Handwriting identification method

Abstract

The invention relates to the technical field of teaching, and discloses a handwriting recognition method, which comprises the following steps that S1, Chinese character writing specifications are dynamically displayed through AR, a user participates in practice in real time, and the practice is analyzed in real time; s2, a multi-sensory collaborative teaching practice module which synchronously integrates visual, auditory and tactile feedbacks; s3, forming a review strategy based on an intelligent memory enhancement mechanism of the user behavior data; s4, gamifying an interactive interface, and supporting real-time stroke detection and error correction; and S5, a personalized learning path generation module adapted to the learning progress and error mode of the user. According to the method, the writing track of the user is analyzed in real time through the double-branch convolutional neural network, three-level error classification is output, and AR dynamic feedback is combined, so that the stroke sequence detection rate is 92%, the font analysis delay is less than 20 ms, and the error correction efficiency is improved; the multi-sensory collaboration module dynamically distributes visual, auditory and tactile weights by using an attention mechanism to adapt to different learning preferences, so that the memory retention rate is improved, and the environment adaptive capacity is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of teaching technology, and specifically to a method for handwriting recognition. Background Art

[0002] In the fields of intelligent education and digital teaching technology, it has become an important research direction to use technologies such as computer vision and multimodal interaction to improve the intelligent level of Chinese character writing teaching. Existing handwriting recognition and teaching technologies mainly collect writing trajectories through cameras or pressure sensors, and combine traditional machine learning algorithms or single neural network models to achieve writing standard detection, but there are significant technical bottlenecks in terms of real-time performance, accuracy, and personalized adaptation.

[0003] Specifically, existing technical solutions mainly use a single-branch convolutional neural network (CNN) to analyze writing trajectories, and can only extract two-dimensional features of glyph structures, and it is difficult to simultaneously consider the temporal dependence of stroke order and the spatial feature analysis of glyph structures. For example, when detecting Chinese characters with complex stroke orders or symmetrical structures such as "yong" and "ban", traditional algorithms cannot effectively distinguish between problems of incorrect stroke writing order and unbalanced structural proportions, and can only output a binary classification result of "correct / incorrect", and the lack of accuracy in error classification leads to the system being unable to generate targeted error correction strategies.

[0004] The inventors of the present application found during the implementation of the technical solution of the present application that the lack of the ability to jointly analyze the temporal and spatial features of writing trajectories limits the real-time performance and accuracy of stroke order detection and glyph structure analysis, and the lack of a dynamic adaptive adjustment mechanism for multimodal feedback makes it difficult to meet the needs of personalized teaching. Summary of the Invention

[0005] In view of the deficiencies of the prior art, the present invention provides a method for handwriting recognition, which solves the problem that the ability to jointly analyze the temporal and spatial features of writing trajectories is insufficient and it is difficult to meet the needs of personalized teaching.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for handwriting recognition, comprising the following steps: S1. Dynamically display Chinese character writing norms through AR, allow users to participate in practice in real time, and perform real-time analysis on the practice; S2. A multi-sensory collaborative teaching practice module that synchronously integrates visual, auditory, and tactile feedback; S3. An intelligent memory reinforcement mechanism based on user behavior data to form a review strategy; S4. A gamified interaction interface that supports real-time stroke detection and error correction; S5. A personalized learning path generation module that adapts to the user's learning progress and error patterns.

[0007] By adopting the above technical solutions, AR dynamically displays the Chinese character writing standards and combines the dual-branch convolutional neural network to perform spatiotemporal feature decoupling analysis of the writing trajectory, achieving a stroke order detection rate of 92%, a character structure analysis delay of less than 20ms, and a three-level error classification, providing support for differentiated error correction and improving error correction efficiency; the multi-sensory collaborative teaching module uses the attention mechanism to dynamically allocate the weights of visual, auditory, and tactile signals, which increases the memory retention rate by 35% and enhances the environmental robustness, adapting to different learning preferences and improving the memory retention rate.

[0008] Preferably, when the AR displays the Chinese character writing standards, a dual-branch convolutional neural network is used to perform real-time analysis of the user's writing trajectory, wherein the dual-branch structure respectively detects the stroke order and the character structure, outputs the error classification results, and forms a hierarchical error classification system.

[0009] Preferably, the intelligent memory enhancement mechanism uses a dynamic time warping algorithm to determine the accuracy and standardization of the user's writing, combines the forgetting probability function to predict the review time point, and recommends exercises. The forgetting function probability density function is as follows: , (k>0,b>0,a>1), t is the time interval, k,α,b are dynamic adjustment parameters.

[0010] Preferably, in the gamified interactive interface, the Hungarian algorithm is used to optimize the correspondence between the drag strokes and the standard template. If a sequence or position deviation is detected, a correction path is generated through the back propagation algorithm.

[0011] Preferably, the personalized learning path generation module adopts a reinforcement learning algorithm with a reward function: Dynamically adjust the training content weight, including and To adjust the hyperparameters.

[0012] Preferably, the multi-sensory collaborative teaching module allocates weights of visual, auditory, and tactile signals through an attention mechanism, and the calculation formula is: in Score the importance of each modal feature.

[0013] Preferably, a knowledge graph construction module is also included, which extracts the semantic features of Chinese characters through the BERT model, establishes a three-dimensional vector relationship between form, sound and meaning using the TransE algorithm, and screens high-frequency related nodes based on the PageRank algorithm for content recommendation.

[0014] Preferably, the real-time stroke detection uses a Gaussian kernel density estimation algorithm to generate a writing trajectory heat map, and the formula is: in is the coordinate point, is the smoothing parameter.

[0015] Preferably, the method is deployed in an edge computing framework, and the cloud-based pre-trained model is compressed into a lightweight version through knowledge distillation, with local inference latency less than 50ms.

[0016] Preferably, it also includes an anomaly detection module that uses an isolation forest algorithm to identify sudden error patterns and trigger targeted error correction strategies.

[0017] The present invention provides a method for handwriting recognition, which has the following beneficial effects: 1. The present invention uses a dual-branch convolutional neural network to analyze the user's writing trajectory in real time and output a three-level error classification. Combined with AR dynamic feedback, it achieves a stroke sequence detection rate of 92%, a character shape analysis delay of less than 20ms, and improved error correction efficiency; the multi-sensory collaboration module uses the attention mechanism to dynamically allocate visual, auditory, and tactile weights to adapt to different learning preferences, improve memory retention, and enhance environmental adaptability.

[0018] 2. The present invention uses a dynamic time warping algorithm and a forgetting probability function to predict review nodes, and links the knowledge graph to build a systematic memory network, which increases the 7-day memory retention rate from 62% to 81% and reduces the error rate of easily misspelled words by 47%; the personalized path module based on reinforcement learning uses a three-stage course framework and dynamic weight adjustment to improve learning efficiency by 35%, reduce the number of consecutive errors by 63%, and achieve customized practice and system stability.

[0019] 3. This invention combines the Hungarian algorithm with the back-propagation algorithm to achieve precise stroke matching and smooth error correction, with a processing delay of less than 16ms. The gamification design increases participation by 58%. It adopts a knowledge distillation compression model and deploys it on the edge computing framework. The model volume is compressed by 98.2%, the inference delay is less than 50ms, and it supports offline operation and privacy protection, significantly improving real-time performance and reliability.

[0020] 4. This invention uses BERT, TransE, and PageRank to construct a Chinese character knowledge graph, reducing the error rate of confused characters by 27%. The isolation forest algorithm realizes real-time detection of three types of anomalies with a response speed of 4.2 seconds and reduces the number of consecutive errors by 63%, forming a "detection-intervention" closed loop. Combined with multi-module linkage, dynamic adjustment of personalized error correction strategies can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 A flow chart of a handwriting recognition method of the present invention; DETAILED DESCRIPTION

[0022] The following will clearly and completely describe the technical solution of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0023] Please see the attached Figure 1 , an embodiment of the present invention provides a method for handwriting recognition, comprising the following steps: S1. Use AR to dynamically display Chinese character writing standards, allowing users to participate in real-time practice and conduct real-time analysis of the practice; S2, a multi-sensory collaborative teaching practice module, which simultaneously integrates visual, auditory and tactile feedback; S3, intelligent memory reinforcement mechanism based on user behavior data to form review strategies; S4, gamified interactive interface, supports real-time stroke detection and error correction; S5. Personalized learning path generation module, adapted to user learning progress and error patterns.

[0024] Specifically, S1. Dynamically display Chinese character writing standards through augmented reality (AR): AR is used to intuitively demonstrate the stroke order of Chinese characters to users, clearly showing the starting, running, and ending positions of each stroke, allowing users to grasp the order of writing. The system displays Chinese character writing standards, including details such as stroke length, inclination angle, and spacing, to help users form correct writing cognition. A dedicated mini-theater is set up for each Chinese character, showing the composition and writing process of the Chinese character in the form of animation, allowing users to master the writing method while watching interesting animations. During the demonstration, users participate in the practice, and relevant practice data is collected and analyzed in real time.

[0025] S2. Multi-sensory collaborative teaching module, which simultaneously integrates visual, auditory and tactile feedback: Vision: Presents the stroke order, structural form and animation scenes of Chinese characters, allowing users to observe the writing process and overall appearance of Chinese characters with their eyes; Auditory: Provides pronunciation guidance for Chinese characters. At the same time, during the writing process, corresponding voice prompts are given for correct or incorrect operations, such as encouraging voice when writing correctly and corrective prompt voice when writing incorrectly; Tactile: Users can feel the brushstrokes and strength when writing Chinese characters by actually writing, achieve the tactile experience of manual writing, and deepen the impression of Chinese character writing.

[0026] S3. An intelligent memory enhancement mechanism based on user behavior data uses a dynamic time warping algorithm combined with the forgetting law of the Ebbinghaus forgetting curve. Based on the user's learning behavior data, such as learning interval time, number of exercises, accuracy of answering questions, and standardization of writing, a scientific review mechanism is designed to determine the frequency of word recurrence. The system automatically collects and analyzes the user's data in writing practice, identifies the user's easily mistaken and easily confused characters, and pushes these easily mistaken and easily confused difficult characters to the user in a timely manner for review and practice at the time when the memory weakens, ensuring that the user forms long-term learning memory.

[0027] S4. The gamified interactive interface supports real-time stroke detection and error correction, and adopts a fun and gamified practice format. For example, by dragging the correct strokes to form a complete Chinese character, users can learn to write through hands-on operation and avoid boring mechanical practice. While the user is dragging the strokes, the system performs real-time stroke detection to determine whether the stroke order and position are correct. If the stroke order or position is correct, the interface will give positive prompts, such as "Great!" "Correct", etc.; if it is wrong, it will give appropriate guidance, such as the prompt "The position of this stroke is wrong, try again", to help users correct the error in time.

[0028] S5. Personalized learning path generation module adapts to the user's learning progress and error patterns. The system continuously tracks the user's learning progress and records the user's mastery of each Chinese character, including writing proficiency, error types and other data, and generates a customized writing practice path for the user. For Chinese characters that have been mastered, the frequency of practice is appropriately reduced; for Chinese characters that have not been fully mastered, the amount of practice is increased. Based on the user's error patterns, the easily misspelled and easily confused characters are classified, and the corresponding practice content is pushed to the user in a targeted manner to help the user strengthen weak links and improve writing skills.

[0029] 1. When AR displays standard Chinese character writing, users practice in real time. A convolutional neural network analyzes the user's writing trajectory in real time, detects the stroke order and glyph structure through a dual-branch structure, and outputs the error classification results.

[0030] Specifically, when displaying standard Chinese character writing in AR, a two-branch convolutional neural network architecture performs real-time analysis of the user's writing trajectory. This network consists of a stroke order detection branch and a glyph structure analysis branch. It employs five layers of depthwise separable convolution with batch normalization and LeakyReLU activation. After the third convolutional layer, it bifurcates into two dedicated detection branches. The stroke order detection branch processes the spatiotemporal sequence data of the writing trajectory using 1D temporal convolution. The output layer uses a Softmax classifier to predict the probability of correct stroke order and sets a dynamic threshold mechanism to trigger real-time error correction prompts when the confidence level falls below 0.7. The glyph structure analysis branch processes the two-dimensional feature map, employing a spatial attention mechanism to focus on key stroke areas. The output includes stroke position offset, morphological similarity, and an overall structural balance score. The outputs of the two branches are weighted and integrated through a feature fusion layer to establish a multi-level error classification system that includes Class I (severe), Class II (moderate), and Class III (minor) errors. Differentiated AR correction prompts are generated based on the error level.

[0031] 2. The multi-sensory collaborative teaching module allocates the weights of visual, auditory, and tactile signals through the attention mechanism. The calculation formula is: in Score the importance of each modal feature.

[0032] Specifically, the visual channel uses a lightweight convolutional neural network to process eye tracking data. By calculating a two-dimensional spatial attention distribution matrix and comprehensively analyzing the user's gaze intensity and pupil changes, it derives a visual importance score. This score is composed of a weighted combination of the interface's gaze coverage and the variance of pupil changes, accurately reflecting the user's visual attention level.

[0033] The auditory channel analyzes speech features based on Mel-frequency cepstral coefficients and constructs an acoustic attention vector through voice activity detection and response latency measurement. The auditory importance score comprehensively considers the user's voice interaction frequency and response speed to accurately assess auditory engagement.

[0034] The tactile channel uses data from the pressure sensor array to analyze writing force distribution and trajectory accuracy. The tactile importance score combines two key indicators: pressure uniformity and writing error, to objectively quantify the quality of the user's tactile engagement.

[0035] The system uses an improved attention weight calculation model, normalizing the basic importance scores of each sensory channel through a softmax function. It also introduces an environmental adaptation correction term to automatically compensate for the interference effects of varying environmental conditions on each sensory channel. The final weight calculation uses a reliability-weighted approach to achieve a balance between the basic attention weight and the environmental correction term.

[0036] In terms of multimodal fusion control, the system achieves refined output adjustment: the visual channel dynamically adjusts the transparency of the AR display and the animation playback speed according to the weight to ensure the best visual effect; the auditory channel adaptively adjusts the voice broadcast rate and volume gain to optimize the auditory experience; the tactile channel accurately controls the vibration feedback intensity and response delay to provide appropriate tactile prompts.

[0037] 3. The intelligent memory enhancement mechanism matches user trajectories with standard templates through the Dynamic Time Warping (DTW) algorithm and combines it with the forgetting probability function: Predict review time points, where t is the time interval and k, α, and b are dynamic adjustment parameters.

[0038] Specifically, the user and standard trajectories are represented as multidimensional sequences containing coordinates, pressure, and timestamps. The dynamic time warping (DTW) algorithm is used to calculate the exercise standardization and accuracy using weighted Euclidean distance. The computational complexity is reduced by using a curved window constraint and a symmetric penalty strategy. The curved window constraint sets a fixed maximum offset r, forcing the path to search within a strip area near the diagonal. The symmetric penalty strategy adds a penalty term λ to prevent asymmetric excessive bending.

[0039] At the same time, a probability density function is constructed based on the Ebbinghaus forgetting curve, where k decays as the number of exercises increases. The system dynamically adjusts according to the irregularity of practice and the error rate, and b increases with the improvement of the recent accuracy rate. When P(t)<0.6, a review is triggered, and the priority is determined by the probability of forgetting, the difficulty of Chinese characters, and the recent accuracy rate. The mechanism adopts a layered review strategy (primary multi-sensory guidance, intermediate gamification interaction, and advanced rapid review) and combines the knowledge graph to associate the same radical / homophone for collaborative review. At the same time, it links the multi-sensory teaching module to adjust the feedback weight and output the review requirements to the personalized learning path module. Through edge computing, offline model deployment and low-latency reasoning are realized. After testing, the user's 7-day memory retention rate increased from 62% to 81%, the long-term error rate of easily misspelled characters decreased by 47%, and the review efficiency was significantly optimized.

[0040] 4. In the gamified interactive interface, the Hungarian algorithm is used to optimize the correspondence between dragged strokes and standard templates. If order or position deviation is detected, a corrective path is generated through the back-propagation algorithm.

[0041] Specifically, the system achieves precise stroke matching through the deep fusion of the Hungarian algorithm and a neural network. The system constructs a bipartite graph model that combines user-dragged stroke elements with standard template strokes. It uses the composite weight calculation formula Wij = α·Dpos+β·Dshape+γ·Dtime (where Dpos is the Euclidean distance of position, Dshape is the DTW distance of shape, Dtime is the operation time difference, and α, β, and γ are adaptive weight parameters) to achieve optimal correspondence matching. The real-time monitoring system simultaneously tracks three core metrics for optimal correspondence matching: order deviation (SO), position deviation (PD), and morphological difference (SD). A three-level warning mechanism is implemented, including visual cues, voice reminders, and forced corrections. When severe deviations (SO ≥ 0.5, PD ≥ 30px, or SD ≥ 0.6) are detected, an intelligent error correction path is generated based on the constrained backpropagation algorithm.

[0042] The error correction path generation process uses the error function E=λ1·SO+λ2·PD+λ3·SD, and calculates the gradient A momentum term, ΔP(t) = η·∇E + μ·ΔP(t-1), is added (with η = 0.01 learning rate and μ = 0.9 momentum coefficient). This ultimately generates a smooth correction path optimized using cubic spline interpolation, with the single-step correction amplitude strictly limited to 20px. The system is equipped with a comprehensive gamified feedback system, including green highlight particles to mark correct strokes, red pulsating arrows to indicate deviating strokes, and a reward system with a base score of 100×(1-min(1,E)) and combo bonuses. Parameters such as the number of strokes (5-15), the allowable error range (±10-±30px), and the operation time limit (10-30 seconds) are dynamically adjusted based on user performance.

[0043] This solution utilizes optimization measures such as JIT compilation acceleration, incremental computing, and GPU acceleration to ensure the system maintains high real-time performance with an average processing latency of less than 16ms, while significantly improving user operation accuracy by 42% and learning efficiency by 35%. The system's innovative dynamic difficulty adjustment mechanism and rich gamified feedback design significantly enhance the user experience while ensuring educational effectiveness, creating a fun and engaging interactive experience. A composite weight calculation model and three-level early warning mechanism enable the system to accurately identify various handwriting deviations and guide users to naturally improve their handwriting through a smooth and gradual correction path.

[0044] 5. Personalized learning: Push relevant practice content to users in a targeted manner The personalized learning path generation module adopts the reinforcement learning algorithm with the reward function: Dynamically adjust the training content weight, including and To adjust the hyperparameters.

[0045] Specifically, the personalized learning path generation module adopts a deep reinforcement learning framework and realizes intelligent teaching strategy optimization through the Actor-Critic dual network structure. The system constructs a multi-dimensional state space S, integrating key features such as user portraits (age, learning style), real-time learning data (accuracy, time consumption, error patterns) and historical performance (memory curve, progress trend); the action space A includes decision outputs in three dimensions: difficulty level of practice content (L1-L5), multimodal teaching strategy (visual / auditory / tactile dominated) and dynamic practice volume (5-20 Chinese characters / time). The core reward function adopts an improved composite design R=ω1·(1-e^(-λ1·Accuracy))+ω2·tanh(λ2·Progress)-ω3·(ErrorCount)^(1 / 3)+ω4·EngagementScore, where ω1-ω4 are dynamic weight parameters (initial values ​​0.4, 0.3, 0.2, 0.1), λ1 and λ2 are adjustable shape parameters (default 1.5 and 0.8), and the EngagementScore is calculated in real time using behavioral data such as eye tracking and operation frequency.

[0046] In terms of system implementation, the proximal policy optimization (PPO) algorithm is used for strategy update, and an innovative three-stage course learning framework is designed: the initial stage (ω1=0.6) focuses on accuracy training, the middle stage (ω2=0.5) accelerates the learning progress, and the late stage (ω3=0.4) focuses on correcting stubborn errors.

[0047] pass Automatic hyperparameter optimization is implemented, coupled with a weekly meta-learning update mechanism based on user population data, to continuously optimize key parameters such as λ1 and λ2. To ensure system stability, multiple safeguards are implemented: a policy reset is automatically triggered when the reward value R decreases by more than 10% three times in a row; safety constraints are established for automatic downgrades if the difficulty level increases by ≤ 2 and the error rate exceeds 40%; and a visual dashboard is provided to monitor ω coefficient changes and influencing factors in real time.

[0048] 6. Construct three common modules to assist teaching 6.1 Knowledge Graph Construction Module The semantic features of Chinese characters are extracted through the BERT model, and the three-dimensional vector relationship between form, sound, and meaning is established using the TransE algorithm. High-frequency related nodes are filtered based on the PageRank algorithm for content recommendation.

[0049] Specifically, 1. Semantic feature extraction of Chinese characters: Optimized application of the BERT model: Using the Chinese-BERT-wwm pre-trained model, Chinese characters are converted into Unicode encoding to generate initial word vectors, and semantic dependencies are captured through a 12-layer Transformer encoder to output 768-dimensional dynamic word vectors; a dictionary enhancement mechanism is introduced to distinguish polysemous word meanings, and combined with fine-tuning of educational corpus, feature extraction of commonly misspelled and confused characters is optimized.

[0050] Three-dimensional shape-sound-meaning vector modeling: an innovative extension of the TransE algorithm: defining an entity set including Chinese character ontology, glyph features, pinyin, and semantic categories, as well as shape-meaning, sound-meaning, and shape-sound association relationships; extracting glyph and pinyin vectors through CNN and FastText respectively, and using BERT to output semantic vectors and reduce dimensionality; improving the TransE algorithm, modeling with three-dimensional relationship constraints, and using Margin-BasedRankingLoss optimization to force the distance of positive samples to be smaller than that of negative samples.

[0051] Screening of high-frequency associated nodes: Adaptation of the PageRank algorithm to educational scenarios: Constructing a knowledge graph containing Chinese characters, glyph components, pinyin, and semantic tags, defining edge weights based on co-occurrence frequency, confusion probability, and semantic similarity; injecting personalized weights into the associated edges of easily misspelled characters based on user error data, recommending homophones, homoradicals, and frequently confused objects based on PageRank scores, and dynamically adjusting recommended content according to learning progress.

[0052] System integration and application effects: The knowledge graph recommendation results are linked to personalized learning paths, multi-sensory teaching and intelligent memory enhancement modules to provide semantic-level support; after testing, the accuracy of Chinese character association recommendations increased by 38%, the error rate of user confused characters decreased by 27%, and learning efficiency increased by 22%, realizing an upgrade from isolated character learning to systematic knowledge network construction.

[0053] 6.2 Real-time Stroke Detection Module The Gaussian kernel density estimation algorithm is used to generate the writing trajectory heat map. The formula is: in is the coordinate point, is the smoothing parameter.

[0054] Specifically, is the coordinate point, is the smoothing parameter, by adjusting It can focus on local details or capture overall trends. When implementing, it first performs noise filtering and time synchronization preprocessing on the trajectory data, then generates a heat map through block calculation and GPU acceleration, and uses threshold segmentation to extract the main strokes, calculates the center of mass to evaluate the deviation, and identifies stroke order, position and shape problems. The linkage with the AR module can superimpose the heat map to highlight the deviation area and trigger the correction animation. Combined with multimodal teaching, it can provide feedback through tactile vibration and auditory voice, and can also store abnormal trajectories in the error library to assist intelligent memory reinforcement review, while dynamically optimizing through reinforcement learning. Parameters, to achieve single stroke analysis delay <20ms, stroke order detection rate of 92% real-time accurate detection, and support multi-user concurrency and adaptation of different writing styles. Combined with knowledge distillation, lightweight deployment on edge devices, it converts writing trajectories into intuitive thermal features, providing key data support for real-time error correction and personalized teaching.

[0055] 6.3 Anomaly Detection Module The isolation forest algorithm is used to identify sudden error patterns and trigger targeted error correction strategies.

[0056] Specifically, the anomaly detection module uses the isolation forest algorithm to construct a binary tree structure to randomly divide the feature space. When the score exceeds the threshold of 0.7, it is judged as an anomaly. The module constructs a multi-dimensional feature system including time series, pattern, trajectory, and context. After normalization, discretization and sliding window preprocessing, it realizes real-time detection through incremental online learning. It can identify types such as sudden accuracy drop (Type I), high-frequency repeated errors (Type II), and trajectory feature abnormalities (Type III), triggering a three-level response system including early warning (AR highlight + voice prompt), intermediate (multi-sensory reinforcement + wrong question collection), and emergency (forced pause + teacher warning). It also dynamically optimizes the strategy weights based on reinforcement learning, and works with the knowledge graph and edge computing module to achieve anomaly detection delay of less than 30ms, recognition rate of 91%, and response speed of 4.2 seconds, reducing the number of consecutive errors by 63% and the recurrence rate of similar errors by 51%, forming a closed-loop error correction link from detection to intervention.

[0057] 7. Method deployment The method is deployed on an edge computing framework and compresses the cloud-based pre-trained model into a lightweight version through knowledge distillation, with local inference latency less than 50ms.

[0058] Specifically, the method is deployed in an edge computing framework and designed with a hierarchical architecture. The device layer integrates intelligent writing devices to collect data. The edge management layer uses KubeEdge to achieve cluster management and message queue communication, and encapsulates the handwriting recognition logic into a Docker container. A "teacher-student" model architecture is constructed through knowledge distillation. The teacher model is ResNet50 + BiLSTM + 12-layer BERT, and the student model uses MobileNetV3 + TinyBERT. Combining a hybrid distillation loss function that includes task loss, knowledge distillation loss, and feature transfer loss, after compression such as quantization, pruning, and weight sharing, the number of parameters is reduced to 3.2 million, and the volume < 15MB. On the edge device side, heterogeneous computing of CPU + GPU + DSP, Winograd convolution algorithm, TensorRT layer fusion, etc. are used for optimization, achieving 18ms for stroke order detection and 22ms for glyph structure analysis in single-model inference, with an end-to-end latency < 50ms. And through data localization processing, privacy is guaranteed, and offline operation is supported. Compared with the model deployed in the cloud, the model compression is 98.2%, the power consumption is reduced by 95.4%, the latency is reduced by > 60%, and the accuracy loss is only 1.4%, significantly improving the real-time performance and reliability of the system. Example 1

[0059] Dynamic display of Chinese character writing norms based on AR 1. Use AR to display standardized strokes (such as the animation of the eight methods of the character 'yong'), and users can follow up in real time; 2. Capture the user's handwriting in real time, obtain the coordinates, and convert them into images; Example 1

[0059] Dynamic display of Chinese character writing norms based on AR 1. Use AR to display standardized strokes (such as the animation of the eight methods of the character 'yong'), and users can follow up in real time; 2. Capture the user's handwriting in real time, obtain the coordinates, and convert them into images; 3. Project the converted images into a two-branch convolutional neural network for analysis; 4. Network architecture: 1) Before bifurcation: The first 3 layers are shared depthwise separable convolutions (reducing the amount of computation), the activation function is LeakyReLU, and batch normalization is performed; 2) After bifurcation: Among them, the temporal branch uses 1D temporal convolution to process the stroke order, uses a softmax layer to predict the probability of each stroke, and takes the maximum probability as the probability of the correct handwriting; the spatial branch uses 2D convolution to process the two-dimensional feature map, uses an attention mechanism to analyze the structure, and outputs the stroke position offset, morphological similarity, and overall structure balance index score; 3) Create a feature fusion layer based on the stroke order correct probability output by the two branches and the scores of the three indicators for weighted integration, and establish a hierarchical error classification system including three types of errors: type I (severe) (similarity < 0.3), type II (medium) (similarity < 0.6), and type III (slight) (similarity ≥ 0.6).

[0060] 5. Dynamic Error Correction: Based on the probabilities output by the softmax layer, the highest probability is called the stroke accuracy confidence. If the stroke order confidence is less than 0.7, a stroke order correction prompt is triggered. If the stroke order confidence is ≥ 0.7, differentiated AR correction solutions are generated based on the error level. For Category I, the main focus is on mandatory structural display, for Category II, a red light warning is issued, and for Category III, only a text prompt is provided. Example 2

[0061] Attention weight allocation mechanism for multi-sensory collaborative teaching Technical implementation: Integration of visual, auditory, and tactile signals: The visual channel generates a score by calculating gaze intensity through eye tracking , the auditory channel generates a score based on the frequency of voice interaction The tactile channel generates a score by analyzing the uniformity of force through the pressure sensor , using the softmax formula Dynamically assign weights to each modality. If the visual weight is the highest, the clarity of the AR animation is enhanced. If the tactile weight is low, the intensity of the pen-tip vibration feedback is increased, so that multimodal feedback can be dynamically adjusted according to the user's attention. Example 3

[0062] Intelligent memory enhancement based on dynamic time warping and forgetting function The user and standard trajectories are represented as multidimensional sequences containing coordinates, pressure, and timestamps. The dynamic time warping (DTW) algorithm is used to calculate the matching cost using weighted Euclidean distance, and the computational complexity is reduced by using a curved window constraint and a symmetric penalty strategy. The curved window constraint sets a fixed maximum offset r, forcing the path to search within a strip area near the diagonal. The symmetric penalty strategy adds a penalty term λ to prevent asymmetric excessive bending.

[0063] 1. Calculate the distance between the user's practice trajectory and the standard trajectory.

[0064] (1) Constructing a distance matrix Assume that the two time series of user and standard trajectories are , , i=1,2,3 I, Represents the coordinates, pressure, and timestamp of the T time series respectively

[0065]

[0066] Calculate the weighted Euclidean distance including position (coordinates), shape (pressure), and time (timestamp):

[0067] (2) Initialize the cumulative distance matrix Construct the matrix DP, where DP( ) represents the minimum cumulative distance from the starting point to ( ):

[0068] (3) Fill in by dynamic programming Recursive formula:

[0069] (4) Backtrack the optimal path Backtrack from DP( ) and select the path DI that minimizes the cumulative distance.

[0070] And set the bending window constraint to set a fixed maximum offset r, forcing the path to search within a strip area near the diagonal; the symmetric penalty strategy adds a penalty term λ to prevent asymmetric over-bending. The specific settings are as follows: Set a fixed maximum offset r, forcing the path to search within a strip area near the diagonal ∣ ∣ ≤ r In the dynamic programming recursive formula, add a penalty term λ for non-diagonal moves (such as horizontal or vertical moves):

[0071] 2. Combine the forgetting probability function: (k decays as the number of practice times increases, α decreases as the error rate and non-standard rate increase, and b increases as the recent correct rate increases) Predict the review nodes, and obtain the standard rate and correct rate by combining the difference between the user's practice trajectory and the actual trajectory. If DI is less than a fixed threshold (this fixed threshold is the system default setting), it is considered correct and standard. Otherwise, it is considered wrong and non-standard.

[0072] 3. According to different P values, formulate a hierarchical review strategy. When P(t) < 0.6, push error-prone characters, and adopt a hierarchical review strategy of "primary multi-sensory guidance + intermediate gamified jigsaw puzzles + advanced time-limited memorization", and联动 the knowledge graph to recommend characters with the same radical (such as "清 / 情 / 请") for collaborative memory, improving the 7-day memory retention rate to 81%.

[0073] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for handwriting recognition, characterized in that: The following steps are involved: S1. Use AR to dynamically display Chinese character writing standards, allowing users to participate in real-time practice and conduct real-time analysis of the practice; S2, a multi-sensory collaborative teaching practice module, which simultaneously integrates visual, auditory and tactile feedback; S3, intelligent memory reinforcement mechanism based on user behavior data to form review strategies; S4, gamified interactive interface, supports real-time stroke detection and error correction; S5. Personalized learning path generation module, adapted to user learning progress and error patterns.

2. A handwriting recognition method according to claim 1, characterized in that: When the AR displays the Chinese character writing standards, a dual-branch convolutional neural network is used to perform real-time analysis of the user's writing trajectory. The dual-branch structure detects the stroke order and the character structure respectively, outputs the error classification results, and forms a hierarchical error classification system.

3. The handwriting recognition method according to claim 1, characterized in that: The intelligent memory enhancement mechanism uses a dynamic time warping algorithm to determine the accuracy and standardization of the user's writing, and combines it with the forgetting probability function to predict the review time point and recommend exercises. The forgetting function probability density function is as follows: , (k>0,b>0,a>1), t is the time interval, k,α,b are dynamic adjustment parameters.

4. The handwriting recognition method according to claim 1, characterized in that: In the gamified interactive interface, the Hungarian algorithm is used to optimize the correspondence between the dragged strokes and the standard template. If a sequence or position deviation is detected, a correction path is generated through the back propagation algorithm.

5. A handwriting recognition method according to claim 4, characterized in that: The personalized learning path generation module adopts a reinforcement learning algorithm with a reward function: Dynamically adjust the training content weight, where and are the tuning hyperparameters.

6. A handwriting recognition method according to claim 5, characterized in that: The multi-sensory collaborative teaching module allocates the weights of visual, auditory, and tactile signals through the attention mechanism. The calculation formula is: in Score the importance of each modal feature.

7. A handwriting recognition method according to claim 6, characterized in that: It also includes a knowledge graph construction module, which extracts the semantic features of Chinese characters through the BERT model, uses the TransE algorithm to establish a three-dimensional vector relationship between form, sound, and meaning, and filters high-frequency related nodes based on the PageRank algorithm for content recommendation.

8. The handwriting recognition method according to claim 1, characterized in that: The real-time stroke detection uses the Gaussian kernel density estimation algorithm to generate a writing trajectory heat map, and the formula is: in is the coordinate point, is the smoothing parameter.

9. The handwriting recognition method according to claim 1, characterized in that: The method is deployed in an edge computing framework and compresses the cloud-based pre-trained model into a lightweight version through knowledge distillation, with local inference latency less than 50ms.

10. The handwriting recognition method according to claim 1, characterized in that: It also includes an anomaly detection module that uses the isolation forest algorithm to identify sudden error patterns and trigger targeted error correction strategies.

Images