial coordinate mapping is performed on the potential suspicious point image to obtain a suspicious point pixel coordinate set, and local content backtracking is performed on the standard operation image based on the suspicious point pixel coordinate set to obtain a suspicious point context image block. The context image blocks of the doubts are semantically aligned to obtain an aligned element sequence. Based on the aligned element sequence, bidirectional index matching is performed with the answer element set to obtain a doubt-element mapping table. Based on the aforementioned question-element mapping table, the solution logic chain is expanded in reverse to obtain the logic breakpoint trajectory, and the logic breakpoint trajectory is reconstructed by causal dependency to obtain the error propagation subgraph. The error propagation subgraph is expanded by multiple order neighborhoods to obtain an expanded error topology. The intersection constraint between the expanded error topology and the knowledge point association graph is then applied to obtain the error association network.

[0011] Furthermore, the propagation of the aforementioned error network is traced to obtain latent error reports, including: The error association network is analyzed for topology and the importance of nodes is calculated to obtain a set of key error nodes. Based on the set of key error nodes, the reverse path is traced to obtain the main path of error propagation. The error propagation main path is expanded by multiple levels of neighborhood and the influence weight is accumulated to obtain an expanded influence node set. The expanded influence node set is then mapped and aligned with the knowledge point association graph to obtain the knowledge defect association area. The knowledge defect association area is classified and labeled according to the cognitive dimension to obtain a defect type matrix. Based on the defect type matrix and the main error transmission path, a temporal correlation integration is performed to obtain a hidden error report.

[0012] The present invention also provides a job analysis and error diagnosis system, comprising: The acquisition module is used to scan and acquire physical homework test papers to obtain original homework images, and to perform edge enhancement and distortion correction on the original homework images to obtain standard homework images; The positioning module is used to segment and locate the standard operation image to obtain the answer area map, and to separate the content of the answer area map to obtain the answer element set; The extraction module is used to extract features from the set of answer elements to obtain feature groups, and to perform structural analysis on the feature groups to obtain the solution logic chain. The detection module is used to detect whether there are potential doubts in the solution logic chain. If so, it performs error correlation analysis on the standard operation image based on the potential doubt map to obtain an error correlation network, and tracks the propagation of the error correlation network to obtain a hidden error report.

[0013] This invention provides a method for homework analysis and error diagnosis, comprising the following steps: scanning and acquiring physical homework test papers to obtain original homework images, and performing edge enhancement and distortion correction on the original homework images to obtain standard homework images; segmenting and locating the standard homework images to obtain answer area maps, and separating the answer area maps into content to obtain answer element sets; extracting features from the answer element sets to obtain element feature groups, and performing structural analysis on the element feature groups to obtain solution logic chains; detecting whether there are potential doubts in the solution logic chains, and if so, performing error correlation analysis on the standard homework images based on the potential doubt maps to obtain error correlation networks, and tracking the propagation of the error correlation networks to obtain implicit error reports. This method solves the technical problem that most systems are still limited to answer matching or simple rule judgment, lack the ability to model the problem-solving logic chains, and cannot effectively identify implicit errors in students' reasoning process. It achieves the technical effect of constructing error correlation networks and tracking the propagation path of errors in the logic chains, distinguishing between surface errors and root causes, revealing the causal relationship between multiple errors, and generating implicit error reports with teaching guidance value. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1This is a schematic diagram of the steps of a job analysis and error diagnosis method in one embodiment of the present invention; Figure 2 This is a schematic diagram of the steps of a job analysis and error diagnosis system in one embodiment of the present invention; Figure 3 This is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0016] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0017] The embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. The step numbers in the following embodiments are set only for ease of explanation, and there is no limitation on the order between the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0018] The following describes in detail, with reference to the accompanying drawings, a job analysis and error diagnosis method according to an embodiment of the present invention. First, the job analysis and error diagnosis method according to an embodiment of the present invention will be described with reference to the accompanying drawings.

[0019] Figure 1 This invention provides a method for job analysis and error diagnosis, comprising the following steps: Step S1: Scan and collect the physical work test paper to obtain the original work image, and perform edge enhancement and distortion correction on the original work image to obtain the standard work image.

[0020] Specifically, the first step in the process is to scan and collect physical worksheets to obtain original work images. This means converting the paper work into digital image files using scanning equipment, such as common flatbed scanners or document cameras. After obtaining the original work images, edge enhancement and distortion correction are performed to obtain standard work images. To achieve this, the edge features of the original work images must first be analyzed, and image processing algorithms are used to identify the positions of the four edges. Here, the Canny edge detection algorithm can be used to determine the edge contours. Once the edges are accurately identified, perspective transformation technology is applied to adjust the image distortion caused by incorrect scanning angles, restoring the originally tilted or distorted image to a normal perspective and proportion, ensuring that the text and graphic information are as close to the original as possible. For example, when processing a worksheet containing math problems, if the scan result shows trapezoidal distortion due to the paper being placed at an angle, the above steps can convert the distorted parts into rectangular areas, ensuring that each question and its answer are in the correct relative position. Then, the edge enhancement process is further refined, that is, the image edges are sharpened to emphasize boundary information, making subsequent block positioning more accurate. Specifically, based on the identified edges, methods such as the Sobel operator can be used to enhance these edges, increasing the contrast between the answer area and the background, thus making it easier to distinguish different content sections. For example, when processing a physics experiment report containing charts and extensive text descriptions, the enhanced edges not only help clearly define the scope of each chart but also help accurately separate the text paragraphs corresponding to the questions, laying a solid foundation for subsequent content separation, feature extraction, and structural analysis. In this way, through a series of meticulous image processing operations, high-quality standard assignment images are ultimately obtained, providing reliable input data for subsequent steps.

[0021] Step S2: The standard operation image is segmented and located to obtain the answer area map, and the answer area map is separated to obtain the answer element set.

[0022] Specifically, the standard assignment image is segmented and localized to obtain the answer area map. This process first uses template matching or connected component analysis to identify the preset positions or answer boundaries of each question on the test paper based on the corrected layout structure of the standard assignment image, thereby dividing the entire image into several independent blocks, each corresponding to a specific answer position. On this basis, through coordinate mapping and region cropping operations, each local image containing the student's answer is extracted and combined to form a complete answer area map. Next, the answer area map is separated to obtain the answer element set. This step requires fine-grained analysis of the mixed content (such as handwritten text, mathematical formulas, geometric figures, arrow labels, etc.) in each answer area. For example, a deep learning-based instance segmentation model (such as Mask R-CNN) is used to perform pixel-level classification and mask generation for different types of visual elements. Then, an OCR engine is used to identify text content and symbol content respectively, and the outline and relative position information of non-text graphic elements are preserved. Finally, all identifiable answer components are organized in a structured form into the answer element set. For example, when processing a standard image of a junior high school math homework assignment, if the answer area of ​​a geometry proof question simultaneously contains handwritten reasoning statements, drawn auxiliary lines, and angle markers, the answer area image corresponding to the question can be accurately extracted through the above-mentioned block positioning. After content separation, the handwritten statements are converted into text strings, the auxiliary lines are converted into vector line segment data, and the angle markers are converted into specific identifiers, which together constitute the answer element set of the question, providing atomic-level input data for subsequent feature extraction and logical chain construction.

[0023] Step S3: Extract features from the set of answer elements to obtain feature groups, and perform structural analysis on the feature groups to obtain the solution logic chain.

[0024] Specifically, feature extraction is performed on the answer element set to obtain element feature groups. This process first requires analyzing various types of data contained in the answer element set, such as text strings, vector line segment data, and specific identifiers. Natural language processing techniques are used to parse the keywords and semantic information in the text. For example, handwritten reasoning statements are converted into high-dimensional vector representations using word embedding models, and geometric attribute parameters, such as the length and angle of auxiliary lines, are calculated using image recognition algorithms. The specific type of angle symbol is determined using pattern recognition technology, thereby assigning a unique feature description to each element and forming an element feature group. Based on this, structural analysis is performed on the element feature group to obtain the solution logic chain. This further explores the intrinsic connections between these features. By constructing graph structures or sequence models, the logical relationships between the elements in the solution process are simulated. For example, in geometric proof problems, the existence and significance of auxiliary lines and their relationship with other elements are deduced based on the text description. The accuracy of the reasoning steps is verified by combining vector data, and the angle symbol is considered as a condition constraint to improve the entire proof chain, ultimately forming a coherent and logically rigorous solution logic chain. For example, when processing the standard homework image of the aforementioned junior high school math homework, for a question involving proof of triangle congruence, feature extraction can be performed on the handwritten reasoning statements in the answer element set to identify the vector expressions corresponding to key terms such as "side-angle-side" and "angle-side-angle". For auxiliary lines and angle symbols, their lengths, angle values, and identifying properties are extracted respectively. Then, in the structural analysis stage, these feature elements are linked together in a logical order. For example, first confirm the congruence conditions indicated by the handwritten reasoning statements, then verify whether the auxiliary lines meet these conditions based on the vector data, and finally use the angle symbols as supplementary proof conditions. In this way, a complete logical chain of the solution can be constructed, ensuring that each step of reasoning from the question to the answer has sufficient evidence to support it.

[0025] Step S4: Detect whether there are potential doubts in the solution logic chain. If so, perform error correlation analysis on the standard operation image based on the potential doubt map to obtain an error correlation network, and track the propagation of the error correlation network to obtain a hidden error report.

[0026] Specifically, to detect potential doubts in the solution logic chain, the first step is to compare each node and its connection relationship in the solution logic chain item by item according to the preset subject rule base and logical consistency constraints. For example, in a middle school math geometry proof problem, if a step in the solution logic chain references the "side-angle-side" congruence criterion, but the lengths of the two sides or the included angle on which it depends have not been effectively proven or marked in the previous steps, then that node is marked as a potential doubt. When one or more such inconsistent, missing premises, or axiom-violation nodes are identified, a corresponding potential doubt graph is generated. This graph records the location, type, and related logic chain fragments of the doubt in a structured manner. Next, based on the potential doubt graph, error association analysis is performed on the standard operation image to obtain an error association network. This process involves tracing back the original answer area related to the doubt in the standard operation image, extracting its corresponding answer elements (such as handwritten statements, graphic symbols, etc.), and combining the spatial arrangement and semantic dependencies between contextual elements to construct a multi-dimensional association graph, i.e., the error association network, which includes error nodes, missing supporting evidence points, and potential sources of misleading information. Subsequently, the system propagates and traces the error network to obtain latent error reports. Specifically, it simulates the diffusion path of the error's impact along the edges of the network. For example, an incorrectly labeled angle might lead to an incorrect congruence determination, resulting in an incorrect conclusion. The system traces the complete propagation chain from the initial drawing error to the final inaccurate conclusion, integrating all affected logical nodes and their root causes into a structured latent error report. For instance, in the aforementioned middle school math homework, if a student mistakenly uses a non-angled angle as an included angle when proving the congruence of two triangles, the system first detects this violation of the "side-angle-side" definition in the solution's logical chain and marks it as a potential problem. Then, it checks the corresponding angle labeling area in the standard homework image, finding that the actual labeled angle is not the included angle between the two known sides. This establishes a connection between the graphic error and the logical error. Through propagation tracing, it reveals how this misunderstanding of the graphic leads to the failure of all subsequent derivations, ultimately outputting a latent error report indicating that "the congruence determination logic was broken due to an incorrect angle labeling."

[0027] In a specific embodiment, edge enhancement and distortion correction are performed on the original job image to obtain a standard job image, including: The original work image is subjected to grayscale noise reduction processing to obtain a noise-reduced work image; Edge features are extracted from the noise-reduced image to obtain an edge-enhanced image, and local sharpening is performed on the edge-enhanced image to obtain a sharpened image. The sharpened image is subjected to distortion region detection to obtain distortion region markers, and geometric correction processing is performed based on the distortion region markers to obtain a standard image.

[0028] Specifically, edge enhancement and distortion correction are performed on the original work image to obtain a standard work image. This process includes several sequentially executed operational steps. First, grayscale noise reduction processing is performed on the original work image to obtain a denoised work image. This step involves converting the original work image in color or RGB format to a grayscale image and applying a non-local mean denoising algorithm or a Gaussian filter to suppress random noise, scan spots, and paper texture interference in the image. This reduces pixel-level disturbances while preserving the main structural information, thereby generating a visually cleaner and more stable denoised work image. Next, edge features are extracted from the denoised image to obtain an edge-enhanced image. This step uses the Canny operator or the Sobel gradient operator to calculate the gradient magnitude and direction of each pixel in the image, identify significant boundaries such as character strokes, graphic outlines, and question separators, and enhance these boundaries as highlighted pixels to form an edge-enhanced image that highlights key structures. Subsequently, the edge-enhanced image is locally sharpened to obtain a sharpened image. Specifically, unsharp masking is used to construct a blurred copy around the edge region and then subtract it from the original image to enhance high-frequency details, making the originally blurred character edges or fine geometric lines clearer and sharper, thereby improving the accuracy of subsequent positioning and recognition. After obtaining the sharpened image, the process continues by detecting distortion regions, resulting in distortion region markers. This step analyzes whether the geometry of known structures such as rectangular answer boxes, exam paper borders, or question numbers in the image is distorted. For example, Hough transform is used to detect the parallelism and intersection angles of line segments. If quadrilateral vertices deviate from the ideal rectangular layout, it is determined that the region has perspective distortion or bending deformation, and distortion region markers containing coordinate ranges and distortion types are generated. These markers are usually stored in the form of binary masks or bounding box sets to indicate the specific locations that need correction. Finally, geometric correction processing is performed based on the distortion region markers to obtain the standard image. This process selects the appropriate correction strategy according to the deformation mode described by the distortion region markers: for perspective distortion caused by tilted shooting angle, four-point perspective transformation (Homography) is used to map irregular quadrilaterals back to standard rectangles; for local bending caused by paper curling, a mesh deformation model (such as thin plate spline interpolation) can be combined to perform non-rigid correction on the local area, so that the entire image restores a flat and regular layout structure. After the above series of processing steps, the final output standard work image has clear edges, low noise background and accurate geometric alignment, which can truly restore the layout and content of the original paper work.For example, when processing the original scanned image of a middle school math homework assignment, if the student is using a creased notebook and uneven lighting during scanning causes some areas to be dark and edges to be blurry, the system first converts it to grayscale and removes noise to eliminate the interference of paper texture on OCR. Then, it highlights the answer boxes for each question and the student's hand-drawn triangles through edge extraction, and sharpens these edges to make small angle symbols and auxiliary lines clearly visible. Next, it detects that the answer area in the lower right corner is deformed into a trapezoidal shape due to paper curling. Based on this, the system generates a distortion area marker and applies perspective transformation to stretch the area back to a standard rectangle. In the final standard homework image, all questions are neatly arranged, text edges are sharp, and graphic proportions are correct, providing high-quality input for subsequent block positioning and content separation. The entire process strictly follows the noun order and technical logic defined in the claims, ensuring that each operation is executable and reproducible, and the application scenario always focuses on the digital processing of physical homework papers without introducing any external assumptions or functional descriptions.

[0029] In a specific embodiment, the standard work image is segmented and located to obtain an answer area map, including: The standard work image is subjected to line detection by Hough transform to obtain a reference line set, and the reference line set is adaptively projected to obtain a projection density map. Density clustering is performed on the projected density map using maximum suppression to obtain a clustered region map, and the clustered region map is further subdivided into equal-interval subdivisions to obtain a raster matrix map; Based on the raster matrix diagram, the clustering region diagram is subjected to morphological processing to obtain a region segmentation diagram, and the boundary of the region segmentation diagram is traced to obtain a contour chain code diagram. The contour chain code map is fused into regions to obtain a merged region map, and the merged region map is constrained and optimized to obtain the answer region map.

[0030] Specifically, the standard assignment image is segmented and localized to obtain the answer area map. This process consists of a series of tightly linked executable steps. First, the standard assignment image is subjected to line detection using the Hough transform to obtain a reference line set. This step uses the Hough line detection algorithm to identify all line segments with significant gradient responses in the standard assignment image, including structured lines such as the exam paper border, question separators, and answer lines, and stores these lines in parametric form (such as ρ-θ coordinates) as the reference line set. Next, the reference line set is adaptively projected to obtain a projection density map. Specifically, the horizontal or vertical projection direction is dynamically selected according to the directional distribution of the reference lines. Each reference line is projected onto the corresponding coordinate axis along its normal vector direction, and the cumulative number of projection points within a unit interval is counted to generate a projection density map reflecting the density of the lines. This map is numerically represented as a one-dimensional or two-dimensional density distribution array. Subsequently, the projected density map is subjected to density clustering through maximum suppression to obtain a clustered region map. This step finds local maxima in the projected density map and removes noise peaks based on a set threshold. Then, with each effective maximum as the center, the density is expanded to both sides until the density drops to the valley, forming several continuous intervals. These intervals correspond to the boundaries of potential answer blocks and are finally combined to form a clustered region map. Then, the clustered region map is subdivided into equal intervals to obtain a raster matrix map. That is, according to the span of the clustered region and the preset typical answer row height or column width, each clustered region is evenly divided into several sub-units to construct a regular raster structure covering the entire image, forming a raster matrix map. This map organizes spatial location information in a row and column index manner. Based on this, morphological processing is performed on the clustered region map based on the raster matrix map to obtain a region block map. This operation uses the raster matrix map as a structuring element to perform a closing operation (dilation followed by erosion) on the clustered region map to fill in the region breaks caused by handwriting occlusion or scanning omissions, and remove isolated noise points, so that adjacent answer regions are connected into complete blocks, and the output is a region block map. Next, boundary tracing is performed on the region block map to obtain a contour chain code map. An eight-neighbor boundary tracing algorithm (such as Freeman chain code) is used to traverse the outer contour of each connected region pixel by pixel, record the movement direction sequence, and form a contour chain code map describing the shape and position of each answer region.Next, the outline chain code map is fused into regions to obtain a merged region map. This step calculates the distance, area ratio, and semantic context (such as whether they belong to multiple lines of answers under the same question number) between adjacent outlines. Outline splicing is performed on neighboring regions that meet the merging conditions. For example, multiple lines of handwritten answers scattered under the same major question are merged into a whole region, thereby generating a merged region map that is more in line with the actual answering logic. Finally, the merged region map is constrained and optimized to obtain the answer area map. Here, prior knowledge of the test paper template (such as the position of the question number and the fixed size of the answer box) or subject layout rules (such as the graphic area of ​​math questions is usually located on the right side of the text) are introduced to perform geometric verification and semantic filtering on the candidate blocks in the merged region map. Abnormal regions that do not meet the constraints (such as fragments that are too small or cross-question interference blocks) are removed, and the boundary coordinates are fine-tuned to align with the standard grid. Finally, a clear and accurate answer area map is output. For example, when processing the standard image of the aforementioned junior high school math homework, if the test paper contains five major questions, with the third question being a geometry proof question, and the student has drawn auxiliary graphics below it and written the reasoning process in three lines, the system first uses Hough transform to detect the top title line of the test paper, the horizontal lines separating each question, and the border lines of the graphics, forming a reference line set; then, these lines are projected onto the vertical axis to generate a projection density map where a high-density peak appears at the position of the third question; after maximum suppression, five main cluster regions are identified, and then they are subdivided into 15 grid units according to typical row height; morphological processing merges the three consecutive grids corresponding to the third question into a connected component, and boundary tracing generates a chain code representation of its irregular contour; in the region fusion stage, it is determined that the contour is spatially adjacent to and semantically related to the graphic region below, so it is merged; finally, through constraint optimization, it is confirmed that the merged region does not cross the starting line of the fourth question and is retained as a valid answer area. The final answer area map accurately covers the text and graphic answer content of the question, providing reliable input for subsequent content separation.

[0031] In a specific embodiment, the answer area map is subjected to content separation to obtain a set of answer elements, including: Texture features are extracted from the answer area map to obtain a texture feature map, and the printed and handwritten areas are segmented based on the texture feature map to obtain a printed area map and a handwritten area map; The printed area map is locally contrast-enhanced to obtain an enhanced printed area map, and the enhanced printed area map is then subjected to connected component segmentation and semantic clustering to obtain a title element matrix. Handwriting feature analysis is performed on the handwritten area map to obtain a handwriting feature vector set, and text and formula area classification is performed based on the handwriting feature vector set to obtain a text area map and a formula area map. The text region map and formula region map are subjected to structured analysis to obtain the analysis element group, and the analysis element group is semantically associated and integrated with the question element matrix to obtain the answer element set.

[0032] Specifically, the answer area image is separated into answer element sets. This process first extracts texture features from the answer area image to obtain a texture feature map. This step quantifies the texture complexity, directionality, and regularity within the neighborhood of each pixel by calculating texture descriptors such as Local Binary Pattern (LBP), Gray-Level Co-occurrence Matrix (GLCM), or Gabor filter response, generating a texture feature map that can distinguish different writing styles and printing characteristics. Next, based on the texture feature map, printed and handwritten regions are segmented to obtain printed and handwritten region maps. Specifically, a Support Vector Machine (SVM) or a lightweight convolutional neural network is used to classify each pixel or superpixel block in the texture feature map. Based on the difference that printed text typically has uniform, high-contrast, and periodic textures, while handwritten text exhibits irregularity, low-frequency variations, and stroke continuity, the image is divided into two types of regions, which are then output as binarized printed and handwritten region maps, respectively. Subsequently, local contrast enhancement is performed on the printed region map to obtain an enhanced printed region map. This operation employs adaptive histogram equalization (CLAHE) or the Retinex algorithm to dynamically adjust brightness and contrast within a local window of the printed area map, enhancing the distinction between character edges and background, especially addressing the blurry areas in the question stem caused by uneven scanning illumination, thereby generating an enhanced printed area map with clearer visuals and higher OCR recognition rate. Then, connected component segmentation and semantic clustering are performed on the enhanced printed area map to obtain the question element matrix. Here, the eight-neighbor connected component labeling algorithm is first used to aggregate foreground pixels into independent character blocks and remove noise points with too small an area. Then, the OCR engine is called to perform text recognition on each character block. Finally, based on the question number markers (such as "1.", "(2)"), keywords (such as "known", "proof"), and spatial arrangement relationships in the recognition results, character blocks belonging to the same question are aggregated into a semantic unit and organized into a structured matrix in row and column form, i.e., the question element matrix. At the same time, handwriting feature analysis is performed on the handwritten area map to obtain a handwriting feature vector set. This step extracts the centerline of the handwritten trajectory through skeletonization and calculates the curvature, length, number of intersections, stroke order, connectivity, and other geometric and topological properties of each stroke to form a high-dimensional handwriting feature vector set. Then, based on this handwriting feature vector set, text and formula regions are classified to obtain text region maps and formula region maps. Here, random forests or small neural networks are used to distinguish handwritten handwriting: if the features are short straight lines, closed loops, low intersection, and regular character spacing, they are classified as text; if they are multiple intersecting line segments, geometrically closed figures, or symbolic structures (such as square roots, triangle outlines, and angle markings), they are classified as formulas, and corresponding text region maps and formula region maps are generated accordingly. Next, the text region maps and formula region maps are subjected to structured analysis to obtain analytical element groups.For text region graphs, horizontal projection is first used to detect line blocks, and then vertical projection or sliding window segmentation is performed within each line. A handwriting recognition model (such as CRNN) is called to recognize characters, and the text is organized into a text sequence by line. For formula region graphs, contour extraction algorithms (such as Canny+ contour tracking) are used to obtain the graphic boundaries, and mathematical symbols (such as "∠", "△", "∥") are recognized in combination with a symbol template library. Then, a tree of relationships such as inclusion, adjacency, and parallelism between symbols is constructed through topological analysis. Finally, the text and formulas are converted into element records with type labels, coordinate information, and semantic structure, and merged into a parsed element group. Finally, the parsed element group is semantically associated and integrated with the question element matrix to obtain the answer element set. This integration process establishes a cross-modal mapping relationship based on the spatial proximity (e.g., below the same question number), logical dependency (e.g., the text description "as shown in the figure" points to a certain figure), and sequential consistency (e.g., the formula immediately follows the text derivation) of the original answer area diagrams. Heterogeneous elements such as the question stem, student handwritten text, numbers, formula symbols, and geometric figures are uniformly encoded into a structured set with a global ID, type label, knowledge point affiliation, and logical sequence number—that is, the answer element set. For example, in the aforementioned junior high school math homework, a student handwritten "∵AB=AC,∴∠B=∠C" along with an isosceles triangle figure in the answer area of ​​a geometry proof problem. The system first uses texture analysis to separate the printed question stem "As shown in the figure, in △ABC…" into a printed area map and the student's answer into a handwritten area map. After contrast enhancement, the printed area is identified by OCR as the question number "Question 3" and "Given AB=AC", and clustered into a question element matrix. The handwritten area is analyzed by handwriting, and the text is categorized into a text area map, while the triangle and its angle markings are categorized into a formula area map. After structured parsing, the text is identified as "AB=AC" and "∠B=∠C", and the graphics are identified as vertices A, B, C and base angle symbols. Finally, based on the positional relationship between the two within the same answer block, the system integrates the question elements and the student's answer elements to form a complete answer element set containing question stem information, handwritten reasoning, and geometric objects, providing atomic-level input for subsequent feature extraction and logical chain construction.

[0033] In a specific embodiment, feature extraction is performed on the answer element set to obtain element feature groups, and structural parsing is performed on the element feature groups to obtain the answer logic chain, including: Semantic hierarchical analysis is performed on the set of answer elements to obtain a hierarchical feature map, and deep semantic parsing is performed on the hierarchical feature map to obtain a semantic element matrix; Based on the semantic element matrix, knowledge points are mapped onto the answer element set to obtain a knowledge point association graph, and the knowledge point association graph is recursively expanded to obtain element feature groups. Temporal correlation analysis is performed on the feature group of the elements to obtain a reasoning time sequence diagram, and the reasoning time sequence diagram is divided into logical units to obtain a sequence of logical steps; Based on the logical step sequence, dependency relationships are constructed for the feature group to obtain a derivation dependency tree, and the path of the derivation dependency tree is restored to obtain the solution logic chain.

[0034] Specifically, feature extraction is performed on the set of answer elements to obtain feature groups, and structural analysis is performed on the feature groups to obtain the answer logic chain. This process first performs semantic hierarchical analysis on the set of answer elements to obtain a hierarchical feature map. This step, based on a pre-defined subject semantic ontology (such as the junior high school mathematics knowledge system), categorizes each element in the answer element set (including textual and numerical elements and formula and graphic elements) according to its semantic hierarchy. For example, "AB=AC" is categorized into the path "Geometric properties → Isosceles triangle → Equal sides", and "∠B=∠C" is categorized into the path "Angle relationships → Equal base angles". The hierarchical relationship of all elements is organized in a tree-like or directed acyclic graph form to form a hierarchical feature map. Next, deep semantic analysis is performed on the hierarchical feature map to obtain a semantic element matrix. Specifically, each node in the map is traversed, and its semantic type, logical role (such as premise, conclusion, intermediate derivation), numerical content, and dependence strength with other nodes are extracted. This information is encoded into a fixed-dimensional vector and arranged in matrix form according to the original answer order, i.e., the semantic element matrix. Subsequently, knowledge point mapping is performed on the answer element set based on the semantic element matrix to obtain a knowledge point association graph. This operation utilizes a knowledge point base defined in the curriculum standards (such as entries like "properties of isosceles triangles" and "criteria for congruent triangles") to match the elements represented by each row in the semantic element matrix to the most relevant knowledge point identifiers, recording the matching confidence and coverage. This constructs a many-to-many mapping graph with knowledge points as nodes and answer elements as edges, i.e., a knowledge point association graph. Then, the knowledge point association graph is recursively expanded to obtain element feature groups. A depth-first traversal algorithm is used here, starting from each knowledge point node and recursively tracing back all associated original answer elements and their semantic attributes. These elements are aggregated by knowledge point, deduplicated, and appended with semantic tags, coordinate indices, and type identifiers, ultimately generating structured element feature groups. Based on this, temporal association analysis is performed on the element feature groups to obtain a reasoning temporal graph. This step infers the sequential dependencies of each element in the problem-solving process based on the writing order, spatial arrangement (e.g., top to bottom, left to right), and the presence or absence of logical connectors (e.g., "∵" and "∴") in the original answer area. It then constructs a reasoning sequence diagram reflecting the student's thought process by connecting element nodes with temporal or logical sequential relationships using directed edges. Next, the reasoning sequence diagram is logically segmented to obtain a sequence of logical steps. Specifically, subgraphs with complete semantic loops (e.g., a triple of "condition → reasoning → conclusion") are identified, segmented as logical units, and arranged chronologically to form an ordered sequence of logical steps. Finally, dependency relationships are constructed on the element feature groups based on the logical step sequence to obtain a derivation dependency tree. The path of the derivation dependency tree is then restored to obtain the solution logic chain.In the dependency relationship construction stage, the conclusion element in each logical step is the parent node, and the prerequisite elements it depends on are the child nodes. A tree structure is built up layer by layer to ensure that each derivation has a clear source of evidence, forming a derivation dependency tree. In the path reconstruction stage, starting from the root node of the tree (the final answer or conclusion), the path is traced back along the dependency edges to the leaf nodes (initial known conditions or graphical attributes). The complete reasoning path is recorded, and each logical step on the path is linearized sequentially. The final output is a coherent, traceable, and clearly structured logical chain of solutions. For example, in the aforementioned junior high school math homework, a student answers an isosceles triangle proof problem. The answer element set includes the text "∵AB=AC" and "∴∠B=∠C", as well as the vertices A, B, and C and the base angle markings in the figure. The system first performs semantic hierarchical analysis, classifying "AB=AC" into the "side equality" level and "∠B=∠C" into the "angle equality" level, forming a hierarchical feature map. After deep analysis, a semantic element matrix is ​​generated, indicating that the two are the premise and conclusion, respectively. Knowledge point mapping associates this with the knowledge point "the two base angles of an isosceles triangle are equal," and recursively expands this to obtain a feature group containing the knowledge point ID and corresponding elements. Temporal analysis establishes a sequential relationship based on the writing order and the "∵…∴…" structure, dividing the system into two logical units. Dependency relationship construction uses "∠B=∠C" as the parent node, and "AB=AC" and the existence of the triangle in the figure as child nodes, forming a derivation dependency tree. After path reconstruction, the solution logic chain is output: "Given AB=AC → Triangle ABC is an isosceles triangle → Base angles ∠B and ∠C are equal." The entire process strictly follows the terminology and operational order defined in the claims, ensuring the executability of the technical implementation and consistency with the application scenario.

[0035] In a specific embodiment, the dependency relationship of the feature group is constructed based on the logical step sequence to obtain a derivation dependency tree, including: Semantic role labeling is performed on each logical step in the logical step sequence to obtain a step predicate-argument structure set, and cross-modal semantic alignment is performed between the step predicate-argument structure set and the feature group to obtain a step-feature binding mapping table. Based on the step-feature binding mapping table, a directed acyclic graph is constructed for the logical step sequence to obtain an initial derivation graph. Then, the dominant node identification and subtree merging operations are performed on the initial derivation graph to obtain the derivation dependency tree.

[0036] Specifically, dependency relationships are constructed on the feature groups based on the logical step sequence to obtain a derivation dependency tree. This process first involves semantic role labeling of each logical step in the logical step sequence to obtain a step predicate-argument structure set. Specifically, for each reasoning unit expressed in natural language or symbolic form in the logical step sequence (e.g., "∵AB=AC,∴△ABC is an isosceles triangle"), a pre-trained semantic role labeling model is invoked to identify the core predicate (e.g., "is", "equals", "infers") and various arguments surrounding the predicate, including premise (Arg0), conclusion (Arg1), and basis (ArgM-CAU), etc., and each logical step is transformed into a structured set of predicate-argument triples, thereby forming a step predicate-argument structure set. Next, cross-modal semantic alignment is performed between the step predicate-argument structure set and the feature group to obtain a step-feature binding mapping table. This operation calculates the similarity between the semantic content expressed by each argument (e.g., "AB=AC") and the semantic embedding vectors of each element in the feature group, and combines spatial location consistency (e.g., being within the same answer area block) and type matching rules (e.g., text arguments are only aligned with text and number elements, and angular descriptions are only aligned with formula and graphic elements) to precisely bind each argument to the corresponding unique element instance in the feature group. This ultimately generates a structured mapping table recording "logical step number—argument role—element ID," i.e., the step-feature binding mapping table. Subsequently, a directed acyclic graph is constructed based on the step-feature binding mapping table for the logical step sequence to obtain the initial derivation graph. This step uses each step in the logical sequence as a graph node. If the element bound to the conclusion argument of one step is exactly the element bound to the premise argument of another step, a directed edge is drawn from the former to the latter, indicating that the latter depends on the output of the former. For steps that directly reference the original problem statement conditions or graphical attributes, no incoming edges are set, serving as the starting point for reasoning. The graph structure constructed in this way ensures acyclicity and reflects the actual reasoning flow, which is the initial derivation graph. Based on this, the dominator node identification and subtree merging operations are performed on the initial derivation graph to obtain the derivation dependency tree. Dominator node identification determines the dominance relationship of each node in the reasoning chain by traversing all paths in the graph—if all paths from the root node to a certain node must pass through node X, then X is its dominator node. The subtree merging operation targets multiple child nodes with the same dominator node and close semantic relationship (e.g., two derivation branches using the same formula in parallel), merging them into a composite subtree, retaining only one parent connection, thereby eliminating the multi-parent node structure and converting the directed acyclic graph into a strict tree structure. In the final output derivation dependency tree, each non-root node has one and only one direct parent node, which fully presents the dependency hierarchy and reasoning path between each logical step in the solution process.For example, in the aforementioned junior high school math homework, the student's answer involves three logical steps: S1 is "AB=AC" (given condition), S2 is "△ABC is an isosceles triangle" (derived from S1), and S3 is "∠B=∠C" (derived from S2 and the properties of isosceles triangles). The system performs semantic role labeling on S2, obtaining the predicate "is", Arg0 is "△ABC", and Arg1 is "isosceles triangle"; and labels S3, obtaining the predicate "equal", Arg0 is "∠B", Arg1 is "∠C", and ArgM-CAU is "properties of isosceles triangles". Through cross-modal alignment, "AB=AC" in S1 is bound to the handwritten text element E3 in the feature group, "△ABC" in S2 is bound to the graphic element G1, and the angle symbol in S3 is bound to the graphic elements A5 and A6, forming a step-feature binding mapping table. When constructing the initial derivation graph, since the premise of S2 depends on the conclusion of S1, S1→S2; the causal basis of S3 depends on the conclusion of S2, so S2→S3. Since the graph is a linear chain structure, the dominating nodes are identified and confirmed that S1 dominates S2 and S2 dominates S3. Without complex merging, the output is directly the derivation dependency tree.

[0037] In a specific embodiment, error association analysis is performed on the standard operation image based on the potential suspicious point map to obtain an error association network, including: Spatial coordinate mapping is performed on the potential suspicious point image to obtain a suspicious point pixel coordinate set, and local content backtracking is performed on the standard operation image based on the suspicious point pixel coordinate set to obtain a suspicious point context image block. The context image blocks of the doubts are semantically aligned to obtain an aligned element sequence. Based on the aligned element sequence, bidirectional index matching is performed with the answer element set to obtain a doubt-element mapping table. Based on the aforementioned question-element mapping table, the solution logic chain is expanded in reverse to obtain the logic breakpoint trajectory, and the logic breakpoint trajectory is reconstructed by causal dependency to obtain the error propagation subgraph. The error propagation subgraph is expanded by multiple order neighborhoods to obtain an expanded error topology. The intersection constraint between the expanded error topology and the knowledge point association graph is then applied to obtain the error association network.

[0038] Specifically, error association analysis is performed on the standard operation image based on the potential suspicious point map to obtain an error association network. This process first maps the potential suspicious point map to spatial coordinates to obtain a set of suspicious point pixel coordinates. This step involves parsing all regions marked as suspicious points in the potential suspicious point map and converting their positions on the image plane into a specific list of pixel coordinates. For example, the (x, y) coordinates of the four vertices of the bounding box of each suspicious point region or all foreground pixels are recorded to form a set of suspicious point pixel coordinates. Next, local content backtracking is performed on the standard operation image based on the set of suspicious point pixel coordinates to obtain suspicious point context image blocks. Specifically, each suspicious point pixel coordinate is used as the center, and the corresponding local image region is cropped from the standard operation image according to a preset size (e.g., 100×100 pixels), preserving its original grayscale and structural information to generate a set of image blocks reflecting the visual content around the suspicious points, i.e., suspicious point context image blocks. Subsequently, semantic alignment of the suspicious context image blocks is performed to obtain an aligned element sequence. This operation calls the same recognition model as the previous content separation stage, and performs lightweight element extraction again on each suspicious context image block, including handwritten text recognition, formula symbol detection, and graphic contour parsing. The recognition results are then semantically matched with existing elements in the answer element set. For example, the "∠B=∠C" recognized in the image block is aligned to the angle equivalent statement element with ID E12 in the answer element set. These successfully matched elements are then arranged according to spatial writing order to form an aligned element sequence. Then, bidirectional index matching is performed based on the aligned element sequence and the answer element set to obtain a suspicious point-element mapping table. Here, by comparing element IDs, the global identifier of each element in the aligned element sequence in the answer element set is located, and a "suspicious point region number" is established. The bidirectional mapping relationship of "element ID" ultimately outputs a structured question-element mapping table. Based on this, the solution logic chain is expanded in reverse order using this question-element mapping table to obtain a logic breakpoint trajectory. This step starts from the element associated with each question in the mapping table, locates its corresponding logical step node in the solution logic chain, and traces back up the derivation dependency tree to the inference starting point or down to the conclusion node. If a step is found to have missing premises, a conclusion that conflicts with axioms, or a misuse of inference rules, this location is marked as a logic breakpoint, and all breakpoints are connected in the order of inference to form a trajectory, i.e., the logic breakpoint trajectory. Next, the logic breakpoint trajectory undergoes causal dependency reconstruction to obtain an error propagation subgraph. This operation, based on the parent-child dependency relationship in the derivation dependency tree, expands downwards from each logic breakpoint as the source, encompassing all subsequent logic step nodes affected by it, and connects causally related nodes with directed edges to construct a local... The subgraph reflects how errors propagate from initial doubts to the final conclusion; this is the error propagation subgraph. Finally, multi-order neighborhood expansion is performed on the error propagation subgraph to obtain an expanded error topology. Intersection constraints are then applied between this expanded error topology and the knowledge point association graph to obtain the error association network. Multi-order neighborhood expansion refers to expanding adjacent logical steps (including sibling nodes and indirectly dependent nodes) outwards along the derivation dependency tree from each node in the error propagation subgraph as the center, capturing potentially implicitly affected reasoning steps and forming a broader-coverage expanded error topology. Subsequently, the knowledge points corresponding to each node in this topology (mapped from the feature groups) are intersected with the pre-constructed knowledge point association graph. Only knowledge points and their associated edges that exist in both graphs are retained, while connections unrelated to the subject knowledge system are removed, ultimately outputting a structurally rigorous and semantically reliable error association network.For example, in the aforementioned junior high school math homework, if a student incorrectly uses the "Side-Side-Angle (SSA)" condition when proving triangle congruence, the system detects this violation of geometric axioms in the solution logic chain and generates a potential point of doubt diagram; after spatial coordinate mapping, it obtains the position of this step in the standard homework image and crops out the part containing "AB=DE, AC=DF, The context image block of "∠B=∠E" is generated; the semantic alignment of the elements identifies these three elements and matches them with E5, E6, and E7 in the answer element set to form an aligned element sequence; a bidirectional index generates a doubt point-element mapping table; the reverse path expansion reveals that this step is directly used as a premise for the congruence judgment, but it is marked as "SSA is not true" in the knowledge point base, so it is marked as a logical breakpoint; causal dependency reconstruction incorporates the erroneous step and its subsequent conclusion "BC=EF" into the error propagation subgraph; multi-order neighborhood expansion further incorporates the subsequent step of proving similar triangles with "BC=EF"; finally, the intersection with the knowledge point association graph (including nodes such as "congruence judgment" and "similar triangles") is taken to confirm that all errors belong to the category of geometric reasoning, and the output error association network clearly reveals that "SSA misuse" is the root error, causing subsequent derivations to fail.

[0039] In a specific embodiment, the propagation and tracing of the error association network are performed to obtain a hidden error report, including: The error association network is analyzed for topology and the importance of nodes is calculated to obtain a set of key error nodes. Based on the set of key error nodes, the reverse path is traced to obtain the main path of error propagation. The error propagation main path is expanded by multiple levels of neighborhood and the influence weight is accumulated to obtain an expanded influence node set. The expanded influence node set is then mapped and aligned with the knowledge point association graph to obtain the knowledge defect association area. The knowledge defect association area is classified and labeled according to the cognitive dimension to obtain a defect type matrix. Based on the defect type matrix and the main error transmission path, a temporal correlation integration is performed to obtain a hidden error report.

[0040] Specifically, the propagation and tracing of the error association network yields latent error reports. This process first involves topological analysis and node importance calculation of the error association network to obtain a set of key error nodes. This step uses graph theory algorithms (such as PageRank, betweenness centrality, or in-degree-out-degree weighted scoring) to traverse all nodes in the error association network, calculating the structural influence of each node in the error propagation path. For example, nodes located at the intersection of multiple propagation paths or upstream in the inference chain receive higher weights. Nodes with weights exceeding a preset threshold are then selected to form the set of key error nodes. Next, reverse path tracing is performed based on this set of key error nodes to obtain the main error propagation path. Specifically, starting from each key error node, the path is backtracked along the directed edges in the error association network to the initial point of suspicion or a known condition node. The path is recorded, and multiple backtracking paths are merged and pruned, retaining the longest shared path that covers the most error impact. The final output is a main error propagation path reflecting the core error transmission mechanism. Based on this, multi-order neighborhood expansion and influence weight accumulation are performed on the main error propagation path to obtain an expanded influence node set. This operation centers on each node in the main error propagation path and expands outwards to adjacent nodes of order one, two, and even three (including sibling nodes, indirectly dependent nodes, and affected subsequent conclusion nodes). The original weights of each node in the main path are accumulated to neighboring nodes according to a distance decay rule (e.g., exponential decay), forming an expanded set of affected nodes with cumulative impact scores. Then, this expanded set of affected nodes is mapped and aligned with the knowledge point association graph to obtain the knowledge defect association region. Here, the intersection of the knowledge point ID bound to each node in the expanded set of affected nodes and the node identifier in the knowledge point association graph is extracted, and their topological connections are preserved, thus defining a subgraph region focused on a specific knowledge domain, namely the knowledge defect association region. Subsequently, the knowledge defect association region is classified and labeled according to cognitive dimensions to obtain a defect type matrix. This step, based on cognitive classification systems in educational psychology (such as Bloom's Taxonomy or the dimensions specific to mathematics, such as "conceptual understanding," "rule application," and "logical reasoning"), assigns one or more cognitive labels to each knowledge point node in the knowledge deficiency association area. It then combines this with the error manifestation (e.g., confusing SSA with SAS is "rule misuse," ignoring implicit conditions is "information extraction deficiency") to generate a structured labeling result, ultimately organizing it into a deficiency type matrix with rows and columns corresponding to knowledge points and cognitive dimensions. Finally, based on the deficiency type matrix and the main error propagation path, a temporal correlation integration is performed to obtain the implicit error report. This integration process arranges the cognitive labels in the deficiency type matrix according to the node order on the main error propagation path, and adds the influence weight of each node, the original answer content fragment, and the corresponding standard knowledge point description, forming a structured text containing the root cause of the error, the propagation path, the affected knowledge points, the cognitive deficiency type, and the temporal evolution logic—that is, the implicit error report.For example, in the aforementioned junior high school math homework, if the error network reveals that a student's misuse of the Side-Side-Angle (SSA) method to determine congruence led to errors in the subsequent two steps of derivation, the system first calculates the importance of each node and finds that the "SSA application" node has the highest betweenness centrality, thus being selected into the set of key error nodes; tracing back to the source reveals that the error stems from "AB=DE, AC=DF, The error propagation path from "∠B=∠E" to "△ABC≌△DEF" is identified; multi-level neighborhood expansion incorporates subsequent steps using "BC=EF" to prove similar triangles, and accumulates influence weights; after aligning with the knowledge point association graph, the knowledge defect association area consisting of the two knowledge points "Triangle Congruence Criterion" and "Properties of Similar Triangles" is locked; cognitive dimension classification labels mark the former as "Rule Misuse" and the latter as "Broken Reasoning Chain"; finally, the implicit error report integrates this information in chronological order and outputs: "The student incorrectly applied the SSA criterion in step 2 (Rule Misuse), causing the congruence conclusion to be invalid, which in turn caused the similarity reasoning in step 3 based on the incorrect side length equality to fail (Broken Reasoning Chain). The root cause is the failure to grasp the core concept that 'SSA cannot determine congruence'."

[0041] The above describes the job analysis and error diagnosis method in the embodiments of the present invention. The following describes the job analysis and error diagnosis system in the embodiments of the present invention. Please refer to [link / reference]. Figure 2 One embodiment of the job analysis and error diagnosis system in this invention includes: The acquisition module 21 is used to scan and acquire physical homework test papers to obtain original homework images, and to perform edge enhancement and distortion correction on the original homework images to obtain standard homework images. The positioning module 22 is used to perform block positioning on the standard operation image to obtain the answer area map, and to perform content separation on the answer area map to obtain the answer element set; The extraction module 23 is used to extract features from the answer element set to obtain element feature groups, and to perform structural analysis on the element feature groups to obtain the answer logic chain; The detection module 24 is used to detect whether there are potential doubts in the solution logic chain. If there are, it performs error correlation analysis on the standard operation image based on the potential doubt map to obtain an error correlation network, and tracks the propagation of the error correlation network to obtain a hidden error report.

[0042] In this embodiment, the specific implementation of each module in the above system embodiment is described in the above method embodiment, and will not be repeated here.

[0043] Reference Figure 3 This invention also provides a computer device whose internal structure can be as follows: Figure 3As shown, the computer device includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores the data corresponding to this embodiment. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0044] Those skilled in the art will understand that Figure 3 The structures shown are merely block diagrams of some structures related to the present invention and do not constitute a limitation on the computer devices on which the present invention is applied.

[0045] An embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method. It is understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0046] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the present invention and embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM, etc.

[0047] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0048] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for job analysis and error diagnosis, characterized in that, Includes the following steps: The physical work test paper is scanned and collected to obtain the original work image, and the edge enhancement and distortion correction of the original work image are performed to obtain the standard work image; The standard operation image is segmented and located to obtain the answer area map, and the answer area map is separated to obtain the answer element set; Feature extraction is performed on the set of answer elements to obtain feature groups, and structural analysis is performed on the feature groups to obtain the solution logic chain; If there are potential issues in the solution logic chain, then an error correlation analysis is performed on the standard operation image based on the potential issue graph to obtain an error correlation network. The error correlation network is then propagated and tracked to obtain a hidden error report.

2. The job analysis and error diagnosis method according to claim 1, characterized in that, The original work image is subjected to edge enhancement and distortion correction to obtain a standard work image, including: The original work image is subjected to grayscale noise reduction processing to obtain a noise-reduced work image; Edge features are extracted from the noise-reduced image to obtain an edge-enhanced image, and local sharpening is performed on the edge-enhanced image to obtain a sharpened image. The sharpened image is subjected to distortion region detection to obtain distortion region markers, and geometric correction processing is performed based on the distortion region markers to obtain a standard image.

3. The job analysis and error diagnosis method according to claim 1, characterized in that, The standard work image is segmented and located to obtain the answer area map, including: The standard work image is subjected to line detection by Hough transform to obtain a reference line set, and the reference line set is adaptively projected to obtain a projection density map. Density clustering is performed on the projected density map using maximum suppression to obtain a clustered region map, and the clustered region map is further subdivided into equal-interval subdivisions to obtain a raster matrix map; Based on the raster matrix diagram, the clustering region diagram is subjected to morphological processing to obtain a region segmentation diagram, and the boundary of the region segmentation diagram is traced to obtain a contour chain code diagram. The contour chain code map is fused into regions to obtain a merged region map, and the merged region map is constrained and optimized to obtain the answer region map.

4. The job analysis and error diagnosis method according to claim 1, characterized in that, The answer area diagram is split into content to obtain a set of answer elements, including: Texture features are extracted from the answer area map to obtain a texture feature map, and the printed and handwritten areas are segmented based on the texture feature map to obtain a printed area map and a handwritten area map; The printed area map is locally contrast-enhanced to obtain an enhanced printed area map, and the enhanced printed area map is then subjected to connected component segmentation and semantic clustering to obtain a title element matrix. Handwriting feature analysis is performed on the handwritten area map to obtain a handwriting feature vector set, and text and formula area classification is performed based on the handwriting feature vector set to obtain a text area map and a formula area map. The text region map and formula region map are subjected to structured analysis to obtain the analysis element group, and the analysis element group is semantically associated and integrated with the question element matrix to obtain the answer element set.

5. The job analysis and error diagnosis method according to claim 1, characterized in that, Feature extraction is performed on the set of answer elements to obtain feature groups, and structural parsing is performed on the feature groups to obtain the solution logic chain, including: Semantic hierarchical analysis is performed on the set of answer elements to obtain a hierarchical feature map, and deep semantic parsing is performed on the hierarchical feature map to obtain a semantic element matrix; Based on the semantic element matrix, knowledge points are mapped onto the answer element set to obtain a knowledge point association graph, and the knowledge point association graph is recursively expanded to obtain element feature groups. Temporal correlation analysis is performed on the feature group of the elements to obtain a reasoning time sequence diagram, and the reasoning time sequence diagram is divided into logical units to obtain a sequence of logical steps; Based on the logical step sequence, dependency relationships are constructed for the feature group to obtain a derivation dependency tree, and the path of the derivation dependency tree is restored to obtain the solution logic chain.

6. The job analysis and error diagnosis method according to claim 5, characterized in that, Based on the potential problem map, error correlation analysis is performed on the standard operation image to obtain an error correlation network, including: Spatial coordinate mapping is performed on the potential suspicious point image to obtain a suspicious point pixel coordinate set, and local content backtracking is performed on the standard operation image based on the suspicious point pixel coordinate set to obtain a suspicious point context image block. The context image blocks of the doubts are semantically aligned to obtain an aligned element sequence. Based on the aligned element sequence, bidirectional index matching is performed with the answer element set to obtain a doubt-element mapping table. Based on the aforementioned question-element mapping table, the solution logic chain is expanded in reverse to obtain the logic breakpoint trajectory, and the logic breakpoint trajectory is reconstructed by causal dependency to obtain the error propagation subgraph. The error propagation subgraph is expanded by multiple order neighborhoods to obtain an expanded error topology. The intersection constraint between the expanded error topology and the knowledge point association graph is then applied to obtain the error association network.

7. The job analysis and error diagnosis method according to claim 1, characterized in that, The propagation and tracing of the aforementioned error network yields a hidden error report, including: The error association network is analyzed for topology and the importance of nodes is calculated to obtain a set of key error nodes. Based on the set of key error nodes, the reverse path is traced to obtain the main path of error propagation. The error propagation main path is expanded by multiple levels of neighborhood and the influence weight is accumulated to obtain an expanded influence node set. The expanded influence node set is then mapped and aligned with the knowledge point association graph to obtain the knowledge defect association area. The knowledge defect association area is classified and labeled according to the cognitive dimension to obtain a defect type matrix. Based on the defect type matrix and the main error transmission path, a temporal correlation integration is performed to obtain a hidden error report.

8. A job analysis and error diagnosis system, characterized in that, The method for performing job analysis and error diagnosis according to any one of claims 1 to 7 includes: The acquisition module is used to scan and acquire physical homework test papers to obtain original homework images, and to perform edge enhancement and distortion correction on the original homework images to obtain standard homework images; The positioning module is used to segment and locate the standard operation image to obtain the answer area map, and to separate the content of the answer area map to obtain the answer element set; The extraction module is used to extract features from the set of answer elements to obtain feature groups, and to perform structural analysis on the feature groups to obtain the solution logic chain. The detection module is used to detect whether there are potential doubts in the solution logic chain. If so, it performs error correlation analysis on the standard operation image based on the potential doubt map to obtain an error correlation network, and tracks the propagation of the error correlation network to obtain a hidden error report.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images