Digital-English full character three-dimensional writing recognition method and system
Patent Information
- Application Number
- CN202211600895.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-13
- Publication Date
- 2026-08-21
- Estimated Expiration
- 2042-12-13
AI Technical Summary
因此，基于惯性传感器构建一套面向英文大小写+数字的62字符3D书写语义识别系统难度很大
[0049] The present invention provides a method and system for three-dimensional handwriting recognition of digits and English characters. By employing variational mode decomposition, time series-complex network conversion, and wavelet transform to enhance the signal of the IMU signal to be recognized, the accuracy of trajectory calculation based on inertial sensors is improved, thereby achieving high-precision human-computer interaction and remote control.
Smart Images
Figure CN116434324B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of character recognition, and in particular to a method and system for recognizing three-dimensional handwriting of numbers and English characters. Background Technology
[0002] With the rise and development of Augmented Reality (AR), Virtual Reality (VR), wearable devices, smart homes, and human-machine interaction, gesture recognition for human-machine interaction has become a hot topic and an essential element in the smart terminal industry. Since inertial sensors / inertial measurement units (IMUs) are unaffected by external environmental factors such as light, obstructions, and noise during data acquisition, gesture recognition based on inertial sensors has a very broad application prospect. Low-cost inertial sensors have advantages such as small size, low power consumption, and strong wearability, and can be widely used in various life scenarios. However, due to their low cost, their quality and performance vary greatly, and the acquired IMU data often contains serious error interference, making it difficult to achieve satisfactory results in tasks such as trajectory reconstruction, motion tracking, and gesture recognition. In recent years, research on inertial sensor gesture recognition tasks has often focused on basic functions such as classifying simple gestures, and the recognition methods cannot adapt to the gesture habits of different individuals, resulting in poor scalability.
[0003] As one of the most challenging tasks in the field of gesture recognition, IMU-based aerial handwriting recognition has broad application prospects. Leveraging low-cost IMUs integrated into mobile phones, watches, and wristbands, users can express semantic information by writing characters. However, the number of characters to be recognized is large; 26 uppercase and lowercase English letters plus 10 Arabic numerals require at least 62 characters to constitute a complete semantic expression system. Therefore, building a 62-character 3D handwriting semantic recognition system based on inertial sensors for uppercase and lowercase English letters plus numbers is very difficult.
[0004] Distinguishing between 62 characters is extremely difficult under conditions of varying writing habits, irregular handwriting, and significant sensor errors. First, many characters have identical uppercase and lowercase forms, such as Cc, Oo, Ss, and Zz. Second, the IMU captures the writing process in the air, including cursive strokes, resulting in high similarity in the movement of some characters, such as et, Vr, kR, 9-q, hn, T-7, and i-7. Third, individual writing habits vary greatly, making characters that would otherwise be easily distinguishable appear very similar, such as b-6, ax, and Mu. Fourth, because users cannot see the movement trajectory in real time during writing, or because the writing process is often casual, the writing is often irregular, leading to very similar movement for many characters, such as 4-6, 1-l, Uu, Ww, Pp, Db, and 0-6. Due to the combined effect of these factors, achieving ultra-high precision handwriting recognition based on IMU is very difficult. Summary of the Invention
[0005] The purpose of this invention is to provide a method and system for three-dimensional handwriting recognition of digits and English characters, which can accurately realize handwriting recognition of digits and English characters based on any IMU and any writing habits, thereby enabling high-precision human-computer interaction and remote control.
[0006] To achieve the above objectives, the present invention provides the following solution:
[0007] A method for recognizing full-character 3D handwriting of numbers and English letters, comprising:
[0008] Acquire the IMU signal to be identified;
[0009] The IMU signal to be identified is enhanced using variational mode decomposition, time series-complex network conversion, and wavelet transform.
[0010] The 3D motion trajectory is determined based on the enhanced IMU signal to be identified.
[0011] Optionally, the IMU signal to be identified is enhanced using variational mode decomposition, time series-complex network conversion, and wavelet transform, specifically including:
[0012] Variational mode decomposition is performed on the IMU signal to be identified to obtain multiple IMF components;
[0013] Transform each IMF component into a complex network;
[0014] The randomness of each IMF component is determined based on the deterministic indices of complex networks.
[0015] Wavelet thresholding is applied to the IMF component with the strongest randomness in each axis IMU signal.
[0016] The IMU signal is reconstructed using the denoised IMF component and the undenoised IMF component, thus enhancing the IMU signal to be identified.
[0017] Optionally, the IMU signal to be identified is enhanced using variational mode decomposition, time series-complex network conversion, and wavelet transform, specifically including:
[0018] Determine the XYZ three-axis velocity in the world coordinate system based on the current enhanced IMU signal to be identified;
[0019] Determine whether the difference between the XYZ three-axis velocities of the current enhanced IMU signal in the world coordinate system at the start and end times is less than the difference between the XYZ three-axis velocities of the previous enhanced IMU signal in the world coordinate system at the start and end times.
[0020] If it is less than, then check if the upper bound of the loop has been reached;
[0021] If the upper bound of the cycle is not reached, then the signal is enhanced by using variational mode decomposition, time series-complex network conversion method and wavelet transform, and the process returns to the step of determining the XYZ three-axis velocity in the world coordinate system based on the IMU signal to be identified after the current signal enhancement.
[0022] If the upper bound of the loop is reached, the enhanced IMU signal to be identified is output.
[0023] If it is equal to or greater than, output the IMU signal to be identified after the previous signal enhancement.
[0024] Optionally, the step of determining the 3D motion trajectory based on the enhanced IMU signal to be identified further includes:
[0025] The first recognition model is determined based on the 3D motion trajectory and the 2D convolutional neural network; the second recognition model is determined based on the IMU signal to be recognized and the 1D convolutional neural network.
[0026] The recognition results are determined by using cooperative learning and competitive learning on the first and second recognition models.
[0027] Optionally, the step of using cooperative learning and competitive learning on the first and second recognition models to determine the recognition result specifically includes:
[0028] The difference between the distributions of the recognition results of the first and second recognition models is determined by using JS divergence.
[0029] Determine the cooperation loss function based on the aforementioned differences;
[0030] The competitiveness of the first identification model and the competitiveness of the second identification model are determined respectively;
[0031] The collaborative model of the first and second identification models is determined based on their competitiveness; the collaborative model includes either the student model or the teacher model.
[0032] Optionally, determining the cooperation loss function based on the difference specifically includes:
[0033] Using the formula Loss cooperation =JS(y spatial ,y time Determine the cooperation loss function Loss cooperation ;
[0034] Among them, y spatial The recognition result of the first recognition model is y. time The result of the second recognition model is JS(), which is the JS divergence function.
[0035] Optionally, the step of using cooperative learning and competitive learning on the first and second recognition models to determine the recognition result further includes:
[0036] The competitiveness of the first identification model and the competitiveness of the second identification model are normalized.
[0037] The competition regularization term is determined based on the competitiveness after normalization of the first identification model and the competitiveness after normalization of the second identification model;
[0038] The cooperative model for determining the first and second identification models is determined based on the competition regularization term.
[0039] Optionally, determining the competition regularization term based on the competitive advantage normalized by the first identification model and the competitive advantage normalized by the second identification model specifically includes:
[0040] Using formula Determine the competition regularization term R competition ;
[0041] in, The first identification model is the normalized competitiveness. This represents the normalized competitiveness of the second identification model.
[0042] A three-dimensional handwriting recognition system for numbers and English characters, comprising:
[0043] The IMU signal acquisition module is used to acquire the IMU signal to be identified.
[0044] The signal enhancement module is used to enhance the IMU signal to be identified by using variational mode decomposition, time series-complex network conversion method and wavelet transform.
[0045] The recognition model determination module is used to determine the 3D motion trajectory based on the enhanced IMU signal to be recognized, and to determine the first recognition model based on the 3D motion trajectory and the 2D convolutional neural network; and to determine the second recognition model based on the IMU signal to be recognized and the 1D convolutional neural network.
[0046] The recognition result determination module is used to determine the recognition result by applying cooperative learning and competitive learning to the first recognition model and the second recognition model.
[0047] A three-dimensional handwriting recognition system for numbers and English characters includes: at least one processor, at least one memory, and computer program instructions stored in the memory, which implement the method when the computer program instructions are executed by the processor.
[0048] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:
[0049] The present invention provides a method and system for three-dimensional handwriting recognition of digits and English characters. By employing variational mode decomposition, time series-complex network conversion, and wavelet transform to enhance the signal of the IMU signal to be recognized, the accuracy of trajectory calculation based on inertial sensors is improved, thereby achieving high-precision human-computer interaction and remote control. Attached Figure Description
[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0051] Figure 1 This is a flowchart illustrating a three-dimensional handwriting recognition method for numbers and English characters provided by the present invention;
[0052] Figure 2 This is a schematic diagram illustrating the effects of different writing habits and non-standard writing provided in Example 1.
[0053] Figure 3 This is a schematic diagram illustrating the effects of different writing habits and non-standard writing provided in Example 2.
[0054] Figure 4 This is a schematic diagram illustrating the principle of signal enhancement.
[0055] Figure 5 This is a schematic diagram of the trajectory calculation results without signal enhancement.
[0056] Figure 6 This is a schematic diagram of the trajectory calculation results after signal enhancement;
[0057] Figure 7 A schematic diagram of the collaborative and competitive learning framework process;
[0058] Figure 8 A schematic diagram illustrating the temporal dynamics and spatial morphological characteristics of the original signal's trajectory.
[0059] Figure 9 This is a schematic diagram showing the temporal dynamics and spatial morphological characteristics of the original signal's trajectory. Detailed Implementation
[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0061] The purpose of this invention is to provide a method and system for three-dimensional handwriting recognition of digits and English characters, which can accurately realize handwriting recognition of digits and English characters based on any IMU and any writing habits, thereby enabling high-precision human-computer interaction and remote control.
[0062] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0063] Because individuals have vastly different writing habits and stroke orders, and their writing is often highly irregular, the movement characteristics of different characters may be similar, such as... Figure 2 As shown. Similarly, due to the influence of different individuals' writing habits and stroke order, the movement characteristics of the same character may be completely different, such as... Figure 3 As shown, this makes the distinction between different characters even less than the distinction between different ways of writing the same character.
[0064] Different users have different writing habits, which means that motion signals such as acceleration and angular velocity recorded by inertial sensors cannot faithfully represent the semantic information corresponding to the writing process. However, no matter how the writing process changes, the final trajectory will always be similar, which means that the trajectory can reflect the most intuitive and essential characteristics of the handwriting process.
[0065] Figure 1This is a schematic diagram of a three-dimensional handwriting recognition method for numbers and English characters provided by the present invention, as shown below. Figure 1 As shown, the present invention provides a three-dimensional handwriting recognition method for numbers and English characters, comprising:
[0066] S101, acquire the IMU signal to be identified;
[0067] S102, variational mode decomposition, time series-complex network conversion method and wavelet transform are used to enhance the IMU signal to be identified;
[0068] S102 specifically includes:
[0069] Variational mode decomposition (VMD) is performed on the IMU signal to be identified to obtain multiple IMF components. The IMU signal is a six-axis IMU signal, that is, each axis IMU signal is decomposed into 5 variational mode functions (IMFs). The 6-axis signal is decomposed into 6 groups of IMF components (5 in each group, for a total of 30 IMFs).
[0070] Each IMF component is transformed into a complex network.
[0071] The randomness of each IMF component is determined based on the deterministic index of complex networks.
[0072] Wavelet thresholding is used to denoise the IMF component with the strongest randomness in each axis IMU signal.
[0073] The IMU signal is reconstructed using the denoised IMF component and the undenoised IMF component, thus enhancing the IMU signal to be identified.
[0074] S102 also includes:
[0075] The XYZ three-axis velocities in the world coordinate system are determined based on the enhanced IMU signal to be identified.
[0076] Determine whether the difference between the XYZ three-axis velocities of the current enhanced IMU signal in the world coordinate system at the start and end times is less than the difference between the XYZ three-axis velocities of the previous enhanced IMU signal in the world coordinate system at the start and end times.
[0077] If it is less than, then it is determined whether the upper limit of the loop has been reached; that is, when the difference between the XYZ three-axis velocity of the IMU signal to be identified in the world coordinate system after the current signal enhancement is less than the difference between the XYZ three-axis velocity of the IMU signal to be identified in the world coordinate system after the previous ...
[0078] If the upper bound of the cycle is not reached, then the signal is enhanced by using variational mode decomposition, time series-complex network conversion method and wavelet transform, and then the step of determining the XYZ three-axis velocity in the world coordinate system based on the IMU signal to be identified after the current signal enhancement is returned.
[0079] If the upper bound of the loop is reached, the enhanced IMU signal to be identified is output.
[0080] If it is equal to or greater than, output the IMU signal to be identified after the previous signal enhancement.
[0081] According to S102 and Figure 4 It can be seen that the signal enhancement provided by the present invention can be divided into a three-stage signal enhancement module of "variational mode decomposition-complex network-wavelet transform", a dynamic property analysis module, and a cyclic feedback judgment module; the three modules are connected in sequence and continuously loop until the stopping condition is met.
[0082] After multiple iterations, the signal is enhanced multiple times, and each enhancement improves the dynamic properties. In fact, the selected dynamic properties are only representative; while the "dynamic cyclic feedback filtering" method enhances the IMU signal and improves the selected dynamic properties, many unselected and unrepresented dynamic properties are also simultaneously enhanced. Ultimately, the IMU signal enhanced by the "dynamic cyclic feedback filtering" method can achieve more accurate trajectory calculation. Figure 5 and Figure 6 The comparison of trajectory calculation results can prove the accuracy of trajectory solution.
[0083] For a specific character, even with different writing habits and stroke counts, the generated trajectory will always be similar. It can be observed that, compared to the raw IMU signal, the motion trajectory reflects the most intuitive, essential, and significant spatial morphological features between different handwritten characters. This feature is particularly important for distinguishing characters with similar motion characteristics. Taking the character 'a' as an example, the motion direction corresponding to its stroke order is very similar to that of characters '9', 'd', and 'u'. This makes the peaks and troughs of the IMU signals for different characters very close in position, with the only difference being the height of the peaks and troughs. This difference is easily masked by sensor errors and differences in motion speed. However, after the IMU signal is decoded into a trajectory, different characters exhibit obvious morphological differences, such as... Figure 8 As shown. However, trajectory features also have certain limitations. For example, the characters 'a' and 'x' have completely different writing styles and motion characteristics, but the resulting trajectories are similar, such as... Figure 9 As shown. At this point, since the original IMU signal records the dynamic properties of the handwriting process over time, the two characters can be easily distinguished based on the signal waveform characteristics.
[0084] S103, determine the 3D motion trajectory based on the enhanced IMU signal to be identified;
[0085] Following S103 are:
[0086] A first recognition model is determined based on the 3D motion trajectory and a 2D convolutional neural network; wherein the projection of the 3D motion trajectory onto XY, XZ, and YZ axes is constructed as a three-channel input image, which serves as the input to the first recognition model. A second recognition model is determined based on the IMU signal to be recognized and a 1D convolutional neural network.
[0087] The recognition model is determined by employing cooperative and competitive learning methods for both the first and second recognition models. This process allows multiple parallel "recognition modules" to take turns acting as "teacher models" to guide the learning of other modules. A "competition mechanism" forces different "recognition modules" to compete with each other, ultimately leading to a collective improvement in recognition accuracy.
[0088] The recognition model and the second recognition model are used to determine the recognition results through cooperative learning and competitive learning, specifically including:
[0089] The difference between the distributions of the recognition results of the first and second recognition models is determined by using JS divergence.
[0090] Determine the cooperation loss function Loss based on the aforementioned differences. cooperation =JS(y spatial ,y time );
[0091] Among them, y spatial The recognition result of the first recognition model is y. time The result of the second recognition model is JS(), which is the JS divergence function.
[0092] The competitiveness of the first identification model and the competitiveness of the second identification model are determined respectively;
[0093] like Figure 7 As shown, the process of determining competitiveness is as follows:
[0094] Both recognition models add a fully connected layer as a predictor when outputting classification results to predict the competitiveness of the recognition module. Let the competitiveness of the morphological feature-dependent recognition module be... The competitiveness of the recognition module based on dynamic features is To ensure that the predicted competitiveness accurately reflects the effectiveness of the corresponding identification module, we supervise the prediction task by using the reciprocal of the squared error of each identification module as the label. Identification modules with high competitiveness are called "strong identification models," and those with low competitiveness are called "weak identification models."
[0095] The competitiveness of the first and second recognition models is normalized; then, these normalized values are used as weights to perform a weighted sum of the two classification results, yielding the final classification result. To make the final classification result more accurate, the model will incorporate [various parameters] during training. and Adjust to a suitable ratio to better integrate the results of the two "recognition modules".
[0096]
[0097]
[0098] To prevent the "weak recognition model" from misleading the "strong recognition model" and to encourage the "weak recognition model" to learn more from the "strong recognition model," we calculated the loss for each "recognition module." cooperation Different weights are assigned to the opposing "recognition module," reflecting its competitiveness. For a "strong recognition model," the "weak recognition model" performs poorly and has low competitiveness; therefore, the "strong recognition model" receives a higher loss. cooperation The model is small and doesn't need to learn too much from the "weak recognition model." It's worth noting that this approach still preserves the channel for the "strong recognition model" to learn from the "weak recognition model," which is crucial for the alternating improvement of the performance of the two "recognition modules." To better highlight the structural characteristics of this cooperative competitive learning framework, the "strong recognition model" is referred to as the "teacher model," and the "weak recognition model" as the "student model."
[0099] This model can determine the "strong recognition model (teacher model)" and the "weak recognition model (student model)" by predicting the competitiveness of each "recognition module," and thus determine the degree to which different "recognition modules" learn from each other (the "student model" learns more from the "teacher model," and the "student model" learns less from the "teacher model"). However, the model cannot always determine which "recognition module" is the "teacher model." The model often predicts similar competitiveness for two "recognition modules," that is:
[0100] To further ensure that the model can distinguish between the "student model" and the "teacher model" from the two "identification models," that is, to force the model to amplify the difference in competitiveness between the two modules, a competition regularization term is determined based on the normalized competitiveness of the first identification model and the normalized competitiveness of the second identification model.
[0101] The collaborative model between the first and second identification models is determined based on a competition regularization term; the collaborative model includes either a student model or a teacher model. This regularization term is added to the loss function of the entire model, and as the "competition regularization term" decreases, and The gap will widen, so the model will clearly predict the “teacher model” and the “student model”, rather than giving similar competitiveness predictions for the two “identification models” in a seemingly plausible way.
[0102] The step of determining the competition regularization term based on the competitiveness normalized by the first identification model and the competitiveness normalized by the second identification model specifically includes:
[0103] Using formula Determine the competition regularization term R competition .
[0104] in, The first identification model is the normalized competitiveness. This represents the normalized competitiveness of the second identification model.
[0105] This invention can achieve 99% ultra-high accuracy in recognizing 10 digits, 26 uppercase English characters, and 26 lowercase English characters. By using a "dynamic cyclic feedback filtering" method to enhance the signal of the IMU, this invention can be applied to any inertial measurement unit, even ultra-low-cost IMUs costing less than 0.5 yuan. Furthermore, this method has no requirements regarding the user's handwriting habits; even with very casual and irregular handwriting, the accuracy remains unaffected.
[0106] Corresponding to the above method, the present invention also provides a three-dimensional handwriting recognition system for all numbers and English characters, the system comprising:
[0107] The IMU signal acquisition module is used to acquire the IMU signal to be identified.
[0108] The signal enhancement module is used to enhance the IMU signal to be identified by using variational mode decomposition, time series-complex network conversion method and wavelet transform.
[0109] The recognition model determination module is used to determine the 3D motion trajectory based on the enhanced IMU signal to be recognized, and to determine the first recognition model based on the 3D motion trajectory and the 2D convolutional neural network; and to determine the second recognition model based on the IMU signal to be recognized and the 1D convolutional neural network.
[0110] The recognition result determination module is used to determine the recognition result by applying cooperative learning and competitive learning to the first recognition model and the second recognition model.
[0111] In order to execute the method corresponding to Embodiment 1 above and achieve the corresponding functions and technical effects, the present invention also provides a three-dimensional handwriting recognition system for full-character digital and English writing, including: at least one processor, at least one memory, and computer program instructions stored in the memory, wherein the method is implemented when the computer program instructions are executed by the processor.
[0112] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.
[0113] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A method for recognizing three-dimensional handwriting of numbers and English characters, characterized in that, include: Acquire the IMU signal to be identified; The IMU signal to be identified is enhanced using variational mode decomposition, time series-complex network conversion, and wavelet transform. The 3D motion trajectory is determined based on the enhanced IMU signal to be identified; The IMU signal to be identified is enhanced using variational mode decomposition, time series-complex network transformation, and wavelet transform, specifically including: Variational mode decomposition is performed on the IMU signal to be identified to obtain multiple IMF components; Transform each IMF component into a complex network; The randomness of each IMF component is determined based on the deterministic indices of complex networks. Wavelet thresholding is applied to the IMF component with the strongest randomness in each axis IMU signal. The IMU signal is reconstructed using the denoised IMF component and the undenoised IMF component to determine the enhanced IMU signal to be identified. Determine the XYZ three-axis velocity in the world coordinate system based on the current enhanced IMU signal to be identified; Determine whether the difference between the XYZ three-axis velocities of the current enhanced IMU signal in the world coordinate system at the start and end times is less than the difference between the XYZ three-axis velocities of the previous enhanced IMU signal in the world coordinate system at the start and end times. If it is less than, then check if the upper bound of the loop has been reached; If the upper bound of the cycle is not reached, then the signal is enhanced by using variational mode decomposition, time series-complex network conversion method and wavelet transform, and the process returns to the step of determining the XYZ three-axis velocity in the world coordinate system based on the IMU signal to be identified after the current signal enhancement. If the upper bound of the loop is reached, the enhanced IMU signal to be identified is output. If it is equal to or greater than, output the IMU signal to be identified after the previous signal enhancement.
2. The method for three-dimensional handwriting recognition of numbers and English characters according to claim 1, characterized in that, The step of determining the 3D motion trajectory based on the enhanced IMU signal to be identified further includes: The first recognition model is determined based on the 3D motion trajectory and the 2D convolutional neural network; the second recognition model is determined based on the IMU signal to be recognized and the 1D convolutional neural network. The recognition results are determined by using cooperative learning and competitive learning on the first and second recognition models.
3. The method for three-dimensional handwriting recognition of numbers and English characters according to claim 2, characterized in that, The process of using cooperative learning and competitive learning on the first and second recognition models to determine the recognition results specifically includes: The difference between the distributions of the recognition results of the first and second recognition models is determined by using JS divergence. Determine the cooperation loss function based on the aforementioned differences; The competitiveness of the first identification model and the competitiveness of the second identification model are determined respectively; The collaborative model of the first and second identification models is determined based on their competitiveness; the collaborative model includes either the student model or the teacher model.
4. The method for three-dimensional handwriting recognition of numbers and English characters according to claim 3, characterized in that, The step of determining the cooperation loss function based on the difference specifically includes: Using formula Determine the cooperation loss function Loss cooperation ; Among them, y spatial The recognition result of the first recognition model is y. time The result of the second recognition model is JS(), which is the JS divergence function.
5. The method for three-dimensional handwriting recognition of numbers and English characters according to claim 3, characterized in that, The method of determining the recognition result by employing cooperative learning and competitive learning on the first and second recognition models specifically includes: The competitiveness of the first identification model and the competitiveness of the second identification model are normalized. The competition regularization term is determined based on the competitiveness after normalization of the first identification model and the competitiveness after normalization of the second identification model; The cooperative model of the first and second identification models is determined based on the competition regularization term.
6. The method for three-dimensional handwriting recognition of numbers and English characters according to claim 5, characterized in that, The step of determining the competition regularization term based on the competitiveness normalized by the first identification model and the competitiveness normalized by the second identification model specifically includes: Using formula Determine the competition regularization term R competition ; in, The normalized competitiveness of the first identification model. This represents the normalized competitiveness of the second identification model.
7. A three-dimensional handwriting recognition system for numbers and English letters, used to implement the three-dimensional handwriting recognition method for numbers and English letters as described in any one of claims 1-6, characterized in that, include: The IMU signal acquisition module is used to acquire the IMU signal to be identified. The signal enhancement module is used to enhance the IMU signal to be identified by using variational mode decomposition, time series-complex network conversion method and wavelet transform. The recognition model determination module is used to determine the 3D motion trajectory based on the enhanced IMU signal to be recognized, and to determine the first recognition model based on the 3D motion trajectory and the 2D convolutional neural network; and to determine the second recognition model based on the IMU signal to be recognized and the 1D convolutional neural network. The recognition result determination module is used to determine the recognition result by applying cooperative learning and competitive learning to the first recognition model and the second recognition model.
8. A three-dimensional handwriting recognition system for numbers and English letters, characterized in that, include: The method comprises at least one processor, at least one memory, and computer program instructions stored in the memory, wherein when the computer program instructions are executed by the processor, a three-dimensional handwriting recognition method for all numbers and English characters as described in any one of claims 1-6 is implemented.