A deep learning-based Chinese character visual recognition training system

Abstract

The present application relates to the field of education technology and artificial intelligence technology, in particular to a Chinese character visual recognition training system based on deep learning. The system comprises a handwriting recognition module, a user state monitoring module, a decision arbitration module and a dynamic intervention module. The handwriting recognition module is used for receiving handwriting sequence data input by a user and identifying the handwriting sequence data to obtain handwriting recognition confidence. The user state monitoring module is used for obtaining a cognitive load index quantifying the cognitive load state of the user. The decision arbitration module is used for obtaining a final intervention intensity, which decouples the direct binding relationship between handwriting recognition confidence and teaching intervention intensity. The dynamic intervention module is used for executing teaching feedback with intensity, granularity and frequency adapted to the user according to the final intervention intensity. The present application breaks the vicious negative feedback cycle caused by poor user state leading to strong intervention, which in turn leads to worse user state, and achieves the synergistic improvement of teaching effectiveness and system robustness.

Description

A Deep Learning-Based Chinese Character Visual Recognition Training System Technical Field

[0001] This invention relates to the fields of educational technology and artificial intelligence technology, specifically to a Chinese character visual recognition training system based on deep learning. Background Technology

[0002] In the current field of intelligent assisted teaching, especially in Chinese character writing training systems, the effectiveness of teaching intervention is crucial. To improve learning efficiency, the system usually needs to provide real-time feedback and guidance based on the learner's handwriting. The triggering and intensity of this feedback mainly depend on the handwriting recognition algorithm's evaluation of the user's writing standardization, such as recognition confidence. Traditional training systems are generally designed to directly and linearly correlate handwriting recognition confidence with the intensity of teaching intervention.

[0003] In existing technologies, when a system detects that a user's handwriting is not standardized or contains errors, it immediately initiates a preset error correction intervention. However, this intervention mechanism suffers from a fundamental problem: a lack of awareness of the user's state. It cannot distinguish whether the writing error is caused by a skill deficiency, such as unfamiliarity with the structure of Chinese characters, or by non-skill fluctuations, such as fatigue, lack of concentration, or emotional distress. This indiscriminate intervention logic can easily lead to a vicious cycle of negative feedback between the teaching intervention and the user's state.

[0004] When a user's handwriting quality declines due to non-skill factors such as fatigue, the system will identify it as a writing error and apply stronger intervention. This will further increase the user's cognitive load or trigger negative emotions, resulting in even worse subsequent handwriting performance, which in turn triggers more frequent and stronger interventions from the system. Current technology lacks real-time monitoring and evaluation of the user's cognitive load, causing intervention decisions to rely solely on outcome-based handwriting evaluation. This results in a lack of intelligent and humanistic considerations in teaching behavior, which may not only reduce the actual teaching effect but also impair learners' learning interest and handwriting fluency.

[0005] Therefore, how to provide a Chinese character visual recognition training method that can decouple handwriting recognition evaluation from teaching intervention decision-making and dynamically adjust the intensity of intervention according to the user's real-time cognitive state is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention discloses a deep learning-based Chinese character visual recognition training system. Specifically, the technical solution includes:

[0007] A deep learning-based Chinese character visual recognition training system includes:

[0008] The handwriting recognition module is used to receive handwriting sequence data input by the user, and to process the handwriting sequence data to obtain the handwriting recognition confidence score.

[0009] The user status monitoring module is used to synchronously receive handwriting sequence data and analyze the dynamic characteristic parameters of the handwriting sequence data to obtain the cognitive load index, which quantifies the user's cognitive load status.

[0010] The decision arbitration module receives the handwriting recognition confidence score and cognitive load index, and modulates the intensity of teaching intervention nonlinearly based on the dynamic adjustment model to obtain the final intervention intensity. The final intervention intensity decouples the direct binding relationship between handwriting recognition confidence score and teaching intervention intensity.

[0011] The dynamic intervention module is used to provide users with teaching feedback that is tailored to the intensity, granularity, and frequency based on the final intervention intensity.

[0012] Preferably, the specific processing procedure of the user status monitoring module includes:

[0013] The system guides users to write preset benchmark Chinese characters, collects and processes the handwriting dynamics feature data of the benchmark Chinese characters, and obtains personalized benchmark state parameters.

[0014] During training, real-time dynamic features of the user's current handwriting sequence are extracted;

[0015] The cognitive load index is generated by comparing real-time dynamic characteristics with baseline state parameters.

[0016] Preferably, the dynamic characteristic data of the handwriting includes average writing speed, average writing pressure, average pause time between strokes, and stroke trajectory tremor frequency.

[0017] Preferably, the specific processing procedure of the decision arbitration module includes:

[0018] Based on the confidence level of handwriting recognition, the baseline intervention intensity reflecting the system's direct response to writing errors is calculated;

[0019] Based on the cognitive load index, an empathy modulation factor was calculated to dynamically suppress the baseline intervention intensity.

[0020] The baseline intervention intensity is modulated by multiplying it with the empathy modulation factor to obtain the final intervention intensity.

[0021] Preferably, the calculation process of the benchmark intervention intensity is as follows: multiply the preset maximum system intervention intensity constant by the Nth power of the difference between the handwriting recognition confidence level and 1, where N is a preset intervention sensitivity coefficient greater than 1.

[0022] Preferably, the calculation process of the empathy modulation factor is as follows: based on the sigmoid function, according to the difference between the cognitive load index and the preset cognitive load tolerance threshold, an inhibition factor that smoothly transitions near the cognitive load tolerance threshold is generated, and the inhibition factor is the empathy modulation factor.

[0023] Preferably, the dynamic intervention module has a preset set of mapping rules from intervention intensity to specific intervention behavior, which is used to map the final intervention intensity to discrete feedback units with different granularities and modalities.

[0024] Preferably, the mapping rule set includes:

[0025] When the final intervention intensity is in the preset extremely low range, no intervention is carried out or only positive encouragement is given;

[0026] When the final intervention intensity is within the preset medium range, macro-level feedback is provided;

[0027] When the final intervention intensity is in the preset high range, fine-grained feedback is provided.

[0028] Preferably, the mapping rule set further includes: when the cognitive load index is higher than a preset threshold, resulting in the suppression of the final intervention intensity, executing intervention behaviors unrelated to error correction and aimed at alleviating the user's emotions.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. This invention separates the objective evaluation of handwriting standardization from the intervention decision by introducing the user's real-time cognitive load as an independent dimension for teaching intervention decisions. When the system identifies that the user's low-confidence handwriting is caused by non-skill factors such as fatigue, it will actively reduce the intensity of intervention, breaking the vicious negative feedback loop of strong intervention due to poor user condition, which in turn leads to a worse user condition. This achieves a synergistic improvement in teaching effectiveness and system robustness.

[0031] 2. This invention establishes a personalized state reference system by guiding users to write benchmark Chinese characters and compares dynamic characteristics such as writing speed, pressure, pauses and trembling frequency in real time. It can accurately and reliably quantify and assess the degree to which users deviate from their normal state. This method ensures that the state assessment is based on the user's own habits rather than universal standards, which significantly improves the accuracy of the cognitive load index.

[0032] 3. This invention adopts a two-level derivation mechanism that multiplies the baseline intervention intensity by the empathy modulation factor, so that the intervention decision takes into account both the severity of the error and the user's real-time status. This design logically separates and then organically combines error assessment and status consideration, making the intervention decision-making process more intelligent and humane. It can accurately reflect the writing problem while determining the intervention behavior in an empathetic way.

[0033] 4. This invention has a pre-set mapping rule set from continuous intervention intensity to discrete teaching behavior, which can execute teaching feedback with appropriate intensity, granularity and frequency according to the final intervention intensity. This hierarchical mapping rule makes the system feedback have a sense of hierarchy that matches the severity of the problem, avoiding excessive intervention for minor flaws and ensuring accurate guidance for serious errors, thus realizing the effective allocation of teaching resources. Attached Figure Description

[0034] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0035] Figure 1 is a flowchart of a deep learning-based Chinese character visual recognition training system according to the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0037] Example 1:

[0038] A deep learning-based Chinese character visual recognition training system includes:

[0039] The handwriting recognition module is used to receive handwriting sequence data input by the user, and to process the handwriting sequence data to obtain the handwriting recognition confidence score.

[0040] The user status monitoring module is used to synchronously receive handwriting sequence data and analyze the dynamic characteristic parameters of the handwriting sequence data to obtain the cognitive load index, which quantifies the user's cognitive load status.

[0041] The decision arbitration module receives the handwriting recognition confidence score and cognitive load index, and modulates the intensity of teaching intervention nonlinearly based on the dynamic adjustment model to obtain the final intervention intensity. The final intervention intensity decouples the direct binding relationship between handwriting recognition confidence score and teaching intervention intensity.

[0042] The dynamic intervention module is used to provide users with teaching feedback that is tailored to the intensity, granularity, and frequency based on the final intervention intensity.

[0043] This invention provides a deep learning-based Chinese character visual recognition training system. The system aims to intelligently adjust teaching interventions by sensing the user's cognitive state, thereby solving the vicious negative feedback loop problem between teaching interventions and handwriting recognition caused by the inability to distinguish between skill-based writing errors and non-skill-based state fluctuations in existing technologies. The system includes a handwriting recognition module, a user state monitoring module, a decision arbitration module, and a dynamic intervention module.

[0044] The handwriting recognition module aims to accurately identify and quantify the standardization of a user's handwriting. In this embodiment, the module is a pre-trained neural network model designed to receive handwriting sequence data dynamically input by the user through a touch device. This data stream mainly contains the coordinate sequence information of the handwriting. After processing the received data, the module's final output is the confidence level C for the current handwriting sequence being identified as the target Chinese character, which is a dimensionless value in the range [0,1].

[0045] The user status monitoring module aims to assess the user's cognitive load in real time, independently of character shape recognition. This module operates in parallel with the handwriting recognition module, synchronously receiving the same handwriting sequence data. It does not aim to recognize character shapes, but rather to obtain a cognitive load index Λ that quantifies the user's current cognitive load by analyzing the dynamic characteristic parameters of the handwriting sequence. This parallel processing design ensures that the accuracy assessment of character shape recognition and the real-time assessment of the user status can be performed simultaneously, providing complete and time-aligned input information for subsequent decision-making arbitration. To clearly illustrate the core idea of ​​this invention, the complex user cognitive state is simplified to a single index Λ.

[0046] The decision arbitration module aims to decouple the direct relationship between handwriting recognition confidence and the intensity of teaching intervention. This module receives the recognition confidence C output by the handwriting recognition module and the cognitive load index Λ output by the user status monitoring module. Its core function is to nonlinearly modulate a baseline teaching intervention intensity based on a dynamic adjustment model, thereby calculating and outputting a final intervention intensity I. f ;

[0047] The dynamic intervention module aims to transform the abstract intensity value calculated by the decision arbitration module into concrete and implementable teaching behaviors; this module is based on the received final intervention intensity I. f This provides users with instructional feedback that is appropriate in intensity, granularity, and frequency, forming the final link in the closed loop of the entire system.

[0048] This embodiment introduces the user's real-time cognitive state as an independent dimension for intervention decisions, separating the objective evaluation of handwriting standardization from the subjective decision of whether and how to conduct teaching intervention. When the system identifies that the user's low-confidence handwriting is likely caused by non-skill factors such as fatigue, it will proactively reduce the intensity of intervention, prioritizing the user's writing fluency and learning interest. This breaks the original technical contradiction and achieves a synergistic improvement in teaching effectiveness and recognition robustness.

[0049] Example 2:

[0050] The specific processing steps of the user status monitoring module include:

[0051] The system guides users to write preset benchmark Chinese characters, collects and processes the handwriting dynamics feature data of the benchmark Chinese characters, and obtains personalized benchmark state parameters.

[0052] During training, real-time dynamic features of the user's current handwriting sequence are extracted;

[0053] The cognitive load index is generated by comparing real-time dynamic characteristics with baseline state parameters.

[0054] Based on the system in Example 1, a preferred implementation of the user state monitoring module's processing procedure is described. This procedure aims to ensure the personalization and accuracy of state assessment. The processing procedure includes a user baseline state calibration step. Before the user first uses the system or before each training session, the system guides the user to write a set of preset baseline Chinese characters in a normal, relaxed state. This process does not involve any instructional intervention. During this process, the user state monitoring module collects and processes the dynamic characteristic data of the baseline handwriting, calculates and stores a set of personalized baseline state parameters. After calibration, during the training process, the module extracts the real-time dynamic characteristics of each handwriting sequence currently written by the user.

[0055] The module then compares the extracted real-time dynamic features with pre-stored baseline state parameters to generate a cognitive load index Λ that dynamically reflects the degree of deviation from the user's state. Specifically, to more accurately distinguish between different cognitive states, such as slower writing due to hesitation and faster writing due to impatience, this comparison calculation can be performed by a pre-set asymmetric weighted model, for example, the formula: Where f i Representing one of the four dynamic characteristics, namely velocity, pressure, pause time, and jitter frequency, w i,dir This is a directional weighting coefficient, and its value depends on the direction of the difference between the real-time feature and the baseline parameter. For example, when the difference is positive, w... i,dir Take the preset positive weight w i,positive When the difference is negative, a preset negative weight w is used. i,negative ;∈ is a preset minimal positive number used to avoid the influence of the reference state parameter f. i,baseline Calculation errors caused by a value of zero;

[0056] This set of weighting coefficients w iThe specific values ​​can be determined in the following way: experts in cognitive psychology or education are invited to make initial settings based on experience. Then, the collected handwriting data containing user state labels is used to train the model through a classification algorithm in machine learning, and the weights are optimized in reverse to maximize the distinguishability of the cognitive load index Λ to the user's real state. In another preferred embodiment, in order to further improve the model fidelity, a nonlinear model can be used to fit the relationship between real-time dynamic characteristics and baseline state parameters, so as to more accurately distinguish different cognitive states such as writing slower due to hesitation and writing faster due to impatience.

[0057] This process, which includes a benchmark calibration step, establishes a personalized reference system for subsequent state assessments. It ensures that the cognitive load index calculation is based on the user's normal writing habits, rather than a universal, fixed standard, significantly improving the accuracy and reliability of state assessments. To further enhance robustness, the system can prompt the user to perform the calibration in a relaxed state before it begins and can introduce a dynamic benchmark update mechanism. That is, during subsequent training, when the user's writing state remains stable and good, the benchmark state parameters are automatically fine-tuned to adapt to long-term changes in the user. For example, this dynamic benchmark update mechanism can be set as follows: when the system detects that the user's handwriting recognition confidence is higher than a preset threshold (e.g., 0.95) in N consecutive Chinese character writing tasks (N being a preset positive integer, such as 10), and the cognitive load index Λ is consistently below a low percentage (e.g., 20%) of the cognitive load tolerance threshold Λ0, the system can automatically trigger an update of the benchmark state parameters. The update algorithm can use the exponential moving average method, which incorporates the dynamic characteristic data of the current steady state into the original baseline state parameters with a small weight, thereby achieving a smooth and gradual adjustment.

[0058] Example 3:

[0059] The dynamic characteristics of handwriting include average writing speed, average writing pressure, average pause time between strokes, and stroke trajectory tremor frequency.

[0060] Based on the implementation of Example 2, the dynamic characteristic data of handwriting are further defined. In this example, the data specifically includes four core indicators: average writing speed, average writing pressure, average pause time between strokes, and stroke trajectory tremor frequency. The real-time data of these four indicators are collected and calculated in real time from the handwriting data stream generated when the user interacts with the touch device. Specifically, the average writing speed can be calculated from the distance the pen tip moves per unit time; the average writing pressure is the average pressure value returned by the sensor during writing; the average pause time between strokes is determined by setting a speed... A threshold is used to determine the stationary state of the pen tip and calculate its duration; the frequency of pen stroke trembling can be obtained by high-pass filtering the handwriting coordinate sequence and then analyzing its spectrum using Fast Fourier Transform (FFT); these feature parameters were chosen because they can jointly characterize the user's cognitive and physiological state from different dimensions; for example, the unit of average writing speed can be pixels per second, the unit of average writing pressure can be the relative pressure value in the range of 0-1024 provided by the device sensor, the unit of average pause time between strokes can be milliseconds (ms), and the unit of pen stroke trembling frequency can be Hertz (Hz).

[0061] Average writing speed v reflects the fluency of a user's writing and the speed of their decision-making;

[0062] Average writing pressure p reflects the user's level of tension and muscle control.

[0063] The average pause time τ between strokes reflects the degree of hesitation a user has when conceiving stroke connections;

[0064] The frequency j of stroke trajectory trembling reflects a decrease in muscle control stability caused by user fatigue or tension.

[0065] By comprehensively analyzing the dynamic characteristics of these four dimensions, the user status monitoring module can more accurately capture changes in writing behavior caused by fatigue, frustration, or lack of concentration, providing a solid data foundation for generating a high-precision cognitive load index.

[0066] Example 4:

[0067] The specific processing steps of the decision arbitration module include:

[0068] Based on the confidence level of handwriting recognition, the baseline intervention intensity reflecting the system's direct response to writing errors is calculated;

[0069] Based on the cognitive load index, an empathy modulation factor was calculated to dynamically suppress the baseline intervention intensity.

[0070] The baseline intervention intensity is modulated by multiplying it with the empathy modulation factor to obtain the final intervention intensity.

[0071] Based on the system in Example 1, a preferred implementation method is described for the specific processing of the decision arbitration module; this process optimizes intervention decisions based solely on errors into intelligent decisions that comprehensively consider the degree of error and the user's state through a two-level derivation mechanism;

[0072] Based on the handwriting recognition confidence level C provided by the handwriting recognition module, a baseline intervention intensity I is calculated. b This intensity value is designed to reflect only the system's direct response to the writing error itself; that is, the more serious the error, the stronger the initial intention to intervene.

[0073] Based on the cognitive load index Λ provided by the user status monitoring module, an empathy modulation factor ω is calculated; the function of this factor is to dynamically and non-linearly suppress the baseline intervention intensity according to the user's real-time cognitive state.

[0074] The calculated baseline intervention intensity I b Multiplicative modulation with the empathy modulation factor ω yields the final intervention intensity I. f Its calculation formula is: I f =I b ·ω;

[0075] This step-by-step calculation design clearly separates error assessment and state consideration logically, and organically combines the two through final multiplication modulation. The resulting technical effect is that the system can accurately assess the severity of writing errors while deciding whether and how to intervene in an empathetic manner, thereby achieving intelligent and humanized intervention decision-making.

[0076] Example 5:

[0077] The calculation process for the baseline intervention intensity is as follows: multiply the preset maximum intervention intensity constant of the system by the Nth power of the difference between the handwriting recognition confidence level and 1, where N is a preset intervention sensitivity coefficient greater than 1.

[0078] Based on the implementation method of Example 4, the calculation process of the benchmark intervention intensity is described in detail. This calculation process is derived from error-driven learning theory, and its purpose is to make the intervention intensity and the decrease in recognition confidence exhibit a non-linear amplification relationship. The calculation formula is as follows:

[0079] I b =I m (1-C) n

[0080] Among them, I b As the baseline intervention intensity; I mis the maximum intervention intensity constant of the system, a preset dimensionless upper limit value, which is a configurable parameter set according to the teaching objectives and the acceptance of the user group; C is the handwriting recognition confidence level output by the handwriting recognition module; n is the intervention sensitivity coefficient, a preset parameter greater than 1, which can be optimized and determined by regression analysis of teaching experimental data to achieve the best error amplification effect.

[0081] This formula uses an exponent n greater than 1 to amplify the impact of low confidence on intervention intensity. For example, when the confidence level slightly decreases from 0.9 to 0.8, the increase in intervention intensity is small; however, when the confidence level decreases from 0.2 to 0.1, the intervention intensity increases sharply. This design allows the system to make a stronger initial response to serious errors, which conforms to the basic principles of teaching intervention and provides a reasonable input that is strongly correlated with the degree of error for subsequent modulation steps. The robustness of the model is reflected in its boundary behavior: when the user writes perfectly and the recognition confidence level C→1, the baseline intervention intensity I... b →0, no system intervention; when the user writes completely incorrectly and confidence level C→0, the baseline intervention intensity I... b →I m The system generates an initial intervention intent of maximum strength. This ensures that the model's response is reasonable and controllable throughout the confidence interval.

[0082] Example 6

[0083] The calculation process of the empathy modulation factor is as follows: Based on the sigmoid function, an inhibition factor that smoothly transitions near the cognitive load tolerance threshold is generated according to the difference between the cognitive load index and the preset cognitive load tolerance threshold. The inhibition factor is the empathy modulation factor.

[0084] Based on the implementation method of Example 4, the calculation process of the empathy modulation factor is described in detail. This calculation process draws on the sigmoid function in control theory, aiming to achieve a soft-switching suppression effect with a smooth transition near a specific threshold. The calculation formula is as follows:

[0085]

[0086] Where ω is the empathy modulation factor; k is the state suppression slope coefficient, a preset positive constant, a dimensionless preset positive constant, which is derived from configurable parameters obtained by adjusting user test data to obtain a suitable suppression curve slope; since the cognitive load index Λ and the cognitive load tolerance threshold Λ0 are both dimensionless values ​​calculated through relative changes, in order to ensure that the index k(Λ-Λ0) is dimensionless; Λ is the cognitive load index calculated and provided in real time by the user state monitoring module. In this embodiment, Λ is designed as a dimensionless value in the range [0,∞), where 0 represents complete consistency with the baseline state; Λ0 is the cognitive load tolerance threshold, a key preset threshold;

[0087] The parameter Λ0 is determined by statistically analyzing the distribution of Λ values ​​during normal writing task cycles through user experiments and taking its higher quantile. It represents the upper limit of normal state fluctuations that the system can tolerate. When Λ is much smaller than Λ0, the value of ω approaches 1 and has almost no inhibitory effect on the baseline intervention intensity. When Λ exceeds Λ0, the value of ω will rapidly and smoothly approach 0, thus having a significant inhibitory effect on the baseline intervention intensity.

[0088] This S-shaped function-based calculation method avoids abrupt changes in intervention intensity caused by sudden state transitions, making the withdrawal or reduction of intervention smoother and more natural. Its technical effect is that it effectively executes state suppression decisions while ensuring the consistency and comfort of the user experience. The S-shaped function ensures the smoothness and stability of the modulation process. Its boundary behavior meets design expectations: when the user's cognitive load is far below the threshold (Λ << Λ0), ω → 1, the empathy modulation factor has no inhibitory effect, and the intervention intensity is determined by the writing error itself; when the user's cognitive load far exceeds the threshold (Λ >> Λ0), ω → 0, the empathy modulation factor produces near-complete inhibition, thus effectively preventing strong intervention when the user is in a poor state. This ensures that the system's behavior is safe and humane under various user states.

[0089] Example 7

[0090] The dynamic intervention module has a pre-set set of mapping rules from intervention intensity to specific intervention behaviors, which is used to map the final intervention intensity to discrete feedback units with different granularities and modalities.

[0091] Based on the system in Example 1, a preferred implementation of the internal structure of the dynamic intervention module is described. This module has a pre-defined set of mapping rules from intervention intensity to specific intervention behavior. The function of this rule set is to map the continuous, dimensionless final intervention intensity I output by the decision arbitration module. f Its range is [0, I m This is transformed into discrete teaching feedback units with different granularities and presentation modalities;

[0092] The pre-defined mapping rule set is the functional carrier for the system to transform abstract decisions into concrete actions; the construction of this rule set can be based on the theory of child educational psychology, and combined with statistical analysis of a large number of teaching cases and expert experience for calibration.

[0093] Through this set of mapping rules, the system can ensure that the final output of teaching behaviors has clear teaching significance and conforms to the user's acceptance habits, rather than a simple and crude division of intensity levels; this makes the intervention behavior of the entire system more standardized and scientific, and improves the professionalism and effectiveness of teaching.

[0094] Example 8

[0095] The mapping rule set includes:

[0096] When the final intervention intensity is in the preset extremely low range, no intervention is carried out or only positive encouragement is given;

[0097] When the final intervention intensity is within the preset medium range, macro-level feedback is provided;

[0098] When the final intervention intensity is in the preset high range, fine-grained feedback is provided.

[0099] Based on the implementation method of Example 7, the specific content of the mapping rule set is described by way of example; the rule set is designed as a hierarchical structure to cope with the final intervention intensity I in different intervals. f The specific thresholds for these intervals are relative to the system's maximum intervention intensity constant I. m The definition is based on user experience testing and teaching effectiveness evaluation; for example, the extremely low range can be set as [0, 0.1, ...]. m The middle range is (0.1·I). m 0.6·I m The higher range is (0.6·I). m ,I m These threshold points are determined by conducting A / B tests on users of different skill levels to find the optimal balance point that effectively conveys teaching information without causing user aversion, and can be used as configurable parameters of the system.

[0100] When the final intervention intensity I f When the value is in a preset very low range, the rule set maps this intensity to a feedback unit that does not intervene at all, or only provides positive encouragement after the user has finished writing the entire character.

[0101] When the final intervention intensity I fWhen it is in a preset medium range, the rule set maps it to a feedback unit that provides macro-level feedback, such as highlighting or briefly prompting the structural problems or stroke combinations of the entire character through highlighting or short voice prompts.

[0102] When the final intervention intensity I f When it is in a preset higher range, the rule set maps it to a feedback unit that provides fine-grained feedback, such as correcting the angle, length or start and end of a specific stroke in real time through dynamic stroke guidance or local magnification.

[0103] This hierarchical mapping rule enables the system's teaching feedback to have a sense of hierarchy and relevance that matches the severity of the problem; it avoids excessive intervention in minor flaws and ensures accurate guidance for serious errors, thereby achieving effective allocation of teaching resources and minimizing the cognitive load on users.

[0104] Example 9

[0105] The mapping rule set also includes: when the cognitive load index is higher than a preset threshold, resulting in the suppression of the final intervention intensity, an intervention behavior unrelated to error correction and aimed at alleviating the user's emotions is executed.

[0106] Based on the implementation of Example 7, a special case included in the mapping rule set is further explained; the rule set further includes an intervention strategy aimed at proactively managing user state;

[0107] When the cognitive load index Λ exceeds the preset threshold Λ0... the system will trigger a special mapping rule. As an advanced implementation in this embodiment, to achieve a more targeted state care strategy, the user state monitoring module can also output a multi-dimensional state vector that can distinguish specific states, such as fatigue and frustration V = [v 疲劳 ,v 沮丧 At this point, the special mapping rule will be triggered based on the analysis results of the vector; the intervention behavior corresponding to this rule is not aimed at writing correction, but at alleviating the user's negative emotions and cognitive load; the specific intervention behavior may be playing a fun animation, switching to a more relaxing practice background music, or directly suggesting to the user to take a break in an anthropomorphic way;

[0108] This special rule enables the system to proactively manage user states. Its function is to switch from a teacher's role to a partner's role when a user's state is detected as unfavorable, shifting the intervention goal from correcting errors to adjusting the state. For example, the system analyzes the state vector V to discover fatigue v, which represents fatigue. 疲劳 When the value exceeds a threshold, an intervention suggesting rest may be triggered; while when the value representing frustration (v) is higher... 沮丧When the score exceeds a threshold, background music can be switched or more targeted encouragement can be given. This intervention logic based on multidimensional state analysis can more accurately respond to different negative states of users, fundamentally breaking the vicious cycle of declining state - strong intervention - worse state, ensuring the sustainability of teaching activities, and maintaining users' learning interest. Those skilled in the art will understand that simplifying complex user cognitive states into a single index Λ is to clearly illustrate the core idea of ​​this invention; for example, under the current single index model, although the system can identify that the user's state is not good, it is difficult to distinguish whether the state is physiological fatigue caused by long-term writing or psychological frustration caused by continuous failure. The former may only require a short rest, while the latter requires more targeted encouragement or a reduction in task difficulty; in a more advanced implementation, the user state monitoring module can also output a multidimensional state vector to distinguish different states such as fatigue and frustration, thereby realizing a more targeted state care strategy;

[0109] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention; 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.

[0110] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A deep learning-based Chinese character visual recognition training system, characterized in that, include: The handwriting recognition module is used to receive handwriting sequence data input by the user, and to process the handwriting sequence data to obtain the handwriting recognition confidence score. The user status monitoring module is used to synchronously receive handwriting sequence data and analyze the dynamic characteristic parameters of the handwriting sequence data to obtain the cognitive load index, which quantifies the user's cognitive load status. The decision arbitration module receives the handwriting recognition confidence score and cognitive load index, and modulates the intensity of teaching intervention nonlinearly based on the dynamic adjustment model to obtain the final intervention intensity. The final intervention intensity decouples the direct binding relationship between handwriting recognition confidence score and teaching intervention intensity. The dynamic intervention module is used to provide teaching feedback to users with intensity, granularity and frequency adapted according to the final intervention intensity. The specific processing steps of the user state monitoring module include: guiding the user to write preset benchmark Chinese characters, collecting and processing the handwriting dynamics feature data of the benchmark Chinese characters to obtain personalized benchmark state parameters; extracting the real-time dynamics features of the user's current handwriting sequence during training; comparing and calculating the real-time dynamics features with the benchmark state parameters to generate a cognitive load index. This comparison calculation can be completed by a preset asymmetric weighted model, with the formula: The dynamic characteristic data of the handwriting include average writing speed, average writing pressure, average pause time between strokes, and stroke trajectory tremor frequency. The specific processing procedure of the decision arbitration module includes: calculating the baseline intervention intensity reflecting the system's direct response to writing errors based on the handwriting recognition confidence level; calculating the empathy modulation factor used to dynamically suppress the baseline intervention intensity based on the cognitive load index; multiplying the baseline intervention intensity with the empathy modulation factor to obtain the final intervention intensity. The calculation process of the baseline intervention intensity is as follows: multiplying the preset maximum system intervention intensity constant by the Nth power of the difference between the handwriting recognition confidence level and 1, where N is a preset intervention sensitivity coefficient greater than 1. The calculation process of the empathy modulation factor is as follows: based on the sigmoid function, generating an inhibition factor that smoothly transitions near the cognitive load tolerance threshold according to the difference between the cognitive load index and the preset cognitive load tolerance threshold, the inhibition factor being the empathy modulation factor.

2. The deep learning-based Chinese character visual recognition training system according to claim 1, characterized in that, The dynamic intervention module has a pre-set set of mapping rules from intervention intensity to specific intervention behavior, which is used to map the final intervention intensity to discrete feedback units with different granularities and modalities.

3. The Chinese character visual recognition training system based on deep learning according to claim 2, characterized in that, The mapping rule set includes: when the final intervention intensity is in a preset very low range, no intervention or only positive encouragement is given; when the final intervention intensity is in a preset medium range, macro-level feedback is given; when the final intervention intensity is in a preset high range, fine-grained feedback is given.

4. The deep learning-based Chinese character visual recognition training system according to claim 2, characterized in that, The mapping rule set also includes: when the cognitive load index is higher than a preset threshold, resulting in the suppression of the final intervention intensity, an intervention behavior unrelated to error correction and aimed at alleviating the user's emotions is executed.

Images