A calligraphy teaching method based on point back action disassembly and a stroke mnemonic system

Abstract

The application discloses a calligraphy teaching method based on point-back action disassembly and a stroke formula system, uniformly defines the strokes of regular script as four action phases of "falling, going, dotting and returning", establishes a 28-set standardized action formula system composed of 14 basic stroke formulas and 14 compound stroke formulas, all the formulas follow a unified "dot-back" action logic, auxiliary teaching is combined with music beat rhythm, the cognitive threshold of beginners is significantly reduced, timing data of writing sensors are collected, four-phase boundaries are automatically segmented through three-stage pipeline of differential threshold pre-segmentation, hidden Markov model fine segmentation and dynamic time warping verification, process indicators such as speed-pressure ratio are extracted for quantitative scoring, root cause tracing diagnosis is realized by fusing image character defect features and timing process features, accurate error correction feedback is generated, and graded formulas with adaptive cognitive load are dynamically pushed according to the skill state of learners.

Description

Technical Field

[0001] This invention relates to the field of calligraphy teaching technology, specifically to a calligraphy teaching method and stroke mnemonic system based on point-return action decomposition, belonging to the interdisciplinary application field of smart education and artificial intelligence technology. Background Technology

[0002] Calligraphy teaching has a long history, and traditional teaching methods mainly rely on textual theoretical descriptions, such as "slanted horizontal strokes," "concealed brush tip and reverse entry," and "pauses and breaks in the brushstroke." These methods depend on teachers' oral instruction and learners' self-experience. The above teaching model has the following inherent defects:

[0003] Firstly, the theoretical descriptions are abstract, making it difficult for young beginners to establish a concrete mapping relationship between "language descriptions" and "muscle movements." This results in a high cognitive threshold and a long learning cycle. Traditional calligraphy theory does not break down the stroke writing process into discrete action units with independent biomechanical characteristics, making it difficult for learners to clearly perceive "at which step and in what way to apply force," leading to difficulties in error correction and slow skill progress.

[0004] Secondly, existing intelligent calligraphy evaluation systems compare the final static stroke images with standard calligraphy models at the pixel level, and can only output a conclusive judgment that "the writing is not good". They cannot independently quantify and score each sub-action stage in the dynamic writing process, nor can they reverse the results of character shape defects to locate specific action error nodes. As a result, the problem of learners "knowing their mistakes but finding it difficult to correct them" has not been solved for a long time.

[0005] Third, existing calligraphy mnemonic rhymes are static summaries of experience, which do not differentiate between learners' age, skill level, and stage of internalization of movements. This causes cognitive overload for beginners and introduces redundant interference for experienced learners. There is a lack of an adaptive mechanism to dynamically adjust the granularity of the mnemonic rhymes based on the learner's current skill level.

[0006] In summary, existing technologies lack a calligraphy teaching method and system that integrates standardized decomposition of stroke movements, structured coding of movement formulas, intelligent quantitative evaluation of the writing process, and adaptive push of formulas. Summary of the Invention

[0007] This invention aims to solve the following three core technical problems: First, traditional calligraphy teaching lacks a standardized and operable action decomposition framework, resulting in high comprehension thresholds and low learning efficiency for beginners; second, intelligent calligraphy evaluation systems lack the process diagnostic ability to trace back from the final character shape defects to specific writing sub-action errors; and third, calligraphy mnemonic rhymes cannot adaptively adjust their granularity according to the learner's real-time skill status, making it difficult to achieve optimal auxiliary effects at different learning stages.

[0008] This invention is implemented as follows: It provides a calligraphy teaching method and stroke mnemonic system based on point-return action decomposition, and its core technical solution is as follows:

[0009] (I) Four-phase motion disassembly framework

[0010] The complete writing process of a single stroke in regular script is uniformly defined as four action phases: "falling," "moving," "dotting," and "returning." "Falling" is the initial stroke phase, where the pen tip contacts the paper from a state of being off the paper and applies initial pressure. "Moving" is the stroke phase, where the pen tip moves along the target direction while maintaining stable pressure. "Dotting" is the pause phase, where the pen tip speed rapidly decreases to near zero while simultaneously increasing downward pressure. "Returning" is the finishing stroke phase, where the pen tip gently lifts off the paper in a specific direction. These four phases form the unified basic framework for the action mnemonic coding and process scoring of this invention.

[0011] (II) A system of 28 sets of stroke-specific action formulas

[0012] Based on the four-phase framework, this invention establishes a system of 28 standardized action formulas consisting of 14 basic stroke action formulas and 14 compound stroke action formulas.

[0013] The mnemonic for the 14 basic stroke movements is shown in the table below:

[0014] Serial Number Stroke Name Action formula 1 Right point Fall back 2 Hanging needle vertical Click back to vertical tip 3 Hanging Dew Vertical Click back and vertical scroll 4 vertical hook Click back vertical hook 5 Long horizontal Long horizontal cut 6 short stroke Click back 7 carry Click to retrieve 8 Long stroke Dot return long stroke 9 Vertical stroke Dot return vertical stroke 10 Pressing down Press down and level out 11 hook Bend Hook 12 Horizontal hook Bend point push hook 13 slanted hook Point back diagonal hook 14 Vertical hook Dot return vertical hook

[0015] The 14 compound stroke action formulas are generated by combining the above basic strokes through the dot-back action. The combination rules of compound strokes are as follows: the basic strokes such as horizontal, vertical, left-falling, right-falling, and hooks are connected to the "falling" phase of the next stroke by embedding a return stroke at the end of their "dot" phase, forming compound strokes such as horizontal turn, vertical lift, left-falling dot, and left-falling turn. The action formulas of compound strokes are formed by sequentially splicing the formulas of each basic stroke in the writing order, with the "dot-back" action as the turning anchor point at the connection of the formulas. The above system makes the 28 sets of stroke formulas cover all the basic strokes of regular script, and all formulas follow the unified "dot-back" action logic, which has a high degree of standardization and reproducibility.

[0016] (III) Music rhythm and beat-assisted teaching mechanism

[0017] Each word in the stroke mnemonic is matched with a preset musical beat unit. Learners follow the rhythm of the beat to perform the corresponding writing force action in sequence. The musical beat provides external temporal constraints for the writing action, helps learners perceive the rhythm of force switching between different phases, lowers the understanding threshold, and accelerates the internalization of muscle memory.

[0018] (iv) Writing sensor data acquisition and four-phase automatic segmentation

[0019] The intelligent writing pen adopts an integrated six-axis inertial measurement unit and a pressure sensor. The pressure sensor has a resolution of no less than 2048 levels, and the inertial measurement unit has a sampling frequency of no less than 120Hz. It synchronously collects a time-series data stream consisting of the pen tip's three-dimensional coordinates, pen barrel pitch angle, pen barrel roll angle, pen tip contact pressure, and timestamp, and simultaneously collects a static image after the stroke is completed.

[0020] Automatic phase segmentation is implemented using a three-stage pipeline: differential threshold pre-segmentation, Hidden Markov Model fine segmentation, and dynamic time warping template alignment verification. In the first stage, the instantaneous velocity time series of the pen tip is calculated by first-order difference of the coordinate sequence, and the pressure time series is obtained by first-order difference of the pressure change rate sequence. Based on the physical prior threshold conditions of each phase, the candidate phase boundary point set is marked, and the time series is initially divided into four candidate intervals. In the second stage, a four-state Hidden Markov Model is constructed. Using the sliding window statistical features of velocity, pressure, and pitch angle as observation features, Viterbi decoding is performed on the local time series windows near the candidate boundaries, and the refined four-phase boundary timestamps are output. In the third stage, the dynamic time warping normalized distance of each phase time series segment is calculated with the pre-stored standard action template. When the normalized distance of any phase exceeds the preset threshold, the phase segment is fed back to the second stage to adjust the initial state probability and then decoded again until all phases pass the verification. Finally, a set of time series segments with phase labels is output.

[0021] (v) Process-based quantitative scoring

[0022] For the "dot" action phase, the ratio of the average pen tip speed to the average pen tip pressure within that phase is extracted as the speed-pressure ratio, denoted as . Defined as:

[0023]

[0024] in, This represents the average pen tip velocity within the "point" phase, expressed in millimeters per millisecond. The value represents the average pen tip pressure within the "point" phase, and is set to the normalized level, with a range of [value missing]. ; The speed-pressure ratio is a core process indicator for evaluating the quality of the pen-pausing and preparatory movements in standard writing. It should fall within the confidence interval determined by the standard template statistics.

[0025] For each phase, multi-dimensional features such as mean velocity, mean pressure, rate of change of tilt angle, and dynamic time-normalized distance are extracted. The pre-trained scoring regression model outputs the quantitative score of each phase, presenting a four-dimensional phase score vector to the learner.

[0026] (vi) Cross-modal root cause diagnosis

[0027] Extract the character shape defect feature vector from the final image of the strokes, concatenate it with the temporal process feature vector of each action phase to form a fused feature vector, input it into a pre-trained XGBoost multi-label classifier, output the error-causing weight probability distribution of each action phase, and determine the phase with the highest weight as the root cause node. Based on this, retrieve the corresponding natural language feedback template in the preset knowledge graph to generate accurate error correction text containing specific action correction suggestions.

[0028] (vii) Adaptive mnemonic generation based on skill state machine

[0029] Define a skill state machine comprising three states: "beginner," "intermediate," and "automated." Use writing timing smoothness and phase comprehensive achievement rate as the state transition driving indicators. Defined as:

[0030]

[0031] in, A higher value indicates a smoother writing motion; For a moment Pen tip speed; This is the average speed over the entire journey; This represents the total number of time-series sampling points.

[0032] Phase comprehensive compliance rate Defined as:

[0033]

[0034] in, ; This is an indicator function; it takes the value 1 if the condition inside the parentheses is true, and 0 otherwise. For the first Phase quantization score; The phase passing threshold is set to 60 points by default. These correspond to the four phases: landing, row, point, and return.

[0035] when and When three consecutive writing sessions exceed their respective upgrade thresholds, the skill status is upgraded to a higher level; when When two consecutive writing sessions fall below the downgrade threshold, the skill status is downgraded to a lower level. The "Beginner" status corresponds to pushing a fine-grained four-step mnemonic, the "Advanced" status corresponds to pushing a medium-grained merging mnemonic, and the "Automation" status corresponds to pushing a macro-fusion mnemonic. If the root cause diagnosis module detects a specific phase anomaly, a targeted enhancement mnemonic is added after the regular mnemonic to achieve personalized mnemonic enhancement.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] By unifying all strokes of regular script into the "dot and return" action logic framework, and covering all basic strokes with 28 standardized action formulas, calligraphy teaching content is greatly simplified and replicable. This significantly lowers the learning threshold for beginners and noticeably improves the speed of learning effectiveness. Combined with musical rhythm teaching, abstract calligraphy movements are transformed into perceptible rhythmic exertion, effectively aiding in muscle memory formation and reducing the cognitive load for young children and beginners. It achieves precise automatic segmentation and independent quantitative scoring of the four phases of the writing process: "fall, move, dot, return," upgrading calligraphy assessment from static result scoring to dynamic process scoring. Learners can clearly understand "where..." "One step, in what way, a deviation occurred"—establishing a root cause diagnosis system that traces back from the result of character shape defects to the specific abnormal action phase. This enables precise and interpretable process-based error correction feedback, improving the traditional teaching dilemma of "knowing mistakes but finding it difficult to correct them." The adaptive mnemonic generation mechanism based on the skill state machine can dynamically adjust the mnemonic granularity according to the learner's real-time skill state, accurately matching cognitive load at different learning stages and taking into account the differentiated needs of beginners and proficient learners. The 28 sets of action mnemonic systems are not limited by the practice grid, font, or Chinese character content, and can be adapted to any Chinese character writing. They have a high degree of standardization and have the foundation for widespread promotion and large-scale replication. Attached Figure Description

[0038] Figure 1 This is a diagram showing the overall architecture of the calligraphy teaching system of the present invention;

[0039] Figure 2 Here is a flowchart of the action phase segmentation algorithm;

[0040] Figure 3 A cross-modal mapping architecture diagram for root cause diagnosis;

[0041] Figure 4 Generate engine architecture diagram for adaptive mnemonic devices;

[0042] Figure 5 This is a complete business process diagram of the system. Detailed Implementation

[0043] To further understand the invention's content, features, and effects, the following embodiments are provided, and detailed descriptions are given in conjunction with the accompanying drawings.

[0044] The structure of the present invention will now be described in detail with reference to the accompanying drawings.

[0045] Implementation Method 1:

[0046] Complete teaching process for the basic stroke "right dot"

[0047] Taking the "right dot" stroke in regular script as an example, this paper illustrates the complete implementation process of the teaching method of the present invention.

[0048] The basic stroke action pithy formula corresponding to the "right dot" is "lowering-returning-lifting", wherein "lowering" corresponds to starting the stroke in the "landing" phase, applying force to the right downward after touching the paper; "returning" corresponds to returning the tip to the upper left after pausing the stroke to store momentum in the "dot" phase; "lifting" corresponds to gently lifting the brush off the paper in the "returning" phase. When teaching, the teacher corresponds the above three-step pithy formula to three-beat musical rhythm, and the learner recites "lowering—returning—lifting" along with the rhythm and completes the corresponding force-applying action synchronously.

[0049] After a learner holds an intelligent writing pen integrated with an inertial measurement unit and a pressure sensor to complete writing, the system collects the full time-series data stream, and obtains a pressure change rate sequence by performing first-order difference on the pressure time series. Based on the physical prior rules that: the "landing" phase corresponds to a sharp increase of pressure from zero with a relatively low writing speed, the "moving" phase corresponds to a stable speed at a medium-high level with the pressure change rate close to zero, the "dot" phase corresponds to a rapid decrease of speed close to zero while the pressure reaches a local peak, and the "returning" phase corresponds to a rapid drop of pressure from the peak until the pen leaves the paper, the system marks candidate phase boundaries. Then, a four-state hidden Markov model performs Viterbi decoding on a window near the candidate boundaries to output time stamps of refined phase boundaries. In the dynamic time warping template alignment verification stage, normalized distance calculation is performed between each phase segment and pre-stored standard phase templates of the standard "right dot", and a set of labeled time series segments is output after confirming that the segmentation is valid.

[0050] In the process quantitative scoring stage, the system calculates the velocity-to-pressure ratio for the "dot" phase : if the average velocity of a learner's "dot" phase is too high while the average pressure is insufficient, resulting in the value exceeding the upper limit of the standard confidence interval, the system diagnoses that the learner has an action deviation of "too fast pause speed and insufficient momentum storage pressure". After the cross-modal root cause diagnosis module fuses the stroke image glyph defect features and time-series process features and confirms the root cause node is the "dot" phase, it retrieves the corresponding feedback template from the knowledge graph and generates accurate error correction text, prompting the learner to slow down the pause speed, increase the downward pressing force when reciting the word "returning" before starting to return the brush tip.

[0051] Embodiment 2:

[0052] Action pithy formula combination and teaching for the compound stroke "horizontal fold"

[0053] The "horizontal stroke with a turn" is a compound stroke, composed of two basic strokes: a "long horizontal stroke" and a "vertical stroke". The basic mnemonic for the "long horizontal stroke" is "long horizontal stroke ends", and the basic mnemonic for the "vertical stroke with a drooping tip" is "dot returns to vertical stroke ends". According to the rules of compound stroke combination, the action mnemonic for the "horizontal stroke with a turn" is "long horizontal stroke ends, dot returns to vertical stroke ends". After completing the horizontal stroke with the "long horizontal stroke ends", the "dot" phase is used as the turning point, and then the "dot returns to vertical stroke ends" is connected to complete the vertical stroke and the ending stroke. The teacher will match each word of the compound mnemonic with the continuous musical beat. The learner will follow the beat to complete all the action phases in sequence. There is no need to memorize the turning rules. The learner can complete the writing of the compound stroke by relying only on the two basic stroke mnemonics that they have mastered.

[0054] Implementation Method 3:

[0055] Skill state machine driven adaptive mnemonic push

[0056] During a learner's continuous writing sessions, the system calculates the smoothness of the writing sequence after each session. Phase-based overall compliance rate The skill state machine is updated, with the initial state being "Beginner". The system pushes a fine-grained four-step mnemonic, such as pushing a complete mnemonic of "dot-return-vertical-hook" for "vertical hook", corresponding to a four-beat musical rhythm. When three consecutive conversations meet the requirements... Exceeding the smoothness upgrade threshold and When the pass rate threshold is exceeded, the status upgrades to "Advanced," and the system pushes a granular merging mnemonic, such as merging "point-back" and pushing the two-word mnemonic "point-back-vertical hook." When three consecutive sessions meet the upgrade conditions, the status upgrades to "Automation," and the system pushes a macro-level fusion mnemonic, such as only pushing "one pause, one return," triggering the established muscle memory for automated execution with the simplest information. If two consecutive sessions in the "Advanced" stage... If the status falls below the downgrade threshold, the system will automatically downgrade the status back to "beginner" and resume fine-grained mnemonic push to prevent learning stagnation caused by skill decline. If the root cause diagnosis module detects a specific phase abnormality in any status, the system will add targeted reinforcement mnemonic after the regular mnemonic to achieve personalized mnemonic enhancement.

[0057] Implementation Method Four:

[0058] Lightweight offline deployment implementation method

[0059] For application scenarios with limited network connectivity, a knowledge distillation technique is used to compress the Hidden Markov Model parameter matrix and XGBoost tree model into a lightweight decision rule set, which is deployed on a local embedded processor. The mnemonic generation module stores a limited number of mnemonic templates in a relational database, and a rule engine replaces graph database retrieval to complete mnemonic matching and splicing. The skill state machine only retains two states, "beginner" and "automation," to reduce storage overhead. This implementation method achieves the complete closed-loop operation of the core function of this invention on network-free, low-power terminal devices, and can serve as an effective supplement to the main implementation method in resource-constrained scenarios.

Claims

1. A calligraphy teaching method based on point-to-point action decomposition, characterized in that, Includes the following steps: The writing process of regular script strokes is uniformly defined as four action phases: "falling, moving, dotting, and returning." Among them, "falling" is the phase of starting the stroke and placing it on the paper, "moving" is the phase of moving the stroke and pushing it forward, "dotting" is the phase of pausing the stroke and accumulating momentum, and "returning" is the phase of returning the stroke and finishing the stroke. A database of stroke action formulas was established, which contains 14 basic stroke action formulas and 14 compound stroke action formulas, totaling 28 sets of stroke-specific action formulas. All formulas are based on the "dot return" action logic as the underlying encoding rule. Based on the rhythm of the music, learners are guided to complete the stroke writing phase by phase according to the described action formula; The smart writing pen, which integrates an inertial measurement unit and a pressure sensor, collects sensor timing data during the writing process, and automatically segments and quantifies the four action phases. Based on the scoring results, precise error correction feedback is generated, and action tips with appropriate granularity are dynamically pushed according to the current state of the learner's skill state machine.

2. The method according to claim 1, characterized in that, The mnemonic for the 14 basic stroke actions is as follows: Right dot corresponds to "fall back and lift", hanging needle vertical corresponds to "dot back and vertical tip", drooping dew vertical corresponds to "dot back and vertical close", vertical hook corresponds to "dot back and vertical hook", long horizontal corresponds to "long horizontal close", short left-falling stroke corresponds to "dot back and left-falling stroke", lifting corresponds to "dot back and lifting", long left-falling stroke corresponds to "dot back and long left-falling stroke", vertical left-falling stroke corresponds to "dot back and vertical left-falling stroke", right-falling stroke corresponds to "right-falling stroke and dot flat exit", curved hook corresponds to "curved dot and hook", reclining hook corresponds to "curved dot and push hook", slanted hook corresponds to "dot back and slanted hook", and vertical curved hook corresponds to "dot back and vertical curved hook".

3. The method according to claim 1, characterized in that, The 14 compound stroke action formulas are generated by combining basic strokes through a dot-back action. The combination rules are as follows: the basic strokes of horizontal, vertical, left-falling, right-falling, and hooking strokes embed a return stroke at the end of their "dot" phase and connect with the "fall" phase of the next stroke to form compound strokes such as horizontal turn, vertical lift, left-falling dot, and left-falling turn. The action formulas of compound strokes are formed by sequentially splicing the formulas of each component basic stroke in the writing order, with the "dot-back" action as the anchor point for the connection and turning between each component stroke.

4. The method according to claim 1, characterized in that, The guided steps based on musical rhythm are as follows: each word in the action mnemonic of each stroke is matched one-to-one with a beat unit of a preset musical rhythm. The learner follows the rhythm and recites the corresponding word while simultaneously completing the writing force action of the corresponding phase. The external rhythm constraint helps the learner perceive the rhythm of force switching between phases and promotes the internalization of muscle memory.

5. The method according to claim 1, characterized in that, The step of automatically segmenting the four action phases is implemented using a three-stage pipeline: In the first stage, the writing sensor time series data is subjected to first-order difference to calculate the pen tip instantaneous speed time series and pressure change rate sequence. Based on the speed and pressure threshold conditions corresponding to each phase, the candidate phase boundary point set is scanned and marked to initially divide the time series into four candidate intervals. In the second stage, a four-state hidden Markov model is constructed. The sliding window statistical features of velocity, pressure, and pitch angle are used as the observation feature vectors. Viterbi decoding is performed on the local time window near the candidate boundary to output a refined four-phase boundary timestamp. In the third stage, the dynamic time warping and normalization distance of each phase time segment is calculated with the pre-stored standard action template. When the normalization distance exceeds the preset threshold, the phase segment is fed back to the second stage, the initial state probability is adjusted and the Viterbi decoding is re-executed until all phases pass the verification and the set of time segments with phase labels is output.

6. The method according to claim 1, characterized in that, In the process-oriented quantitative scoring step, the speed-pressure ratio is extracted as the core quality indicator for the "point" action phase. The speed-pressure ratio is defined as the ratio of the average pen tip speed to the average pen tip pressure within that phase. The speed-pressure ratio of standard writing should fall within the confidence interval determined by statistics from the standard template; When the speed-pressure ratio deviates from the confidence interval, the system determines that there is an abnormality in the pen-pausing and accumulating action, and uses this abnormality information as input for root cause diagnosis.

7. The method according to claim 1, characterized in that, The steps for generating accurate error correction feedback include: using a convolutional neural network to extract character shape defect feature vectors from the final stroke image; concatenating the character shape defect feature vectors with the temporal process feature vectors of each action phase to form a fusion feature vector; inputting the fusion feature vector into a pre-trained multi-label classifier; outputting the error-causing weight probability distribution of each action phase; determining the phase with the highest error-causing weight probability as the root cause node; retrieving the corresponding natural language feedback template in a preset knowledge graph; generating error correction text containing specific action correction suggestions; and pushing it to the learner's interface.

8. A calligraphy teaching system implementing the method of any one of claims 1 to 7, characterized in that, include: The multimodal data acquisition module is configured to simultaneously acquire three-dimensional coordinates, pen barrel pitch angle, pen barrel roll angle, pen tip contact pressure timing data, and final stroke image of the writing process through an intelligent writing pen that integrates an inertial measurement unit and a pressure sensor. The action phase segmentation module is configured to perform a three-stage pipeline on the time series data, namely differential threshold pre-segmentation, hidden Markov model fine segmentation, and dynamic time warping template alignment verification, and output a set of time series segments with phase annotations. The process-oriented quantitative scoring module is configured to extract multi-dimensional features and score each phase time segment independently, and output a four-dimensional phase scoring vector. The root cause diagnosis module is configured to fuse the glyph defect features of the image channel with the process features of the temporal channel, and output the error-causing weight probability distribution and accurate error correction feedback text for each action phase. The adaptive mnemonic generation and push module is configured to dynamically generate calligraphy mnemonics based on the learner's skill state machine at the corresponding granular level in the mnemonic knowledge graph and push them to the learner's interface.

9. The system according to claim 8, characterized in that, The adaptive mnemonic generation and push module includes a skill state machine submodule, which includes three states: "Beginner," "Advanced," and "Automated." When the writing sequence smoothness and phase comprehensive pass rate exceed their respective upgrade thresholds in three consecutive writing sessions, the skill state is upgraded to a higher level. When the phase comprehensive pass rate is below the downgrade threshold in two consecutive writing sessions, the skill state is downgraded to a lower level. The "Beginner" state corresponds to pushing fine-grained four-step mnemonics, the "Advanced" state corresponds to pushing medium-grained merging mnemonics, and the "Automated" state corresponds to pushing macroscopic fusion mnemonics. When the root cause diagnosis module detects a specific phase anomaly, it adds a reinforcement mnemonic for that phase anomaly after the regular mnemonics.

10. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the calligraphy teaching method according to any one of claims 1 to 7.

Images