A method for intelligent identification and self-adaptive trace printing of work paper

Abstract

The present application relates to the technical field of image processing, and more particularly to a method for intelligent recognition of work paper and adaptive trace printing, comprising the following steps: obtaining a digital image; performing edge detection and contour extraction on the digital image to identify a closed quadrilateral region representing the boundary of the paper; calculating the aspect ratio of the circumscribed rectangle of the quadrilateral region and combining the image DPI information to calculate the actual physical size; matching the calculated size with a standard paper type in a preset education field paper knowledge graph (EPKG) through a matching algorithm to determine the target paper type; automatically loading a corresponding trace rendering template according to the identified paper type, wherein the trace rendering template contains spatial coordinate anchor points for each question; mapping the correction result data of the intelligent correction system to the corresponding coordinate positions of the trace rendering template to generate an output file with correction traces; and controlling a printing device to perform a printing operation according to the default printing configuration of the paper type.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method for intelligent paper recognition and adaptive trace printing. Background Technology

[0002] In the current process of intelligent education, the grading of student assignments is gradually moving towards automation and digitalization. However, after intelligent grading, accurately "leaving traces" of the grading results and printing them back onto the original paper assignments still largely relies on manual pre-selection of paper size. Because student assignments come from diverse sources and paper specifications are inconsistent, and issues such as distortion, tilting, and obstruction often exist during photography or scanning, current technology lacks the ability to automatically recognize paper types. This necessitates manual intervention from the user to set the paper type, which is not only cumbersome but also prone to matching errors, affecting the accuracy of trace positioning and printing efficiency, thus hindering the development of fully automated grading.

[0003] Chinese Patent Publication No. CN109101202A discloses a paper exam paper marking machine and marking method, comprising: the marking machine includes an image acquisition module, an image preprocessing module, a marking mark image generation module, and a printing device; the image acquisition module acquires the pattern content of the paper exam paper to be printed; the image preprocessing module performs noise reduction, shaping, and correction on the acquired image; the image recognition and judgment module recognizes and judges the acquired pattern according to rules; the network communication module interacts with the backend through the network; the print image generation module generates the print pattern according to the recognition and judgment results; and the printing device prints the generated pattern onto the original paper.

[0004] Therefore, while the proposed solution can generate personalized traceability patterns based on the answers to each test paper, it suffers from the following problems:

[0005] This solution uses a fixed trace layout generation logic and lacks the ability to automatically recognize the physical size of the paper. It cannot automatically switch the corresponding trace template and printing configuration according to the recognized paper type. The trace printing still relies on manually preset paper type or fixed layout, which makes it difficult to solve the problems of trace misalignment and low efficiency caused by inconsistent paper specifications in mixed operation scenarios. Summary of the Invention

[0006] To address this, the present invention provides a method for intelligent paper identification and adaptive trace printing, which overcomes the problems in the prior art where manual intervention is required to set the paper type due to the inability to automatically identify the paper type, and the inability to adaptively match the trace template, resulting in trace misalignment.

[0007] To achieve the above objectives, the present invention provides a method for intelligent identification and adaptive trace printing of printing paper, comprising:

[0008] S1, acquire the digitized image;

[0009] S2, perform edge detection and contour extraction on the digital image to identify the closed quadrilateral region representing the paper boundary;

[0010] S3, calculate the aspect ratio of the circumscribed rectangle of the quadrilateral region, and estimate the actual physical size by combining the image DPI information;

[0011] S4. The calculated size is matched with the standard paper type in the preset educational paper knowledge graph EPKG using a matching algorithm to determine the target paper type.

[0012] S5. Based on the identified paper type, automatically load the corresponding trace rendering template, which contains the spatial coordinate anchor points of each question.

[0013] S6, map the correction result data of the intelligent correction system to the corresponding coordinate position of the trace rendering template, and generate an output file with correction traces;

[0014] S7, control the printing device to perform a printing operation according to the default printing configuration of the paper type.

[0015] Furthermore, the matching algorithm adopts a fusion decision mode, and its confidence output adopts a multi-source weighted calculation method; the matching process of the matching algorithm adopts a dual-channel collaborative decision architecture, including a main channel for geometric matching and an auxiliary channel for CNN classification.

[0016] Furthermore, the process by which the main channel determines the confidence level through geometric matching includes:

[0017] The physical dimensions of the circumscribed rectangle of the quadrilateral region are determined based on edge detection;

[0018] Based on the physical dimensions, a nearest neighbor search with fuzziness tolerance is performed in the paper knowledge graph (EPKG) for the education field;

[0019] The search yielded three key elements for joint calculation: dimensional deviation normalization score, aspect ratio consistency score, and geometric stability score.

[0020] The confidence level is calculated by combining the three factors.

[0021] Furthermore, the calculation process for the normalized score of the size deviation includes:

[0022] Physical size restoration: Read the image DPI information from image metadata or default configuration, and convert the detected paper quadrilateral pixel dimensions in the image into physical measurement dimensions. The conversion formula is as follows.

[0023]

[0024]

[0025] Where Lmeas represents the physical measurement dimension length, Wmeas represents the physical measurement dimension width, Lpixel represents the pixel length, and Wpixel represents the pixel width;

[0026] Standard data retrieval: Based on the initially identified paper type candidates, retrieve the standard physical dimensions corresponding to the paper type from the Education-Specific Paper Knowledge Graph (EPKG);

[0027] Deviation mapping to score: The average value of the relative error of the length and width of the physical measurement paper quadrilateral is calculated by the following formula (1), and the error is mapped to a value between 0 and 1 by the following formula (2) exponential decay function to obtain the normalized score of the dimensional deviation:

[0028] (1)

[0029] (2)

[0030] Where Esize represents the average value of the relative error of the length and width of the quadrilateral of the paper in physical measurement, Lstd represents the standard physical dimension length corresponding to the paper type, Wstd represents the standard physical dimension length corresponding to the paper type, Ssize represents the normalized score of the size deviation, and α represents the sensitivity coefficient.

[0031] Furthermore, the calculation process for the aspect ratio consistency score includes:

[0032] Calculation of measured aspect ratio: The measured aspect ratio of the quadrilateral boundary detected in the image is calculated using the following formula, and the ratio is greater than 1:

[0033]

[0034] Where Rmeas represents the aspect ratio of the measured paper quadrilateral, max(Lmeas,Wmeas) represents selecting the maximum value among the physical measurement of the paper quadrilateral's length and width, and min(Lmeas,Wmeas) represents selecting the minimum value among the physical measurement of the paper quadrilateral's length and width.

[0035] Theoretical aspect ratio matching: Obtain the theoretical aspect ratio of candidate paper types from the Education-Specific Paper Knowledge Graph (EPKG);

[0036] Proportion Difference Score: The absolute difference between the measured proportion and the theoretical proportion is calculated using the following formula (1), and the aspect ratio consistency score is generated using the following formula (2) Gaussian function:

[0037] (1)

[0038] (2)

[0039] Where Δratio represents the absolute difference between the measured ratio and the theoretical ratio, Rstd represents the theoretical aspect ratio of the paper quadrilateral, Sratio represents the aspect ratio consistency score, and β represents the discrimination coefficient.

[0040] Furthermore, the calculation process for the geometric stability score includes:

[0041] Edge straightness fitting: For each of the four edges of the detected paper quadrilateral, perform least squares straightness fitting, calculate the goodness-of-fit R² value for each edge, and calculate the average straightness score using the following formula:

[0042]

[0043] Among them, Score line r represents the average straightness score. i This represents the R² value of the four sides of the quadrilateral on the paper.

[0044] Vertex fitting error calculation: Compare the coordinates of the four detected corner points with the coordinates of the vertices of the ideal quadrilateral formed by the intersection of the four fitted lines, and calculate the Euclidean distance error. Let the average pixel error of the four vertices be denoted as . Set a maximum allowable error, and calculate the vertex score using the following formula:

[0045]

[0046]

[0047] Where Evertex represents the average pixel error of the four vertices, P (j) detected P represents the coordinates of the four detected corner points. (j) ideal Score represents the coordinates of the vertices of an ideal quadrilateral formed by the intersection of four fitted lines. vetex Tvertex represents the vertex score, and Tvertex represents the maximum allowable error. This indicates a function that selects the maximum value.

[0048] Stability comprehensive score: The straightness score and vertex score are weighted and combined using the following formula:

[0049] Sstab=W line •Score line +W vertex •Score vertex

[0050] Where Sstab represents the geometric stability score, W line W represents the internal weight of straightness. vertex This represents the weight inside a vertex.

[0051] Furthermore, the calculation steps for the geometric matching confidence score include:

[0052] Dynamic weight allocation: Based on the image quality, job type, and regional habits of the paper, dynamic weights are assigned to the three scores; if the image clarity does not meet the standard, the proportions of the dynamic weights of the size deviation normalization score and the aspect ratio consistency score are reduced, while the sensitivity weight for edge quality in the dynamic weight of the geometric stability score is increased.

[0053] Weighted summation: The original confidence level is obtained by linearly weighting each score based on the adjusted weights, as shown in the following formula:

[0054] Craw=Wsize•Ssize+Wratio•Sratio+Wstab•Sstab

[0055] Where Craw represents the original confidence level, Wsize represents the dynamic weight of the size deviation normalization score, Wratio represents the dynamic weight of the aspect ratio consistency score, and Wstab represents the dynamic weight of the geometric stability score.

[0056] Nonlinear calibration and penalty: If the dynamic weight of the geometric stability score is lower than the threshold of the dynamic weight of the geometric stability score, it indicates that the geometric quality of the image is extremely poor. Then, a penalty coefficient is applied to the final geometric matching confidence. The formula for calculating the geometric matching confidence is as follows:

[0057]

[0058] Where Confgeo represents the geometric matching confidence, λ represents the penalty coefficient, and Tstab represents the geometric stability score threshold.

[0059] Normalized output: Ensure the final output value falls between 0 and 1, and record this value as the normalized geometric matching confidence score, calculated as follows:

[0060] Confgeo1=min(1,max(0,Confgeo))

[0061] Where Confgeo1 represents the normalized geometric matching confidence score, max(0,Confgeo) represents the maximum value between 0 and the geometric matching confidence score, and min(1,max(0,Confgeo)) represents the minimum value between 1 and max(0,Confgeo).

[0062] Furthermore, the calculation process for the auxiliary channel includes:

[0063] Start the lightweight CNN model, input the entire image or ROI region, and output the probability distribution of each paper pattern;

[0064] The geometric matching confidence score obtained through the main channel is compared with a pre-stored threshold.

[0065] If the geometric matching confidence score is greater than or equal to the threshold, the geometric matching confidence score is adopted, and the geometric matching confidence score is weighted and fused with the preset CNN classification confidence score using the following formula to obtain the comprehensive confidence score:

[0066] Final_Score=Wgeo•Confgeo1+(1−Wgeo)•Confcnn

[0067] Where Final_Score represents the overall confidence score, Wgeo represents the fusion coefficient dynamically adjusted based on the image quality, job type, and regional habits of the paper, and Confcnn represents the CNN classification confidence score;

[0068] If the geometric matching confidence is less than the threshold, the CNN channel is triggered, and the overall confidence is calculated according to the following formula:

[0069] Final_Score=a×Confgeo1+b×Confcnn+c×Context_Boost

[0070] Where a represents the geometric weight coefficient, b represents the neural network weight coefficient, c represents the EPKG boosting weight factor, and Context_Boost represents the EPKG usage frequency boosting score.

[0071] Furthermore, step 6 of the method introduces a dynamic coordinate mapping mechanism, which precisely projects the correction marks onto the spatial coordinate system of the original work image through perspective transformation based on the actual resolution of the input image and the scanning distortion, so as to ensure that the mark position is strictly aligned with the original question position.

[0072] Furthermore, the intelligent paper recognition and adaptive trace-keeping printing system of the method is characterized by comprising:

[0073] An image acquisition unit is used to acquire digital images;

[0074] A paper recognition engine, which is connected to the image acquisition unit, is used to integrate edge detection, geometric analysis and template matching functions;

[0075] The educational paper knowledge graph storage module is connected to the paper recognition engine and is used to store the typical worksheet specifications used in different regions, grades and textbook versions, as well as the standard size, common layout methods and question area distribution patterns of each specification.

[0076] A dual-mode recognition decision-maker is connected to the paper recognition engine and the educational paper knowledge graph storage module. It is used to combine the paper feature extraction results and knowledge graph information to identify and determine the type of homework paper. The dual-mode recognition decision-maker includes a rule engine and a deep learning classifier.

[0077] The trace rendering template library is connected to the dual-mode recognition decision unit and is used to organize the rendering modules of various types of correction traces according to paper type, and to retrieve the matching trace template according to the paper recognition result.

[0078] The dynamic mapping and printing control unit is connected to the dual-mode recognition decision unit and the trace rendering template library to generate documents with correction marks and drive the printer to automatically execute the output.

[0079] The closed-loop feedback module, which is connected to the paper recognition engine and the dynamic mapping and printing control unit, is used to collect post-printing trace effect evaluation data, including user correction records or automatic comparison of image deviations before and after printing; the closed-loop feedback module is also used to iteratively optimize the parameter thresholds and template library accuracy of the paper recognition algorithm.

[0080] Compared with existing technologies, the beneficial effects of this invention are as follows: by introducing a fusion of edge detection, geometric feature analysis, a knowledge graph of paper in the education field, and a lightweight AI classification model, this invention achieves high-precision automatic identification of student homework paper types and supports intelligent matching of various common specifications and their regional layouts; furthermore, by introducing a dynamic coordinate mapping engine to accurately restore the marking marks to the original question positions, it significantly improves the accuracy and automation of the marking printing; at the same time, the introduction of a closed-loop feedback mechanism continuously optimizes the recognition performance, reduces the need for manual intervention, improves marking efficiency and user experience, is suitable for large-scale homework processing scenarios, and has good scalability and industrial application value.

[0081] Furthermore, the matching algorithm described in this invention employs a fusion-based decision-making model and multi-source weighted calculation of confidence output. This enables intelligent evaluation across multiple dimensions, including size deviation normalization, aspect ratio consistency, and geometric stability, avoiding the risk of misjudgment caused by a single indicator. Simultaneously, the dual-channel collaborative decision-making architecture involves mutual verification between the main and auxiliary channels, maintaining high robustness even under complex conditions such as image distortion, tilt, or occlusion, further improving the accuracy and adaptability of paper type recognition.

[0082] Furthermore, the main channel described in this invention employs a nearest neighbor search mechanism with fuzzy tolerance, enabling stable matching of standard paper types even in the presence of measurement errors and image distortion. Simultaneously, by introducing three factors—size deviation normalization score, aspect ratio consistency score, and geometric stability score—to jointly calculate the geometric matching confidence level, a comprehensive evaluation is conducted from three dimensions: absolute size, shape proportion, and boundary quality. This reduces the risk of matching failure and misjudgment due to measurement errors or image distortion, thereby further improving the accuracy, robustness, and adaptability of paper type recognition.

[0083] Furthermore, the size deviation normalization score described in this invention reduces the equipment cost caused by reliance on hardware sensors and the risk of incorrect matching due to single-dimensional judgment by restoring physical dimensions, standard data retrieval, and deviation mapping score, thereby further improving the accuracy and adaptability of paper type matching.

[0084] Furthermore, the aspect ratio consistency score described in this invention reduces the risk of misjudgment of direction due to inconsistent horizontal and vertical placement of the paper images, as well as the risk of incorrect matching due to similar sizes but different proportions, through actual ratio calculation, theoretical ratio matching, and ratio difference scoring. This further improves the accuracy and adaptability of paper type matching.

[0085] Furthermore, the geometric stability score described in this invention, through edge straightness fitting, vertex fitting error calculation, and comprehensive stability scoring, avoids matching errors caused by problems such as tilting, distortion, local occlusion, or edge blurring of the work paper image, improves the self-evaluation ability of the paper boundary detection results and the self-diagnosis ability of low-quality acquired images, thereby further improving the system's automatic recognition ability of work paper images.

[0086] Furthermore, the geometric matching confidence score described in this invention outputs the final confidence score through three elements: dynamic weight allocation, weighted summation, nonlinear calibration and penalty, and normalization fusion. This avoids the distortion of single-dimensional scores caused by problems such as tilting, distortion, local occlusion, or edge blurring of the work paper image, which could lead to overall misjudgment. It effectively improves the adaptability of the matching confidence score assessment to complex work image scenarios and its sensitivity feedback capability to boundary detection quality, thereby further improving the accuracy and robustness of image recognition.

[0087] Furthermore, the auxiliary channel of this invention, by activating a lightweight CNN model to input the entire image or ROI region, avoids the overall system failure that occurs when the main channel cannot effectively match, further improving the fault tolerance and robustness of the paper type recognition architecture, and realizing complementary enhancement under the collaborative decision-making of the main and auxiliary channels.

[0088] Furthermore, the system described in this invention achieves end-to-end adaptive processing from image acquisition, paper type identification, trace template value, image structure mapping to final printing output, reducing operational complexity and improving the adaptive capability of trace correction; and, by integrating a rule engine and a deep learning classifier through a dual-mode recognition decision-maker, and introducing a paper knowledge graph storage module in the field of education, it dynamically calls statistical data based on the region and grade level information of the assignment, relying on the rule engine to output high-confidence results when the image is clear, and using the deep learning classifier to supplement when there is distortion, occlusion or blur, thereby further improving the accuracy and robustness of image recognition. Attached Figure Description

[0089] Figure 1 This is a flowchart of the method for intelligent paper recognition and adaptive trace printing described in this invention;

[0090] Figure 2 This is a system architecture diagram for implementing the method of the present invention;

[0091] Figure 3 This is a flowchart illustrating how the fusion coefficient of the present invention is dynamically adjusted based on the image quality of the paper, the type of work, and regional customs.

[0092] Figure 4 This is a flowchart of the paper recognition and adaptive trace printing process described in an embodiment of the present invention. Detailed Implementation

[0093] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0094] It should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0095] Please see Figure 1 The diagram shows a flowchart of the intelligent paper recognition and adaptive trace-leaking printing method of the present invention, the process of which includes:

[0096] Step S1: Acquire a digitized image;

[0097] Step S2: Perform edge detection and contour extraction on the digitized image to identify the closed quadrilateral region representing the paper boundary;

[0098] Step S3: Calculate the aspect ratio of the circumscribed rectangle of the quadrilateral region, and estimate the actual physical size by combining the image DPI information;

[0099] Step S4: The calculated size is matched with the standard paper type in the preset educational paper knowledge graph EPKG using a matching algorithm to determine the target paper type.

[0100] Step S5: Based on the identified paper type, automatically load the corresponding trace rendering template, which contains the spatial coordinate anchor points of each question;

[0101] Step S6: Map the correction result data of the intelligent correction system to the corresponding coordinate position of the trace rendering template to generate an output file with correction traces;

[0102] Step S7: Control the printing device to perform the printing operation according to the default printing configuration of the paper type.

[0103] Specifically, the matching algorithm adopts a fusion decision mode, and its confidence output uses a multi-source weighted calculation method; the matching process of the matching algorithm adopts a dual-channel collaborative decision architecture, including a main channel for geometric matching and an auxiliary channel for CNN classification.

[0104] Specifically, the process by which the main channel determines the confidence level through geometric matching includes:

[0105] The physical dimensions of the circumscribed rectangle of the quadrilateral region are determined based on edge detection;

[0106] Based on the physical dimensions, a nearest neighbor search with fuzziness tolerance is performed in the paper knowledge graph (EPKG) for the education field;

[0107] The search yielded three key elements for joint calculation: dimensional deviation normalization score, aspect ratio consistency score, and geometric stability score.

[0108] The confidence level is calculated by combining the three factors.

[0109] Specifically, the size deviation normalization score is used to quantify the closeness between the measured size of the image and the standard physical size, the aspect ratio consistency score is used to distinguish papers of similar size but different proportions, and the geometric stability score is used to evaluate the quality of the paper boundaries in the image.

[0110] Specifically, the calculation process for the normalized score of the size deviation includes:

[0111] Physical size restoration: Read the image DPI information from image metadata or default configuration, and convert the detected paper quadrilateral pixel dimensions in the image into physical measurement dimensions. The conversion formula is as follows.

[0112]

[0113]

[0114] Where Lmeas represents the physical measurement dimension length, Wmeas represents the physical measurement dimension width, Lpixel represents the pixel length, and Wpixel represents the pixel width;

[0115] Standard data retrieval: Based on the initially identified paper type candidates, retrieve the standard physical dimensions corresponding to the paper type from the Education-Specific Paper Knowledge Graph (EPKG);

[0116] Deviation mapping to score: The average value of the relative error of the length and width of the physical measurement paper quadrilateral is calculated by the following formula (1), and the error is mapped to a value between 0 and 1 by the following formula (2) exponential decay function to obtain the normalized score of the dimensional deviation:

[0117] (1)

[0118] (2)

[0119] Where Esize represents the average value of the relative error of the length and width of the quadrilateral of the paper in physical measurement, Lstd represents the standard physical dimension length corresponding to the paper type, Wstd represents the standard physical dimension length corresponding to the paper type, Ssize represents the normalized score of the size deviation, and α represents the sensitivity coefficient.

[0120] Specifically, the sensitivity coefficient α ranges from [5, 10]. In this embodiment, the sensitivity coefficient α is assigned a value of 7. It is understood that when the error exceeds 10%, the normalized score for the size deviation will decrease rapidly, thereby achieving a normalized score. This embodiment does not impose specific restrictions on the value of the sensitivity coefficient α, as long as it ensures that the normalized score for the size deviation can be obtained based on the sensitivity coefficient.

[0121] Specifically, the calculation process for the aspect ratio consistency score includes:

[0122] Calculation of measured aspect ratio: The measured aspect ratio of the quadrilateral boundary detected in the image is calculated using the following formula, and the ratio is greater than 1:

[0123]

[0124] Where Rmeas represents the aspect ratio of the measured paper quadrilateral, max(Lmeas,Wmeas) represents selecting the maximum value among the physical measurement of the paper quadrilateral's length and width, and min(Lmeas,Wmeas) represents selecting the minimum value among the physical measurement of the paper quadrilateral's length and width.

[0125] Theoretical aspect ratio matching: Obtain the theoretical aspect ratio of candidate paper types from the Education-Specific Paper Knowledge Graph (EPKG);

[0126] Proportion Difference Score: The absolute difference between the measured proportion and the theoretical proportion is calculated using the following formula (1), and the aspect ratio consistency score is generated using the following formula (2) Gaussian function:

[0127] (1)

[0128] (2)

[0129] Where Δratio represents the absolute difference between the measured ratio and the theoretical ratio, Rstd represents the theoretical aspect ratio of the paper quadrilateral, Sratio represents the aspect ratio consistency score, and β represents the discrimination coefficient.

[0130] Specifically, the calculation of the aspect ratio consistency score also requires the introduction of a series of feature judgment logic. It is understood that if a candidate belongs to series A, the theoretical ratio is fixed at [value missing]. If the candidate belongs to the K series, then the specific integer ratio or specific decimal ratio recorded in the EPKG is read. In this embodiment, there are no specific restrictions on the selection of candidate sequences, as long as the theoretical ratio matching can be achieved by introducing series feature judgment logic.

[0131] Specifically, the calculation process for the geometric stability score includes:

[0132] Edge straightness fitting: For each of the four edges of the detected paper quadrilateral, perform least squares straightness fitting, calculate the goodness-of-fit R² value for each edge, and calculate the average straightness score using the following formula:

[0133]

[0134] Among them, Score line r represents the average straightness score. i This represents the R² value of the four sides of the quadrilateral on the paper.

[0135] Vertex fitting error calculation: Compare the coordinates of the four detected corner points with the coordinates of the vertices of the ideal quadrilateral formed by the intersection of the four fitted lines, and calculate the Euclidean distance error. Let the average pixel error of the four vertices be denoted as . Set a maximum allowable error, and calculate the vertex score using the following formula:

[0136]

[0137]

[0138] Where Evertex represents the average pixel error of the four vertices, P (j) detected P represents the coordinates of the four detected corner points. (j) ideal Score represents the coordinates of the vertices of an ideal quadrilateral formed by the intersection of four fitted lines. vetex Tvertex represents the vertex score, and Tvertex represents the maximum allowable error. This indicates a function that selects the maximum value.

[0139] Stability comprehensive score: The straightness score and vertex score are weighted and combined using the following formula:

[0140] Sstab=W line •Score line +W vertex •Score vertex

[0141] Where Sstab represents the geometric stability score, W line W represents the internal weight of straightness. vertex This represents the weight inside a vertex.

[0142] Specifically, the calculation steps for the geometric matching confidence score include:

[0143] Dynamic weight allocation: Based on the image quality, job type, and regional habits of the paper, dynamic weights are assigned to the three scores; if the image clarity does not meet the standard, the proportions of the dynamic weights of the size deviation normalization score and the aspect ratio consistency score are reduced, while the sensitivity weight for edge quality in the dynamic weight of the geometric stability score is increased.

[0144] Weighted summation: The original confidence level is obtained by linearly weighting each score based on the adjusted weights, as shown in the following formula:

[0145] Craw=Wsize•Ssize+Wratio•Sratio+Wstab•Sstab

[0146] Where Craw represents the original confidence level, Wsize represents the dynamic weight of the size deviation normalization score, Wratio represents the dynamic weight of the aspect ratio consistency score, and Wstab represents the dynamic weight of the geometric stability score.

[0147] Nonlinear calibration and penalty: If the dynamic weight of the geometric stability score is lower than the threshold of the dynamic weight of the geometric stability score, it indicates that the geometric quality of the image is extremely poor. Then, a penalty coefficient is applied to the final geometric matching confidence. The formula for calculating the geometric matching confidence is as follows:

[0148]

[0149] Where Confgeo represents the geometric matching confidence, λ represents the penalty coefficient, and Tstab represents the geometric stability score threshold.

[0150] Normalized output: Ensure the final output value falls between 0 and 1, and record this value as the normalized geometric matching confidence score, calculated as follows:

[0151] Confgeo1=min(1,max(0,Confgeo))

[0152] Where Confgeo1 represents the normalized geometric matching confidence score, max(0,Confgeo) represents the maximum value between 0 and the geometric matching confidence score, and min(1,max(0,Confgeo)) represents the minimum value between 1 and max(0,Confgeo).

[0153] Specifically, the normalized geometric matching confidence score obtained through calculation is input into the dual-mode decision module and weighted and fused with the CNN classification confidence score.

[0154] Specifically, the calculation process for the auxiliary channel includes:

[0155] Start the lightweight CNN model, input the entire image or ROI region, and output the probability distribution of each paper pattern;

[0156] The geometric matching confidence score obtained through the main channel is compared with a pre-stored threshold.

[0157] If the geometric matching confidence score is greater than or equal to the threshold, the geometric matching confidence score is adopted, and the geometric matching confidence score is weighted and fused with the preset CNN classification confidence score using the following formula to obtain the comprehensive confidence score:

[0158] Final_Score=Wgeo•Confgeo1+(1−Wgeo)•Confcnn

[0159] Where Final_Score represents the overall confidence score, Wgeo represents the fusion coefficient dynamically adjusted based on the image quality, job type, and regional habits of the paper, and Confcnn represents the CNN classification confidence score;

[0160] If the geometric matching confidence is less than the threshold, the CNN channel is triggered, and the overall confidence is calculated according to the following formula:

[0161] Final_Score=a×Confgeo1+b×Confcnn+c×Context_Boost

[0162] Where a represents the geometric weight coefficient, b represents the neural network weight coefficient, c represents the EPKG boosting weight factor, and Context_Boost represents the EPKG usage frequency boosting value.

[0163] Specifically, the model is trained on a set of real work images labeled by EPKG. It learns visual patterns that are strongly related to educational scenarios, such as paper texture, binding hole position, printing border, and question number arrangement, to compensate for the failure of pure geometric methods under severe distortion or local occlusion.

[0164] Specifically, step S6 of the method introduces a dynamic coordinate mapping mechanism, which precisely projects the correction marks onto the spatial coordinate system of the original work image through perspective transformation based on the actual resolution of the input image and the scanning distortion, so that the mark position is aligned with the original question position.

[0165] Please see Figure 2 As shown, it is a system architecture framework diagram for implementing the method of the present invention, including:

[0166] The image acquisition unit is used to acquire digital images;

[0167] The paper recognition engine is connected to the image acquisition unit to integrate edge detection, geometric analysis and template matching functions;

[0168] The educational paper knowledge graph storage module is connected to the paper recognition engine and is used to store the typical worksheet specifications used in different regions, grades and textbook versions, as well as the standard size, common layout methods and question area distribution patterns of each specification.

[0169] The dual-mode recognition decision-maker is connected to the paper recognition engine and the educational paper knowledge graph storage module, and is used to identify and determine the type of homework paper by combining the paper feature extraction results and knowledge graph information. The dual-mode recognition decision-maker includes a rule engine and a deep learning classifier.

[0170] The trace rendering template library is connected to the dual-mode recognition decision unit and is used to organize the rendering modules of various types of correction traces according to paper type, and to retrieve the matching trace template according to the paper recognition result.

[0171] The dynamic mapping and printing control unit is connected to the dual-mode recognition decision unit and the trace rendering template library to generate documents with correction traces and drive the printer to automatically execute output.

[0172] The closed-loop feedback module, which is connected to the paper recognition engine and the dynamic mapping and printing control unit, is used to collect post-printing trace effect evaluation data, including user correction records or automatic comparison of image deviations before and after printing; the closed-loop feedback module is also used to iteratively optimize the parameter thresholds and template library accuracy of the paper recognition algorithm.

[0173] Specifically, the image acquisition unit of the present invention is an intelligent shooting terminal or high-speed scanning device that supports automatic correction and illumination compensation, so as to provide high-quality image input that can still be used for recognition even in non-ideal environments.

[0174] Please see Figure 3 The diagram shows a flowchart illustrating the dynamic adjustment of the fusion coefficient according to the image quality, job type, and regional customs of the paper, as described in this invention. The specific process includes:

[0175] The sharpness of the acquired digital images is assessed.

[0176] The high-frequency component energy of the image is calculated using the Laplacian variance algorithm as follows, and a sharpness score is generated:

[0177]

[0178] Where Score_clear represents the sharpness score, N represents the total number of pixels in the Laplacian response graph, and L... i represents the pixel values ​​of the Laplacian response map, and L is the mean value of the response map;

[0179] Based on the sharpness score generated above, the quality correction factor is calculated using the following formula:

[0180]

[0181] Where Kquality represents the quality correction factor, and T_ λ represents the preset sharpness threshold. q The attenuation factor is less than 1;

[0182] Specifically, when the sharpness score is less than the sharpness threshold, the image sharpness is determined to be non-compliant with the standard, and the final weight of the geometric matching confidence score is reduced, thereby increasing the final weight of the CNN classification confidence score.

[0183] Specifically, the attenuation factor λ q The value range of λ is (0, 1). In this embodiment, the attenuation factor λ qThe value is assigned to 0.8. It is understood that those skilled in the art can adjust the attenuation factor λ according to the specific implementation. q Assigning values ​​is done as long as the assigned attenuation factor can correct the sharpness weight.

[0184] The lightweight texture analysis module or OCR preprocessing module is invoked to statistically analyze the ink density and the proportion of unstructured strokes in the image, generate the handwritten density index, and calculate the content correction coefficient using the following formula:

[0185] Kcontent=1-m•Index_handwriting

[0186] Where Kcontent represents the content correction factor, m represents the handwriting sensitivity factor, and Index_handwriting represents the handwriting density index;

[0187] Specifically, when the handwriting density index exceeds a preset handwriting threshold, handwriting interference is deemed severe. In this case, the generated content correction coefficient is negatively correlated with the handwriting density index, thereby improving the final weight of the CNN classification confidence.

[0188] Specifically, the handwriting sensitivity coefficient m ranges from (0,1). In this embodiment, the handwriting sensitivity coefficient m is assigned a value of 0.5. When the handwriting density index is higher, the final weight of the geometric matching confidence score is lower, and the final weight of the CNN classification confidence score is higher. It is understood that those skilled in the art can assign values ​​to the handwriting sensitivity coefficient m according to the specific implementation, as long as the handwriting sensitivity coefficient can effectively adjust the original weight of the geometric matching confidence score.

[0189] Based on user login information or image metadata, obtain the region identifier and query the knowledge graph of paper in the education field;

[0190] Retrieve the historical usage frequency of the paper type in this region that most closely matches the geometric matching result for the current educational level, and generate the prior correction coefficient using the following formula:

[0191]

[0192] Where Kprior represents the prior correction coefficient, λ p This represents the prior enhancement coefficient, Pregion represents the historical usage frequency, and Tprior represents the preset prior probability threshold.

[0193] Specifically, when the historical usage frequency is greater than the prior probability threshold, and the geometrically matched paper type is consistent with the high-frequency paper type, the prior correction coefficient is generated.

[0194] Specifically, the prior enhancement coefficient λ p The value range of λ is (1, +∞), and it is used as the final weight of the geometric matching confidence level. In this embodiment, the prior enhancement coefficient λ p The value is assigned to 1.2. It is understandable that when a region has a very high usage habit for a specific paper type, if the geometric measurement results happen to match this habit, it indicates that the reliability of the geometric measurement has been verified by regional big data. Therefore, the weight of the geometric matching confidence level should be increased to reduce misjudgments caused by random errors. It is understood that those skilled in the art can assign values ​​to the prior enhancement coefficient according to the specific implementation, as long as the prior enhancement coefficient can complete the adjustment of the original weight of the geometric matching confidence level.

[0195] The original weights of the geometric matching confidence scores are calculated using the following formula:

[0196] Wgraw=Wgbase•Kquality•Kprior•Kcontent

[0197] Wherein, Wgraw represents the original weight of the geometric matching confidence score, and Wgbase represents the initial weight of the geometric matching confidence score;

[0198] The original weights of the geometric matching confidence score are normalized using the following formula, and denoted as the final weights of the geometric matching confidence score:

[0199] Wgfinal=max(Wmin,min(Wmax,Wgraw))

[0200] Where Wgfinal represents the final weight of the geometric matching confidence score, Wmin represents the minimum value of the original weight of the preset geometric matching confidence score, and Wmax represents the maximum value of the original weight of the preset geometric matching confidence score.

[0201] The final weights of the CNN classification confidence are calculated using the following formula:

[0202] Wcfinal=1-Wgfinal

[0203] Where Wcfinal represents the final weight of the CNN classification confidence score;

[0204] Specifically, the minimum value Wmin and the maximum value Wmax of the preset geometric matching confidence weights are within the range of (0,1) and Wmin ≤ Wmax. In this embodiment, the minimum value Wmin is assigned a value of 0.2, and the maximum value Wmax is assigned a value of 0.8. It can be understood that by assigning values ​​to the minimum and maximum values ​​Wmin and Wmax of the preset geometric matching confidence weights, the final weights of the geometric matching confidence and the final weights of the CNN classification confidence are calculated and normalized to ensure that their sum is strictly equal to 1, and that individual weight values ​​are limited to the preset minimum and maximum values, preventing excessively large weights in one channel from causing decision imbalance. It is understood that those skilled in the art can assign values ​​to the minimum and maximum values ​​of the preset geometric matching confidence original weights according to the specific implementation, as long as the minimum and maximum values ​​of the preset geometric matching confidence original weights can complete the normalization process of the final weights of the geometric matching confidence and the final weights of the CNN classification confidence.

[0205] Using the dynamic weights obtained above, the confidence scores of the geometric matching channel output and the CNN classification channel output are weighted and summed, and the final comprehensive confidence score is calculated using the following formula:

[0206] Final_Score=(Wgfinal×Confgeo1)+(Wcfinal×Confcnn)

[0207] The overall confidence level is compared with a preset global decision threshold;

[0208] If the overall confidence level is greater than or equal to the global decision threshold, the identification result is confirmed and the process of leaving a record is initiated.

[0209] If the overall confidence level is less than the global decision threshold, the closed-loop feedback module is triggered.

[0210] Specifically, when the overall confidence level is less than the global decision threshold, the closed-loop feedback module is triggered, and the sample is marked as a low-confidence sample, transferred to the manual verification process, and the verified data is fed back to the training set to optimize the tolerance parameters of the CNN model and geometric matching.

[0211] Specifically, in this embodiment, the global decision threshold is set to 0.7. It is understood that those skilled in the art can assign a value to the global decision threshold according to the specific implementation, as long as the global decision threshold can determine whether the overall confidence level meets the identification and confirmation conditions.

[0212] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0213] Please see Figure 4 The flowchart shown is a process for homework paper recognition and adaptive trace printing in an embodiment of the present invention. The scenario selected in this embodiment is: a primary school using an intelligent homework grading system to process math workbooks submitted by students.

[0214] All 50 assignments from the class were scanned in batches using a high-speed scanner. The assignment types included A4 standard practice paper and 16K workbook pages.

[0215] The scanner generates an image for each page of work, and the system automatically starts the recognition process.

[0216] The system uses image processing technology to detect the paper edges, calculates the aspect ratio, and estimates the actual size based on the resolution.

[0217] Compare the actual size with the built-in "Educational Paper Knowledge Graph". If it is close to 21cm × 29.7cm, it is determined to be A4; if it is close to 19.5cm × 27cm, it is determined to be 16K.

[0218] For images whose clarity does not meet the standards, a lightweight AI model is invoked to assist in the judgment;

[0219] Once the paper type is confirmed, the system immediately loads the corresponding trace template. A3-column layout template is used for A4 paper, and 2-column template is used for 16K paper.

[0220] The correction results are precisely positioned according to the template, a PDF is generated, and sent to the printer.

[0221] The printer automatically outputs according to the recognized paper type, and the correction marks for each assignment are accurately placed next to the original question.

[0222] Specifically, in this embodiment, a three-column layout template is used for A4 paper, and a two-column template is used for 16K paper. It is understood that this embodiment of the invention does not impose specific restrictions on the selection of paper type templates, as long as the system can accurately locate the paper type using the template.

[0223] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0224] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent identification and adaptive trace-keeping printing of printing paper, characterized in that, include: S1, acquire digital image; S2, perform edge detection and contour extraction on the digital image to identify the closed quadrilateral region representing the paper boundary; S3, calculate the aspect ratio of the circumscribed rectangle of the quadrilateral region, and estimate the actual physical size by combining the image DPI information; S4. The calculated size is matched with the standard paper type in the preset educational paper knowledge graph EPKG using a matching algorithm to determine the target paper type. S5. Based on the identified paper type, automatically load the corresponding trace rendering template, which contains the spatial coordinate anchor points of each question. S6, map the correction result data of the intelligent correction system to the corresponding coordinate position of the trace rendering template, and generate an output file with correction traces; S7, control the printing device to perform a printing operation according to the default printing configuration of the paper type.

2. The method for intelligent identification and adaptive trace-keeping printing of work paper according to claim 1, characterized in that, The matching algorithm adopts a fusion decision mode, and its confidence output adopts a multi-source weighted calculation method. The matching process of the matching algorithm adopts a dual-channel collaborative decision architecture, including a main channel for geometric matching and an auxiliary channel for CNN classification.

3. The method for intelligent identification and adaptive trace-keeping printing of work paper according to claim 2, characterized in that, The process by which the main channel determines the confidence level through geometric matching includes: The physical dimensions of the circumscribed rectangle of the quadrilateral region are determined based on edge detection; Based on the physical dimensions, a nearest neighbor search with fuzziness tolerance is performed in the paper knowledge graph (EPKG) for the education field; The search yielded three key elements for joint calculation: dimensional deviation normalization score, aspect ratio consistency score, and geometric stability score. The confidence level is calculated by combining the three factors.

4. The method for intelligent identification and adaptive trace-keeping printing of work paper according to claim 3, characterized in that, The calculation process for the normalized score of the dimensional deviation includes: Physical size restoration: Read the image DPI information from image metadata or default configuration, and convert the pixel dimensions of the detected paper quadrilaterals in the image into physical measured dimensions. The conversion formula is as follows: Where Lmeas represents the physical measurement dimension length, Wmeas represents the physical measurement dimension width, Lpixel represents the pixel length, and Wpixel represents the pixel width; Standard data retrieval: Based on the initially identified paper type candidates, retrieve the standard physical dimensions corresponding to the paper type from the Education-Specific Paper Knowledge Graph (EPKG); Deviation mapping to score: The average value of the relative error of the length and width of the physical measurement paper quadrilateral is calculated by the following formula (1), and the error is mapped to a value between 0 and 1 by the following formula (2) exponential decay function to obtain the normalized score of the dimensional deviation: （1） （2） Where Esize represents the average value of the relative error of the length and width of the quadrilateral of the paper in physical measurement, Lstd represents the standard physical dimension length corresponding to the paper type, Wstd represents the standard physical dimension length corresponding to the paper type, Ssize represents the normalized score of the size deviation, and α represents the sensitivity coefficient.

5. The method for intelligent identification and adaptive trace-keeping printing of work paper according to claim 3, characterized in that, The calculation process for the aspect ratio consistency score includes: Calculation of measured aspect ratio: The measured aspect ratio of the quadrilateral boundary detected in the image is calculated using the following formula, and the ratio is greater than 1: Where Rmeas represents the aspect ratio of the measured paper quadrilateral, max(Lmeas,Wmeas) represents selecting the maximum value among the physical measurement of the paper quadrilateral's length and width, and min(Lmeas,Wmeas) represents selecting the minimum value among the physical measurement of the paper quadrilateral's length and width. Theoretical aspect ratio matching: Obtain the theoretical aspect ratio of candidate paper types from the Education-Specific Paper Knowledge Graph (EPKG); Proportion Difference Score: The absolute difference between the measured proportion and the theoretical proportion is calculated using the following formula (1), and the aspect ratio consistency score is generated using the following formula (2) Gaussian function: （1） （2） Where Δratio represents the absolute difference between the measured ratio and the theoretical ratio, Rstd represents the theoretical aspect ratio of the paper quadrilateral, Sratio represents the aspect ratio consistency score, and β represents the discrimination coefficient.

6. The method for intelligent identification and adaptive trace-keeping printing of work paper according to claim 3, characterized in that, The calculation process for the geometric stability score includes: Edge straightness fitting: For each of the four edges of the detected paper quadrilateral, perform least squares straightness fitting, calculate the goodness-of-fit R² value for each edge, and calculate the average straightness score using the following formula: where Score line represents the average straightness score, r i represents the value of the fourth side R2 of the quadrangle of the paper Vertex fitting error calculation: Compare the coordinates of the four detected corner points with the coordinates of the vertices of the ideal quadrilateral formed by the intersection of the four fitted lines, and calculate the Euclidean distance error; denoted as the average pixel error of the four vertices; set the maximum allowable error, and calculate the vertex score using the following formula: Where Evertex represents the average pixel error of the four vertices, P (j) detected P represents the coordinates of the four detected corner points. (j) ideal Score represents the coordinates of the vertices of an ideal quadrilateral formed by the intersection of four fitted lines. vetex Tvertex represents the vertex score, and Tvertex represents the maximum allowable error. This indicates a function that selects the maximum value. Stability comprehensive score: The straightness score and vertex score are weighted and combined using the following formula: Sstab = W line • Score line + W vertex • Score vertex Where Sstab represents the geometric stability score, W line W represents the internal weight of straightness. vertex This represents the weight inside a vertex.

7. The method for intelligent identification and adaptive trace-keeping printing of work paper according to any one of claims 3-6, characterized in that, The steps for calculating the geometric matching confidence score include: Dynamic weight allocation: Based on the image quality, job type, and regional habits of the paper, dynamic weights are assigned to the three scores; if the image clarity does not meet the standard, the proportions of the dynamic weights of the size deviation normalization score and the aspect ratio consistency score are reduced, while the sensitivity weight for edge quality in the dynamic weight of the geometric stability score is increased. Weighted summation: The original confidence level is obtained by linearly weighting each score based on the adjusted weights, as shown in the following formula: Craw=Wsize•Ssize+Wratio•Sratio+Wstab•Sstab Wherein, Craw represents the original confidence level, Wsize represents the dynamic weight of the size deviation normalization score, Wratio represents the dynamic weight of the aspect ratio consistency score, and Wstab represents the dynamic weight of the geometric stability score. Nonlinear calibration and penalty: If the dynamic weight of the geometric stability score is lower than the threshold of the dynamic weight of the geometric stability score, it indicates that the geometric quality of the image is extremely poor. Then, a penalty coefficient is applied to the final geometric matching confidence. The formula for calculating the geometric matching confidence is as follows: Where Confgeo represents the geometric matching confidence, λ represents the penalty coefficient, and Tstab represents the geometric stability score threshold; Normalized output: Ensure the final output value falls between 0 and 1, and record this value as the normalized geometric matching confidence score, calculated as follows: Confgeo1=min(1,max(0,Confgeo)) Where Confgeo1 represents the normalized geometric matching confidence score, max(0,Confgeo) represents the maximum value between 0 and the geometric matching confidence score, and min(1,max(0,Confgeo)) represents the minimum value between 1 and max(0,Confgeo).

8. The method for intelligent identification and adaptive trace-keeping printing of work paper according to claim 2, characterized in that, The calculation process for the auxiliary channel includes: Start the lightweight CNN model, input the entire image or ROI region, and output the probability distribution of each paper pattern; The geometric matching confidence score obtained through the main channel is compared with a pre-stored threshold. If the geometric matching confidence score is greater than or equal to the threshold, the geometric matching confidence score is adopted, and the geometric matching confidence score is weighted and fused with the preset CNN classification confidence score using the following formula to obtain the comprehensive confidence score: Final_Score=Wgeo•Confgeo1+(1−Wgeo)•Confcnn Where Final_Score represents the overall confidence score, Wgeo represents the fusion coefficient dynamically adjusted based on the image quality, job type, and regional habits of the paper, and Confcnn represents the CNN classification confidence score; If the geometric matching confidence is less than the threshold, the CNN channel is triggered, and the overall confidence is calculated according to the following formula: Final_Score=a×Confgeo1+b×Confcnn+c×Context_Boost Where a represents the geometric weight coefficient, b represents the neural network weight coefficient, c represents the EPKG boosting weight factor, and Context_Boost represents the EPKG usage frequency boosting value.

9. The method for intelligent identification and adaptive trace-keeping printing of work paper according to claim 1, characterized in that, Step 6 of the method introduces a dynamic coordinate mapping mechanism, which precisely projects the correction marks onto the spatial coordinate system of the original work image through perspective transformation based on the actual resolution of the input image and the scanning distortion, so that the mark position is aligned with the original question position.

10. A paper intelligent recognition and adaptive traceability printing system using the method of any one of claims 1-9, characterized in that, include: An image acquisition unit is used to acquire digital images; A paper recognition engine, which is connected to the image acquisition unit, is used to integrate edge detection, geometric analysis and template matching functions; The educational paper knowledge graph storage module is connected to the paper recognition engine and is used to store the typical worksheet specifications used in different regions, grades and textbook versions, as well as the standard size, common layout methods and question area distribution patterns of each specification. A dual-mode recognition decision-maker is connected to the paper recognition engine and the educational paper knowledge graph storage module. It is used to combine the paper feature extraction results and knowledge graph information to identify and determine the type of homework paper. The dual-mode recognition decision-maker includes a rule engine and a deep learning classifier. The trace rendering template library is connected to the dual-mode recognition decision unit and is used to organize the rendering modules of various types of correction traces according to paper type, and to retrieve the matching trace template according to the paper recognition result. The dynamic mapping and printing control unit is connected to the dual-mode recognition decision unit and the trace rendering template library to generate documents with correction marks and drive the printer to automatically execute the output. The closed-loop feedback module, which is connected to the paper recognition engine and the dynamic mapping and printing control unit, is used to collect post-printing trace effect evaluation data, including user correction records or automatic comparison of image deviations before and after printing; the closed-loop feedback module is also used to iteratively optimize the parameter thresholds and template library accuracy of the paper recognition algorithm.

Images