An intelligent teaching auxiliary system and method based on a large model

Abstract

This invention discloses an intelligent teaching assistance system and method based on a large model. The method includes: acquiring the original optical image data of assignments to be graded on an assignment placement platform, and performing optical path distortion correction on the original optical image data to generate optical image data; performing a first-level image preprocessing on the optical image data to generate preprocessed image data; performing low-complexity semantic extraction and high-complexity grading reasoning on the preprocessed image data to generate grading results; overlaying the grading results onto the corresponding answer area of ​​the optical image data in the form of simulated handwriting, and driving an output device to generate visual grading marks; collecting and storing the grading result data to construct a multi-dimensional learning profile; and generating a personalized set of practice questions based on the multi-dimensional learning profile. This invention has the advantages of achieving automated assignment grading, improving processing efficiency, automatically adapting to optical distortion, optimizing resource utilization, and providing instant feedback and personalized learning support.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence education technology, specifically to an intelligent teaching assistance system and method based on a large model. Background Technology

[0002] Traditional homework grading relies heavily on manual teacher intervention, often resulting in inefficiency, subjective judgment bias, and excessively long feedback cycles. While artificial intelligence (AI) technology has led to the emergence of large-scale model-based intelligent grading systems in education, existing solutions face significant limitations in practical applications.

[0003] Image acquisition often relies on pre-made answer sheets or special paper, significantly increasing costs and deployment complexity. For ordinary paper-based assignments, issues such as paper wrinkles, uneven ambient lighting, and edge distortion from wide-angle lenses severely reduce the accuracy of optical character recognition, making it difficult to effectively handle handwritten content in real teaching scenarios. In the core grading stage, existing solutions typically use a single large language model, which is prone to logical errors and biases, while also consuming excessive computational resources, hindering widespread application in conventional teaching environments. Grading results are disconnected from subsequent learning analysis, failing to transform real-time grading data into dynamic and actionable learning profiles. Personalized practice recommendation systems are also disconnected from grading functions, making it difficult to form an effective teaching feedback loop. Regarding network conditions, pure cloud architectures experience significant response delays when bandwidth is limited or the network is unstable, while pure edge computing solutions are limited by local computing power and lack intelligent mechanisms to dynamically adjust task allocation based on network conditions, resulting in insufficient adaptability of the system to different teaching environments.

[0004] Therefore, the industry urgently needs an intelligent teaching support system that can break free from special hardware dependence, automatically adapt to optical distortion, achieve efficient collaboration between the edge and the cloud, and possess self-questioning capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent teaching assistance system and method based on a large model, which has the advantages of realizing automated homework grading, improving processing efficiency, automatically adapting to optical distortion, optimizing resource utilization, and providing instant feedback and personalized learning support.

[0006] In a first aspect, the present invention provides an intelligent teaching assistance system based on a large model, comprising: The space optical imaging and optical path compensation unit is used to acquire the original optical image data of the work to be corrected on the work placement platform, and to perform optical path distortion correction on the original optical image data to generate optical image data. The edge intelligent processing unit is communicatively connected to the spatial optical imaging and optical path compensation unit, and is used to perform first-level image preprocessing on the optical image data to generate preprocessed image data. The cloud-based intelligent processing unit is communicatively connected to the edge intelligent processing unit and is used to perform low-complexity semantic extraction and high-complexity correction reasoning on the preprocessed image data to generate correction results. The interactive feedback unit is communicatively connected to the cloud-based intelligent processing unit and is used to overlay the correction results onto the corresponding answer area of ​​the optical image data in the form of simulated handwriting, and drive the output device to generate visual correction marks. The learning analysis unit is connected to the cloud-based intelligent processing unit to collect and store grading result data and construct a multi-dimensional learning profile. The personalized practice generation unit is communicatively connected to the learning analysis unit and is used to generate personalized practice question sets based on the multi-dimensional learning profile.

[0007] Secondly, the present invention also provides an intelligent teaching assistance method based on a large model, applied to the above-mentioned system, comprising the following steps: S1. Obtain the original optical image data of the job to be corrected on the job placement platform, and perform optical path distortion correction on the original optical image data to generate optical image data; S2. Perform first-level image preprocessing on the optical image data to generate preprocessed image data; S3. Perform low-complexity semantic extraction and high-complexity correction reasoning on the preprocessed image data to generate correction results; S4. The correction result is superimposed onto the corresponding answer area of ​​the optical image data in the form of simulated handwriting, and the output device is driven to generate a visual correction mark. S5. Collect and store grading results data to build a multi-dimensional learning profile; S6. Generate a personalized set of practice questions based on the multi-dimensional learning profile.

[0008] As can be seen from the above, the intelligent teaching assistance system and method based on a large model provided in this application solves the problems of low efficiency, reliance on special hardware, optical distortion interference, and lack of dynamic collaboration in traditional manual grading by acquiring and correcting homework image data, performing multi-level intelligent processing, generating grading results and constructing learning profiles. It has the advantages of realizing automated homework grading, improving processing efficiency, automatically adapting to optical distortion, optimizing resource utilization, and providing instant feedback and personalized learning support. Attached Figure Description

[0009] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of the structure of an intelligent teaching assistance system based on a large model, according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating an intelligent teaching assistance method based on a large model, according to an embodiment of the present invention. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] Traditional homework grading methods rely on manual labor, resulting in limitations such as low efficiency, high subjectivity, and delayed feedback. Existing intelligent grading systems often depend on specific paper types or equipment for image acquisition, limiting recognition accuracy. Furthermore, single-model grading lacks reliability, consumes significant computational resources, and often operates independently of grading and student learning analysis, making it difficult to form a closed-loop teaching system. In addition, the edge and cloud collaborative processing mechanisms are still underdeveloped, affecting the system's deployment and operational efficiency in different network environments.

[0013] In this regard, such as Figure 1 As shown, this application proposes an intelligent teaching assistance system based on a large model, comprising: The space optical imaging and optical path compensation unit is used to acquire the original optical image data of the work to be corrected on the work placement platform, and to perform optical path distortion correction on the original optical image data to generate optical image data. The edge intelligent processing unit is communicatively connected to the spatial optical imaging and optical path compensation unit, and is used to perform first-level image preprocessing on optical image data to generate preprocessed image data. The cloud-based intelligent processing unit communicates with the edge intelligent processing unit to perform low-complexity semantic extraction and high-complexity correction reasoning on preprocessed image data, and generate correction results. The interactive feedback unit communicates with the cloud-based intelligent processing unit to overlay the correction results onto the corresponding answer area of ​​the optical image data in the form of simulated handwriting, and drives the output device to generate visual correction marks. The learning analysis unit communicates and connects with the cloud-based intelligent processing unit to collect and store grading result data and construct a multi-dimensional learning profile. The personalized practice generation unit communicates with the learning analysis unit and is used to generate personalized practice question sets based on multi-dimensional learning profiles.

[0014] The space optical imaging and optical path compensation unit is responsible for acquiring operational images through optical means and correcting any distortions that may exist in the images to ensure the image quality for subsequent processing. This unit typically includes an optical lens, an image sensor, and components for correcting optical path deviations.

[0015] The function of an edge intelligence processing unit is to perform preliminary intelligent processing tasks at the device end, close to the data source. This unit typically integrates computing chips and storage modules, enabling it to preprocess image data, reduce the computational burden on cloud servers, and improve response speed.

[0016] The role of the cloud-based intelligent processing unit is to execute complex intelligent processing tasks on a remote server cluster. This unit typically possesses powerful computing capabilities and storage resources, enabling advanced functions such as deep semantic understanding and complex reasoning.

[0017] The interactive feedback unit's function is to present the system-generated correction results in a user-perceptible manner. This unit can drive the printing device to generate correction marks on paper work or display correction information on an electronic screen.

[0018] The purpose of the learning analysis unit is to conduct data-driven analysis and modeling of students' learning progress. This unit assesses students' mastery of knowledge points and identifies weaknesses in their learning by processing grading data.

[0019] The personalized practice generation unit automatically generates customized practice questions based on students' learning characteristics and weaknesses. This unit aims to provide targeted learning resources to help students consolidate their knowledge and improve their skills.

[0020] Raw optical image data refers to image information directly acquired by optical imaging equipment without any processing, and may contain original defects such as uneven illumination and geometric distortion.

[0021] Optical image data refers to image data that has undergone optical path distortion correction. Its geometric distortion and optical defects have been effectively corrected, providing high-quality input for subsequent image processing.

[0022] Preprocessed image data refers to image data that has undergone first-level image preprocessing by the edge intelligent processing unit, such as noise reduction, enhancement, and segmentation, so that semantic extraction and critical reasoning can be performed in the cloud.

[0023] The grading results refer to the feedback information generated by the system after evaluating the student's homework, including but not limited to correct and incorrect markings, scores, comments, and error type analysis.

[0024] A multi-dimensional learning profile refers to a data model constructed through learning analysis units that comprehensively reflects students' learning status and ability characteristics, covering multiple aspects such as knowledge mastery, error type preferences, and learning trajectory.

[0025] Personalized practice question sets refer to targeted practice question sets that are automatically matched or generated by the system based on the specific learning profile of students.

[0026] This embodiment provides an intelligent teaching assistance system. The spatial optical imaging and optical path compensation unit acquires the original optical image data of the assignments to be graded on the assignment placement platform and performs optical path distortion correction on the original optical image data to generate optical image data. Specifically, the spatial optical imaging and optical path compensation unit can use a standard fixed-focus lens and an image sensor to acquire the original optical image data on the assignment placement platform. For example, a regular camera can be fixed above the assignment to capture images. The acquired original optical image data may contain geometric distortions caused by the lens itself or the shooting angle. To address this, preset distortion parameters can be used for fixed correction, such as using software algorithms to perform geometric transformations on the image, remapping the image pixels to the correct positions, thereby generating optical image data.

[0027] Furthermore, the edge intelligent processing unit is communicatively connected to the spatial optical imaging and optical path compensation unit to perform first-level image preprocessing on the optical image data, generating preprocessed image data. This edge intelligent processing unit can be an embedded computing module that transmits data with the spatial optical imaging and optical path compensation unit via wired or wireless means. After receiving the optical image data, this unit can perform a series of image preprocessing operations. For example, mean filtering or Gaussian filtering can be used to reduce noise in the image. Subsequently, simple edge detection algorithms, such as the Sobel operator or the Canny operator, can be used to extract handwriting edge information from the image. Further, a fixed threshold method or a simple segmentation method based on pixel grayscale histograms can be used to separate the answer area from the background in the image, thereby generating preprocessed image data.

[0028] Furthermore, the cloud-based intelligent processing unit communicates with the edge intelligent processing unit to perform low-complexity semantic extraction and high-complexity grading reasoning on preprocessed image data, generating grading results. This cloud-based intelligent processing unit can be a remote server cluster that communicates with the edge intelligent processing unit via the internet. After receiving the preprocessed image data, this unit can use a pre-trained text recognition model (such as an OCR engine) to convert the text content in the image into processable text. For the recognized text, preliminary semantic analysis can be performed using rule-based matching or keyword extraction to determine the correctness of the answer. For questions requiring deeper understanding, such as open-ended questions, a large language model can be invoked for reasoning, generating detailed grading results, including scores and comments, based on preset scoring criteria and reference answers.

[0029] Therefore, the interactive feedback unit communicates with the cloud-based intelligent processing unit to overlay the correction results onto the corresponding answer area of ​​the optical image data in the form of simulated handwriting, and drives the output device to generate visual correction marks. This interactive feedback unit can be a graphics rendering module that receives the correction results sent by the cloud-based intelligent processing unit via a network. This module can render the correction results (such as "√", "×" symbols, scores, or brief comments) into images using preset fonts and colors. Subsequently, these rendered correction images are overlaid onto the corresponding answer area of ​​the original optical image data. For example, a simple weighted average of the pixel values ​​of the correction symbols and the original image can be used to achieve the overlay effect. The overlaid image data can be sent to an output device, such as a regular inkjet printer, which prints a copy with correction marks on a paper document.

[0030] Simultaneously, the learning analysis unit communicates with the cloud-based intelligent processing unit to collect and store grading result data, constructing a multi-dimensional learning profile. This learning analysis unit can be a data analysis server, interacting with the cloud-based intelligent processing unit. This unit receives and stores grading result data for each assignment, such as student scores and answer details. Based on this data, statistical analysis methods can be used, such as calculating the average score rate of students on different knowledge points, or counting the number of errors students make on specific question types. Through these statistical indicators, a preliminary learning profile of the student can be constructed, for example, identifying knowledge points where students score low or question types they frequently make mistakes on.

[0031] Finally, the personalized practice generation unit communicates with the learning analysis unit to generate personalized practice question sets based on multi-dimensional learning profiles. This personalized practice generation unit can be a question bank management system that connects to the learning analysis unit. This unit receives learning profile information provided by the learning analysis unit, such as a student's poor performance on a particular knowledge point. Based on this information, the unit can retrieve practice questions related to that weak knowledge point from a pre-set question bank through simple keyword matching or tag filtering. For example, if the learning profile shows that a student has problems with "score calculation," the system will filter all questions marked "score calculation" from the question bank and combine them into a personalized practice question set.

[0032] Based on the above technical solutions, this system effectively solves the problems of low efficiency and high subjectivity in traditional homework grading. Through spatial optical imaging and optical path compensation, the system can directly process ordinary paper-based homework, overcome image distortion, and improve recognition accuracy. The edge and cloud collaborative processing mechanism balances real-time performance and accuracy, reducing resource consumption. Grading results are superimposed in the form of simulated handwriting, which conforms to teachers' grading habits. At the same time, the system can construct multi-dimensional student learning profiles and generate personalized exercises, forming a teaching closed loop and improving teaching effectiveness.

[0033] In one optional implementation, the spatial optical imaging and optical path compensation unit includes a wide-angle optical lens, an optical path compensation component, and an ambient light sensor. The wide-angle optical lens is positioned above the work platform to capture images over a large field of view, capable of covering the entire work to be corrected on the platform in a single pass. It is characterized by a short focal length and a wide angle of view, but typically comes with some optical distortion, especially edge distortion. The optical path compensation component is optically coupled to the wide-angle optical lens and is equipped with at least one compensation lens. The optical path compensation component is a device used to correct light path deviations in an optical system; its core function is to offset or reduce image distortion by altering the characteristics of optical elements in the optical path. The compensation lens is a key optical element in the optical path compensation component, and its position and curvature are adjustable. By precisely adjusting these parameters, the refraction path of light can be changed, thereby actively correcting image distortion. The ambient light sensor is electrically connected to the wide-angle optical lens to detect the incident angle and intensity of ambient light around the work platform in real time. This is crucial for understanding and predicting the type and degree of optical distortion, as lighting conditions affect image quality and distortion performance.

[0034] Furthermore, the optical path compensation component is electrically connected to the ambient light sensor to dynamically adjust the position and curvature of the compensation lens based on the incident light angle and intensity collected by the ambient light sensor, thereby correcting edge distortions in the original optical image data. Dynamically adjusting the position and curvature of the compensation lens is an active distortion correction mechanism. Based on real-time data collected by the ambient light sensor, the system can calculate the type and degree of distortion that may occur under the current lighting conditions and accordingly drive the compensation lens to make physical adjustments to achieve the best optical compensation effect. Correcting edge distortions in the original optical image data refers to eliminating or significantly reducing geometric distortions, such as barrel or pincushion distortion, in the edge regions of the original optical image data through the aforementioned dynamic adjustment mechanism, ensuring the geometric accuracy of the image and providing high-quality input for subsequent image processing.

[0035] Through the above technical solutions, this system can acquire images of the entire work area at once, improving efficiency. Simultaneously, addressing the inherent edge distortion problem of wide-angle lenses, this system uses a compensation lens in the optical path compensation component for optical correction. More importantly, the ambient light sensor detects changes in ambient light in real time and drives the optical path compensation component to dynamically adjust the position and curvature of the compensation lens. This proactive, real-time distortion correction mechanism effectively copes with complex and changing lighting environments, ensuring that geometrically accurate raw optical image data with significantly reduced edge distortion is obtained under different conditions. This provides high-quality input for subsequent image preprocessing and grading inference, significantly improving the robustness and accuracy of the entire intelligent teaching assistance system in the image acquisition stage.

[0036] In one optional implementation, the specific steps of the first-level image preprocessing performed by the edge intelligent processing unit include: firstly, performing wavelet transform decomposition on the optical image data to separate low-frequency approximation components and high-frequency detail components. Wavelet transform, as a multi-resolution analysis tool, can decompose image signals into sub-bands of different frequencies, where the low-frequency approximation components represent the overall structure and main information of the image, while the high-frequency detail components contain detailed information such as image edges, textures, and noise. Through this decomposition, different frequency components can be processed selectively.

[0037] Subsequently, an adaptive thresholding denoising algorithm is applied to the high-frequency detail components to remove paper texture and optical noise. Since the high-frequency components contain a large amount of noise and unnecessary texture information, the adaptive thresholding denoising algorithm can dynamically adjust the denoising intensity according to local image characteristics, thereby effectively suppressing noise while preserving the handwriting details to the maximum extent. For example, thresholding methods based on the statistical characteristics of wavelet coefficients, such as VisuShrink or SureShrink, can be used. These methods can adaptively determine the optimal threshold based on the distribution characteristics of wavelet coefficients, achieving refined denoising.

[0038] After denoising, the denoised high-frequency detail components and low-frequency approximation components are reconstructed using inverse wavelet transform to generate the denoised image. This step re-synthesizes the processed image components into a complete image. At this point, most of the noise and paper texture in the image have been effectively removed, laying the foundation for subsequent handwriting analysis.

[0039] Building upon this, handwriting edge enhancement based on a deep convolutional neural network is performed on the denoised image. The handwriting edge enhancement employs a residual dense block structure and fuses shallow edge features and deep semantic features through skip connections, ultimately generating an image with enhanced handwriting edges. Deep convolutional neural networks possess powerful feature learning capabilities, enabling the extraction of complex handwriting features from images. The residual dense block structure effectively alleviates the gradient vanishing problem in deep network training by allowing information to flow more freely between network layers and promotes feature reuse. Skip connections directly transfer features from shallow layers (containing rich edge details) to deeper layers (containing more abstract semantic information), thereby enhancing handwriting edges while maintaining the integrity and coherence of the handwriting, resulting in clearer and sharper handwriting edges and providing high-quality input for subsequent recognition.

[0040] Finally, the image after handwriting edge enhancement is automatically segmented into answer regions based on connected component analysis, and the segmented answer region image blocks are output as preprocessed image data. Connected component analysis is an effective image processing technique that can identify regions in an image with the same pixel value and that are interconnected. By binarizing the enhanced image and applying connected component analysis, the individual regions of the student's answer can be accurately identified. By setting reasonable filtering conditions such as area and shape, the answer regions can be accurately separated from the background, generating independent image blocks, ensuring that subsequent processing focuses only on the core answer content.

[0041] Through the above technical solutions, this application effectively solves the problems of noise interference and blurred handwriting edges in the original optical image data. The combination of wavelet transform decomposition and adaptive threshold denoising can accurately separate and remove paper texture and optical noise, significantly improving the signal-to-noise ratio of the image. Handwriting edge enhancement based on deep convolutional neural networks, especially the use of residual dense block structure and skip connections, enables the handwriting edges to be extracted clearly and completely even in complex backgrounds, overcoming the limitations of traditional edge detection methods that are sensitive to noise and prone to losing details. Finally, automatic segmentation of the answer region based on connected component analysis ensures that subsequent processing can obtain clean and accurate answer region image blocks. Overall, these preprocessing steps work together to greatly improve the quality and robustness of the preprocessed image data, providing high-quality input for cloud-based intelligent processing units to perform low-complexity semantic extraction and high-complexity grading inference, thereby significantly improving the accuracy, efficiency, and adaptability to diverse job scenarios of the intelligent grading system.

[0042] In one alternative implementation, the edge intelligence processing unit further includes a local inference cache module, a network status monitoring module, and a task offloading decision module.

[0043] The local inference cache module stores the preprocessed image data output by the edge intelligent processing unit and its corresponding preliminary classification results. This module can be implemented using high-speed cache memory (such as SRAM or DRAM) or non-volatile memory (such as NAND Flash) to ensure fast data access. Its main function is to provide a locally available copy of data for subsequent task offloading decisions when network conditions are poor, or to provide necessary data support when executing tasks locally.

[0044] The network status monitoring module is used to monitor the network connection status and bandwidth quality between the edge intelligent processing unit and the cloud intelligent processing unit in real time. This module can be implemented through various technical means, such as periodically sending probe packets (e.g., ICMP Echo Requests) to the cloud intelligent processing unit to assess network latency and packet loss rate, or estimating the current available bandwidth by measuring the actual data transmission rate. Furthermore, more detailed network statistics, such as network throughput and jitter, can be obtained by utilizing network interfaces provided by the operating system or third-party libraries.

[0045] The task offloading decision module connects to the network status monitoring module and dynamically determines the task processing strategy based on the current network bandwidth quality. This module has a built-in decision logic that judges the current network environment based on preset bandwidth thresholds and selects the optimal task execution location accordingly. Specifically: when the network bandwidth is greater than the first preset threshold, the task offloading decision module will completely package and upload the preprocessed image data and preliminary classification results to the cloud intelligent processing unit to fully utilize the powerful computing resources of the cloud; when the network bandwidth is less than or equal to the first preset threshold but greater than the second preset threshold, to save bandwidth and maintain processing efficiency, the task offloading decision module will only upload high-complexity task data to the cloud intelligent processing unit; when the network bandwidth is less than or equal to the second preset threshold, to avoid long delays caused by network transmission, the task offloading decision module will instruct the edge intelligent processing unit to execute all processing tasks locally.

[0046] Through the above technical solution, the edge intelligent processing unit of this application can intelligently adjust its task processing strategy according to real-time network conditions. When network bandwidth is sufficient, the system can fully utilize the powerful computing capabilities of the cloud-based intelligent processing unit to achieve efficient and accurate graded inference. When network bandwidth is limited, data is stored through a local inference cache module, and the task offloading decision module dynamically selects to execute some or all tasks locally at the edge based on feedback from the network status monitoring module, thereby effectively avoiding service interruptions and performance degradation caused by network transmission delays or interruptions. This adaptive task offloading mechanism significantly improves the robustness and availability of the system, ensuring that the intelligent teaching assistance system can provide a stable and smooth user experience in different network environments, optimizing resource utilization efficiency, and reducing dependence on network infrastructure.

[0047] In one alternative implementation, the cloud-based intelligent processing unit includes a first-level large language model inference engine and a second-level large language model inference engine.

[0048] The first-level large language model inference engine primarily receives preprocessed image data and preliminary classification results uploaded by the edge intelligent processing unit. This engine is configured to call a lightweight large language model with no more than 3 billion parameters to perform preliminary semantic extraction on the answer content, thereby generating an initial semantic vector. This lightweight large language model is typically optimized using techniques such as model pruning, quantization, or knowledge distillation to significantly reduce computational resource consumption and inference latency while maintaining a certain level of semantic understanding, enabling it to efficiently handle a large number of relatively simple grading tasks. Subsequently, the engine inputs the question type identifier, the coordinates of the answer area, and the generated initial semantic vector into an attention-based classification network. This classification network can focus on key information in the input data through an attention mechanism, such as the core vocabulary of the question or the specific structure of the answer area, thereby more accurately assessing the complexity of the task and outputting a corresponding task complexity score.

[0049] The second-level large language model inference engine connects to the first-level engine and its primary responsibility is to receive high-complexity task data with complexity scores exceeding a preset threshold. For these identified high-complexity tasks, the second-level engine performs deep semantic understanding and grading reasoning, ultimately generating graded results. Unlike the first-level engine, the second-level engine typically invokes a general-purpose large language model with a larger number of parameters and stronger reasoning capabilities to handle complex grading scenarios such as open-ended questions, multi-step problem-solving, and logical reasoning. Deep semantic understanding involves a comprehensive analysis of the text's context, logical relationships, and implicit meanings, while grading reasoning combines preset scoring criteria, knowledge point associations, and domain expertise to ensure the accuracy and professionalism of the grading results.

[0050] Through the above technical solution, this application achieves intelligent hierarchical processing of grading tasks with varying complexities. The first-level large language model inference engine can quickly filter and process most low-complexity tasks, significantly improving the overall processing efficiency and response speed of the system, and avoiding unnecessary occupation of high-performance computing resources by simple tasks. Simultaneously, by accurately diverting high-complexity tasks to the second-level large language model inference engine for in-depth processing, the accuracy and quality of complex grading tasks are ensured. This hierarchical processing mechanism not only optimizes the utilization efficiency of computing resources and reduces the system's operating costs, but also effectively improves the performance and user experience of the intelligent teaching support system when handling massive and diverse assignments by applying models of different scales and capabilities in a targeted manner.

[0051] In one optional implementation, the learning analysis unit includes a wrong question classification and storage module, a knowledge status tracking module, a weak point location module, and a learning trajectory recording module.

[0052] The error classification and storage module receives the grading results generated by the cloud-based intelligent processing unit and automatically labels and classifies the errors based on a preset error type classification model. Error types include conceptual errors, calculation errors, comprehension errors, and expression errors. Specifically, the core of this module lies in its error type classification model, which can be a classifier based on Natural Language Processing (NLP) and machine learning techniques. By analyzing student answers, grading feedback, and question types, the model can identify and summarize the underlying causes of errors. For example, conceptual errors may refer to students' confusion about basic definitions, principles, or formulas; calculation errors specifically refer to mistakes or incorrect steps in the calculation process; comprehension errors refer to students' misunderstandings of the question, the application scenario of the knowledge point, or the problem-solving approach; and expression errors may manifest as unclear expression, illogical reasoning, or non-compliance with standards. This detailed error classification provides more instructive and targeted raw data for subsequent learning analysis.

[0053] The knowledge status tracking module is connected to the error classification and storage module. Its function is to dynamically update the student's mastery probability value for each knowledge point based on historical error data using a deep learning model based on recurrent neural networks (RNNs). RNN models are particularly suitable for processing sequential data, such as students' performance and error records on different questions at different times. This model can capture the trends and patterns of changes in students' knowledge mastery status over time. For example, when a student makes the same type of error consecutively on a specific knowledge point, the RNN model will correspondingly decrease the mastery probability value for that knowledge point; conversely, if a student shows consistent accuracy and comprehension on related knowledge points, it will increase their mastery probability value. This dynamic update mechanism allows the learning profile to reflect students' learning progress and fluctuations in knowledge mastery in real time and accurately, rather than simply being a static record of grades.

[0054] The weakness identification module connects to the knowledge status tracking module. Its function is to compare the mastery probability value output by the knowledge status tracking module with preset multi-level thresholds to identify students' weaknesses. For example, a first-level threshold and a second-level threshold can be preset, where the second-level threshold is lower than the first-level threshold. When the mastery probability value of a knowledge point falls below the first-level threshold, that knowledge point is marked as a "weak point," indicating that the student's grasp of that knowledge point is not solid or their understanding is incomplete. If the mastery probability value further falls below the second-level threshold, the knowledge point will be marked as a "serious weakness," indicating that the student has a serious deficiency in that knowledge point and urgently needs strengthening and tutoring. This hierarchical identification mechanism helps the teaching support system prioritize and solve students' most pressing and critical learning problems, achieving optimal resource allocation.

[0055] The learning trajectory recording module records the grading results and error classification results of each student's assignment, organizing them into time-series learning trajectory data. This data includes, but is not limited to, the submission time of each assignment, score, specific error content, detailed error classification (such as conceptual errors, calculation errors, etc.), and grading feedback provided by the cloud-based intelligent processing unit. By recording the complete learning trajectory, this module not only provides rich and continuous historical data for the knowledge status tracking module, ensuring the accuracy of knowledge status updates, but also provides teachers and students with an intuitive review of the learning process, helping them analyze learning habits, progress curves, and the evolution of knowledge mastery.

[0056] Through the aforementioned technical solutions, this application overcomes the limitations of traditional learning analysis, which is limited to scores or simple right-and-wrong statistics. The error classification and storage module can deeply analyze the essence of students' errors, attributing them to concepts, calculations, understanding, or expression, providing accurate diagnostic basis for subsequent personalized tutoring. Building on this, the knowledge state tracking module utilizes the powerful sequence modeling capabilities of recurrent neural networks to dynamically capture the long-term trends and short-term fluctuations in students' knowledge mastery, making the learning profile no longer a static snapshot but a real-time updated and dynamic learning state map. The weakness identification module further transforms these dynamic mastery probability values ​​into actionable weakness and severe weakness identifiers, providing clear targets for the personalized exercise generation unit and ensuring that the generated exercise sets accurately address students' knowledge gaps. Simultaneously, the learning trajectory recording module constructs complete time-series learning data, supporting not only the accuracy of knowledge state tracking but also providing students and teachers with a comprehensive review of the learning process. Overall, these modules work together to enable the system to construct a more refined, dynamic, and predictive multi-dimensional learning profile, significantly improving the accuracy and effectiveness of personalized exercise set generation, ultimately achieving more targeted and efficient intelligent teaching assistance.

[0057] In addition, such as Figure 2 As shown, the present invention also provides an intelligent teaching assistance method based on a large model, comprising the following steps: S1. Obtain the original optical image data of the job to be corrected on the job placement platform, and perform optical path distortion correction on the original optical image data to generate optical image data; S2. Perform the first-level image preprocessing on the optical image data to generate preprocessed image data; S3. Perform low-complexity semantic extraction and high-complexity correction reasoning on the preprocessed image data to generate correction results; S4. Overlay the correction results onto the corresponding answer area of ​​the optical image data in the form of simulated handwriting, and drive the output device to generate visual correction marks. S5. Collect and store grading results data to build a multi-dimensional learning profile; S6. Generate personalized practice question sets based on multi-dimensional learning profiles.

[0058] In one optional implementation, step S1 specifically includes: The incident light angle and intensity are collected using an ambient light sensor; The position and curvature of the compensation lens in the optical path compensation assembly are dynamically adjusted according to the incident light angle and intensity. The original optical image data of the work to be corrected on the work platform is acquired by a wide-angle optical lens, edge distortion is corrected, and optical image data is generated.

[0059] In one optional implementation, step S2 specifically includes: Wavelet transform decomposition is performed on optical image data to separate low-frequency approximate components and high-frequency detail components; An adaptive threshold denoising algorithm is applied to the high-frequency detail components to remove high-frequency components containing paper texture and optical noise. The high-frequency detail components and low-frequency approximation components after denoising are reconstructed by inverse wavelet transform to generate the denoised image. The handwriting edge enhancement based on a deep convolutional neural network is performed on the denoised image. The handwriting edge enhancement adopts a residual dense block structure, which fuses shallow edge features and deep semantic features through skip connections to generate the handwriting edge enhanced image. Automatic segmentation of the answer region based on connected component analysis is performed on the image after handwriting edge enhancement, and the segmented answer region image blocks are output as the preprocessed image data.

[0060] In one optional implementation, step S3 specifically includes: The first-level large language model inference engine receives preprocessed image data and preliminary classification results uploaded by the edge intelligent processing unit, and calls the first lightweight large language model with no more than 3 billion parameters to perform preliminary semantic extraction on the answer content and generate semantic initial vectors. Input the question type identifier, answer area coordinates, and semantic initial vector into an attention-based classification network, and output a task complexity score; Tasks with a complexity score higher than a preset threshold are marked as high-complexity task data and diverted to the second-level large language model inference engine; The second-level large language model inference engine performs deep semantic understanding and correction inference on highly complex task data to generate correction results.

[0061] The following example will provide a more detailed explanation of the above technical solution: This embodiment uses the daily homework grading of mathematics for the second year of junior high school (12 classes, 480 students) in a provincial-level demonstration middle school as an application scenario to explain in detail the specific implementation process and technical effects of the system of the present invention.

[0062] I. Application Scenarios and Deployment Environment: Located in a suburban area, the school suffers from unstable network conditions, with Wi-Fi upload bandwidth in some classrooms fluctuating between 1 and 15 Mbps. Teachers need to grade approximately 40 paper assignments per class each day, with each assignment taking an average of 3 minutes to grade manually, totaling more than 2 hours for the entire class.

[0063] The system deployment is as follows: Each classroom is equipped with one intelligent data acquisition terminal (integrating a spatial optical imaging and optical path compensation unit, an edge intelligent processing unit, and a thermal printer).

[0064] The cloud servers are deployed on the campus private cloud (2 NVIDIA A10 GPU servers, each with 24GB of video memory and 256GB of RAM).

[0065] Students can use regular lined notebooks or exercise books; no special paper or answer sheets are required.

[0066] II. Complete Homework Correction Process: Step S1: One day, the math teacher of Class 3, Grade 8 assigned a unit homework on "Quadratic Equations in One Variable", which consisted of 5 questions (2 multiple choice questions, 2 fill-in-the-blank questions, and 1 application question). The teacher placed all 42 homework books in the automatic paper tray of the smart data collection terminal page by page.

[0067] S11: The built-in ambient light sensor (model TSL2591) of the terminal detects in real time that the incident light intensity on the left side of the work page is 320 lux and on the right side is 180 lux, with incident angles of 15° and -8° respectively (due to uneven lighting from the side windows of the classroom).

[0068] S12: Based on the above data, the PID controller of the optical path compensation component calculates that the compensation lens needs to move 0.23mm along the optical axis, and the radius of curvature is adjusted from 120mm to 108mm to correct edge distortion.

[0069] S13: Wide-angle optical lens (4mm focal length, 80° field of view) scanned at 300 DPI resolution, generating raw optical image data per page (2480×3508 pixels, approximately 8.7MB). After the above dynamic adjustments, the edge sharpness of the generated optical image data was improved by 42% (according to SFR testing, the edge MTF50 improved from 0.28 to 0.40). The total scanning time for 84 pages (42 double-sided jobs) was 126 seconds (average 1.5 seconds / page).

[0070] Step S2: Optical image data is transmitted via USB 3.0 to the edge intelligent processing unit (Rockchip RK3588, NPU computing power 6 TOPS, memory 4GB), and the following sub-steps are executed: S21: The Daubechies-4 wavelet is used for 3-level decomposition to separate the low-frequency approximation component (LL3) and the high-frequency detail component (HL3, LH3, HH3).

[0071] S22: Calculate the standard deviation σ of the high-frequency coefficients in the third layer, which is 27.5. Threshold. High-frequency coefficients with an absolute value less than 104.5 were set to zero (retaining approximately 35% of the coefficients) to remove paper texture and optical noise. The PSNR of the denoised image improved from 28.6 dB to 36.2 dB.

[0072] S23: Reconstruct the denoised high-frequency detail components and low-frequency approximation components using inverse wavelet transform to generate the denoised image.

[0073] S24: A pre-trained residual dense block network was used (trained for 50 epochs on a dataset of 100,000 handwritten mathematical symbols, with a training loss of L1=0.023). The input denoised image (512×512) was processed through 4 residual dense blocks, and shallow edge features and deep semantic features were fused through skip connections. The average gradient magnitude of the handwriting edges was improved by 2.3 times, and the probability of a sloppy "5" being mistaken for a "3" decreased from 24% to 4.5%.

[0074] S25: The enhanced image was binarized (OTSU thresholded) using 4-neighbor connected component labeling, and a total of 215 connected components were detected. 210 valid answer region image blocks were cropped based on the bounding rectangle coordinates (5 were filtered out due to no answers), each block being approximately 150KB, for a total data volume of 31.5MB. The average preprocessing time per page was 2.8 seconds, and the total time for 84 pages was approximately 235 seconds (3.9 minutes).

[0075] Step S3: The network status monitoring module of the edge intelligent processing unit detects that the current classroom Wi-Fi uplink bandwidth is 4.7 Mbps and the RTT is 48ms. The task offloading decision module executes a hybrid strategy based on preset thresholds (Th_high=10 Mbps, Th_low=2 Mbps): low-complexity tasks (multiple choice questions and fill-in-the-blank questions, a total of 168 image patches) are processed locally at the edge; high-complexity tasks (application questions, a total of 42 image patches) are uploaded to the cloud.

[0076] S31: The cloud receives 42 image patches of application problems. The first-level engine calls the Phi-3-mini model with 2.7B parameters to perform preliminary semantic extraction on the answer content, generating a 768-dimensional semantic initial vector. The question type identifier (application problem), the coordinates of the answer area, and the semantic initial vector are input into an attention-based classification network, which outputs a task complexity score C. The C values ​​of the 12 standard calculation problems are between 0.35 and 0.52 (below the threshold of 0.7), while the C values ​​of the 30 practical application modeling problems are between 0.73 and 0.89 (above the threshold).

[0077] S32: Mark 30 tasks with a C value higher than 0.7 as high complexity and redirect them to the second-level engine.

[0078] S33: The second-level engine calls the Llama-3-13B model (13B parameters, 2 A10 GPU tensor parallel processing) to perform deep semantic understanding and grading reasoning on 30 high-complexity application problems. The average reasoning time is 2.1 seconds per problem, generating a score (out of 10) and targeted comments (e.g., "The equality relationship is correct, but the sign is incorrect when solving the equation"). 12 low-complexity calculation problems are directly graded by Phi-3-mini.

[0079] Example of grading results (a word problem): Student's answer: "Let the width be x, the length be x+5, 2(x+x+5)=30, solve for x=5, area=5×10=50 square meters." Llama-3-13B rating: 9 out of 10. Comments: "The equation is correct and the solution steps are complete, but the unit conversion instructions are missing, so deduct 1 point." Overall grading results: 42 assignments with a total of 210 questions. The correct answer rate for multiple-choice questions was 82.1%, the correct answer rate for fill-in-the-blank questions was 73.8%, and the average score for word problems was 7.2 / 10. The total grading time (including edge preprocessing, local grading, uploading, cloud inference, and return upload) was approximately 6.8 minutes, plus 2.1 minutes for scanning, for a total of approximately 9 minutes.

[0080] Step S4: The interactive feedback unit receives the correction results: Handwriting features (stroke width 0.8mm, pressure variation 0.23, tilt 8°) were extracted from 2,000 historical correction samples from the school's mathematics teachers, and cGAN was trained to generate simulated handwriting.

[0081] A red "√" will be generated next to the correct answer; for incorrect answer questions, the wrong steps will be underlined with a wavy line, and a comment will be generated in the blank space: "Incorrect symbol, note that the sign should be changed when moving terms," ​​and "7 / 10" will be marked.

[0082] The composite image is printed onto the original worksheet using a thermal printer (300 DPI), taking approximately 3 seconds per page. The worksheet is then automatically fed out after printing.

[0083] The teacher randomly checked 10 samples and concluded that the similarity between the simulated handwriting and the teacher's own handwriting was over 85%, so there was no need to rewrite the annotations by hand.

[0084] Step S5: Collect and store the grading results data to construct a multi-dimensional learning profile; The learning analysis unit executes the following sub-modules: Error classification and storage module: An error type classification model fine-tuned based on BERT (trained on 2000 labeled samples, with an accuracy of 91.3%) automatically labels incorrect questions. Statistical results: Conceptual errors 18%, computational errors 52%, comprehension errors 24%, and expressive errors 6%.

[0085] The knowledge state tracking module uses a GRU (128-dimensional hidden layer) RNN model as input to update the mastery probability values ​​of seven knowledge points related to "quadratic equations". For example, student Zhang San's original mastery probability value for "quadratic formula" was p=0.52 (weak), which was updated to 0.68 after answering one question correctly; his mastery probability value for "practical application modeling" was p=0.31 (severely weak), which was updated to 0.35 after scoring 5 / 10.

[0086] Weakness identification module: Preset first-level threshold θ1=0.6 (weakness), second-level threshold θ2=0.4 (serious weakness). Class commonality: Average p=0.42 in practical application modeling (weakness). Individual: Zhang San's practical application modeling p=0.35 (serious weakness), root formula p=0.68 (meets standard).

[0087] Learning trajectory recording module: Generates time-series data, recording Zhang San's scores in "Application Modeling" in his last 4 assignments: 3 / 10, 4 / 10, 5 / 10, 5 / 10, showing a slow upward trend.

[0088] Step S6: Generate a personalized practice question set based on a multi-dimensional learning profile; The personalized practice generation unit retrieves and matches questions from the school-based knowledge base (the school's junior high school math question bank, containing 3200 questions): Common weaknesses in the class: Generate 5 practical application problems as supplementary exercises for the class (to be distributed uniformly to the whole class).

[0089] Zhang San's three major weaknesses were addressed by generating three additional personalized practice questions (difficulty level 0.5-0.7, from easy to difficult), each with hints and explanations. The practice questions were pushed to the student's tablet, and the system automatically graded them upon completion (reusing the two-level cloud engine), and the new results were then input into the knowledge tracking model.

[0090] Closed-loop effect: After two weeks (4 assignments + personalized practice), Zhang San's mastery probability in "practical application modeling" increased from 0.35 to 0.61, moving away from the severely weak area; the class average score increased from 7.2 / 10 to 8.5 / 10.

[0091] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A large-scale intelligent teaching support system, characterized in that, include: The space optical imaging and optical path compensation unit is used to acquire the original optical image data of the work to be corrected on the work placement platform, and to perform optical path distortion correction on the original optical image data to generate optical image data. The edge intelligent processing unit is communicatively connected to the spatial optical imaging and optical path compensation unit, and is used to perform first-level image preprocessing on the optical image data to generate preprocessed image data. The cloud-based intelligent processing unit is communicatively connected to the edge intelligent processing unit and is used to perform low-complexity semantic extraction and high-complexity correction reasoning on the preprocessed image data to generate correction results. The interactive feedback unit is communicatively connected to the cloud-based intelligent processing unit and is used to overlay the correction results onto the corresponding answer area of ​​the optical image data in the form of simulated handwriting, and drive the output device to generate visual correction marks. The learning analysis unit is connected to the cloud-based intelligent processing unit to collect and store grading result data and construct a multi-dimensional learning profile. The personalized practice generation unit is communicatively connected to the learning analysis unit and is used to generate personalized practice question sets based on the multi-dimensional learning profile.

2. The system according to claim 1, characterized in that, The space optical imaging and optical path compensation unit includes: A wide-angle optical lens is placed above the work platform; An optical path compensation component is coupled to the optical path of the wide-angle optical lens and is equipped with at least one compensation lens; An ambient light sensor, electrically connected to the wide-angle optical lens, is used to collect the incident light angle and intensity; The optical path compensation component is also electrically connected to the ambient light sensor, and is used to dynamically adjust the position and curvature of the compensation lens according to the incident light angle and intensity collected by the ambient light sensor, so as to correct the edge distortion of the original optical image data.

3. The system according to claim 1, characterized in that, The first-level image preprocessing performed by the edge intelligent processing unit includes: The optical image data is decomposed by wavelet transform to separate the low-frequency approximation component and the high-frequency detail component; An adaptive threshold denoising algorithm is applied to the high-frequency detail components to remove high-frequency components containing paper texture and optical noise. The high-frequency detail components and low-frequency approximation components after denoising are reconstructed by inverse wavelet transform to generate the denoised image. The handwriting edge enhancement based on a deep convolutional neural network is performed on the denoised image; the handwriting edge enhancement adopts a residual dense block structure, which fuses shallow edge features and deep semantic features through skip connections to generate an image with enhanced handwriting edges; Automatic segmentation of the answer region based on connected component analysis is performed on the image after handwriting edge enhancement, and the segmented answer region image blocks are output as the preprocessed image data.

4. The system according to claim 1, characterized in that, The edge intelligence processing unit also includes: A local inference cache module is used to store the preprocessed image data output by the edge intelligent processing unit and its corresponding preliminary classification results; The network status monitoring module is used to detect the network connection status and bandwidth quality between the edge intelligent processing unit and the cloud intelligent processing unit in real time. The task unloading decision module, connected to the network status monitoring module, dynamically determines the task processing strategy based on the current network bandwidth quality: when the network bandwidth is greater than a first preset threshold, the preprocessed image data and preliminary classification results are packaged and uploaded to the cloud intelligent processing unit; when the network bandwidth is less than or equal to the first preset threshold but greater than a second preset threshold, only the high-complexity task data is uploaded; when the network bandwidth is less than or equal to the second preset threshold, all processing tasks are executed locally on the edge intelligent processing unit.

5. The system according to claim 1, characterized in that, The cloud-based intelligent processing unit includes: The first-level large language model inference engine is used to receive preprocessed image data and preliminary classification results uploaded by the edge intelligent processing unit. It calls the first lightweight large language model with no more than 3 billion parameters to perform preliminary semantic extraction on the answer content, generate semantic initial vector, and input the question type identifier, answer area coordinates and semantic initial vector into the attention mechanism-based classification network to output the task complexity score. The second-level large language model inference engine is connected to the first-level large language model inference engine. It is used to receive high-complexity task data with a complexity score higher than a preset threshold, perform deep semantic understanding and correction inference, and generate correction results.

6. The system according to claim 1, characterized in that, The learning analysis unit includes: The error classification and storage module is used to receive the grading results and automatically label the errors based on a preset error type classification model; the error types include conceptual errors, calculation errors, comprehension errors, and expression errors; The knowledge status tracking module is connected to the wrong question classification and storage module, and is used to dynamically update the probability value of students' mastery of each knowledge point based on the students' historical wrong question data using a deep learning model based on recurrent neural networks. The weakness point location module is connected to the knowledge status tracking module and is used to compare the mastery probability value with a preset multi-level threshold. The first knowledge point whose mastery probability value is lower than the first level threshold is marked as a weakness point, and the second knowledge point whose mastery probability value is lower than the second level threshold is marked as a serious weakness point. The learning trajectory recording module is used to record the grading results and error classification results of each student's homework, forming time-series learning trajectory data.

7. A large-model-based intelligent teaching assistance method, applied to the system according to any one of claims 1 to 6, characterized in that, Includes the following steps: S1. Obtain the original optical image data of the job to be corrected on the job placement platform, and perform optical path distortion correction on the original optical image data to generate optical image data; S2. Perform first-level image preprocessing on the optical image data to generate preprocessed image data; S3. Perform low-complexity semantic extraction and high-complexity correction reasoning on the preprocessed image data to generate correction results; S4. The correction result is superimposed onto the corresponding answer area of ​​the optical image data in the form of simulated handwriting, and the output device is driven to generate a visual correction mark. S5. Collect and store grading results data to build a multi-dimensional learning profile; S6. Generate a personalized set of practice questions based on the multi-dimensional learning profile.

8. The method according to claim 7, characterized in that, S1 specifically includes: The incident light angle and intensity are collected using an ambient light sensor; The position and curvature of the compensation lens in the optical path compensation assembly are dynamically adjusted according to the incident light angle and intensity. The original optical image data of the work to be corrected on the work platform is acquired by a wide-angle optical lens, edge distortion is corrected, and optical image data is generated.

9. The method according to claim 7, characterized in that, S2 specifically includes: The optical image data is decomposed by wavelet transform to separate the low-frequency approximation component and the high-frequency detail component; An adaptive threshold denoising algorithm is applied to the high-frequency detail components to remove high-frequency components containing paper texture and optical noise. The high-frequency detail components and low-frequency approximation components after denoising are reconstructed by inverse wavelet transform to generate the denoised image. The handwriting edge enhancement based on a deep convolutional neural network is performed on the denoised image. The handwriting edge enhancement adopts a residual dense block structure and fuses shallow edge features and deep semantic features through skip connections to generate an image with enhanced handwriting edges. Automatic segmentation of the answer region based on connected component analysis is performed on the image after handwriting edge enhancement, and the segmented answer region image blocks are output as the preprocessed image data.

10. The method according to claim 7, characterized in that, S3 specifically includes: The first-level large language model inference engine receives preprocessed image data and preliminary classification results uploaded by the edge intelligent processing unit, and calls the first lightweight large language model with no more than 3 billion parameters to perform preliminary semantic extraction on the answer content and generate semantic initial vectors. Input the question type identifier, answer area coordinates, and semantic initial vector into an attention-based classification network, and output a task complexity score; Tasks with a complexity score higher than a preset threshold are marked as high-complexity task data and diverted to the second-level large language model inference engine; The second-level large language model inference engine performs deep semantic understanding and correction inference on highly complex task data to generate correction results.

Images