Handwritten Chinese character aesthetic degree evaluation method and system
By constructing diverse training datasets and multi-stage training processes, and combining multi-channel feature fusion and attention enhancement modules, the problems of insufficient data and poor model generalization in the aesthetic evaluation of handwritten Chinese characters are solved, and high-precision automatic aesthetic evaluation is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2026-02-12
- Publication Date
- 2026-05-29
AI Technical Summary
Existing handwritten Chinese character aesthetic evaluation technologies rely on manual scoring, which is time-consuming, labor-intensive, and highly subjective. It is difficult to obtain large-scale, high-density continuous aesthetic distribution samples. Traditional models cannot simultaneously take into account relative ranking and absolute scores, and they are difficult to capture the global structural relationships and local geometric features of handwritten Chinese characters, resulting in limited model generalization performance.
By constructing diverse training datasets, continuous aesthetic pseudo-labels are generated. An aesthetic network with multi-channel feature fusion and multi-stage training is adopted, including a multi-channel feature fusion layer, a backbone feature extraction module, an attention enhancement module, and a staged training process. Combined with pseudo-label supervision and self-distillation techniques, the aesthetic evaluation of the model is optimized.
It achieves high-precision and scalable automatic evaluation of the aesthetics of handwritten Chinese characters, improves the model's ranking ability and absolute scoring accuracy, and has versatility and scalability.
Smart Images
Figure CN122116387A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of computer vision, artificial intelligence, and font aesthetic analysis, specifically relating to a method and system for evaluating the aesthetic appeal of handwritten Chinese characters. Its core technologies include aesthetic appeal network construction, continuous aesthetic appeal pseudo-annotation generation, multi-channel feature fusion, and multi-stage training and evaluation strategies. Background Technology
[0002] Current techniques for evaluating the aesthetic appeal of handwritten Chinese characters primarily rely on manual scoring or traditional machine learning methods. These techniques suffer from several problems: manual scoring is time-consuming, labor-intensive, and highly subjective, making it difficult to obtain large-scale, high-density, continuous aesthetic distribution samples; existing models often employ single-task regression or classification, failing to simultaneously consider relative ranking and absolute scores, leading to inconsistencies between predicted scores and human aesthetic judgment; furthermore, handwritten Chinese characters are complex in strokes and diverse in structure, with significant variations in writing style, stroke order, stroke thickness, local noise, and distortion, making it difficult for traditional methods to capture global structural relationships and local geometric features, thus limiting model generalization performance. Therefore, there is an urgent need for a method for evaluating the aesthetic appeal of handwritten Chinese characters that combines automatic data generation and continuous aesthetic scoring to achieve high-precision and scalable aesthetic prediction. Summary of the Invention
[0003] This invention aims to address the shortcomings of existing handwritten Chinese character aesthetic evaluation technologies, such as insufficient data, strong subjectivity of manual scoring, and poor model generalization. By automatically generating diverse samples, continuous pseudo-labeling, multi-channel feature fusion, and multi-stage aesthetic network training, a training system covering a wide range of writing styles and continuous aesthetic distribution is constructed, achieving high-precision and scalable automatic evaluation of handwritten Chinese character aesthetics.
[0004] The core solution of this invention includes the following modules: a method for evaluating the aesthetic appeal of handwritten Chinese characters, comprising the following steps: (1) Obtain diverse samples through public font libraries and handwritten databases. Use standard fonts as high aesthetic benchmarks and apply geometric structure perturbation, stroke detail perturbation and stylization perturbation according to the target score difference. Combine with manually labeled real samples to construct continuous aesthetic pseudo-labels for subsequent model training. (2) Clean, normalize and construct multi-channel images of Chinese characters, including grayscale images, binary masks, skeleton heatmaps and distance transformation maps, to provide high-quality input and structural features for the model; (3) Construct an aesthetic evaluation network; the structure of the aesthetic evaluation network is as follows: starting from the input layer, it is connected in sequence to the multi-channel feature fusion layer, the backbone feature extraction module, the attention enhancement module and the output layer; The multi-channel feature fusion layer is used to fuse multi-channel input features; the backbone feature extraction module is used to extract deep features of handwritten Chinese characters and is an optional structure, which forms different levels of feature representation through hierarchical feature extraction; the attention enhancement module is used to perform weighted enhancement on the features corresponding to key strokes or component regions; and the output layer is used to output an aesthetic score of 0–100. (4) A phased training process is adopted: In the first phase, pseudo-labels are used for basic supervised training to obtain the initial model; in the second phase, the initial model is used as the teacher to perform self-distillation, and the scoring stability is improved by aligning soft labels with features; in the third phase, a light regression is performed on the basis of the distillation results to make the model output more consistent with the continuous aesthetic distribution and obtain the final model. (5) The final model is evaluated by multiple indicators, including: Rank Correlation, Score Correlation (Pearson / Spearman), High Score Absolute Error (MAE), and Distribution Consistency (KL / JS distance), in order to comprehensively measure the model’s aesthetic prediction quality and label fitting ability. (6) Input the handwritten Chinese characters into the final model and evaluate its aesthetic appeal.
[0005] Specifically, step (1) includes: (1.1) Obtain original fonts and handwritten samples from publicly available standard font libraries (such as KaiTi, XingKai, and FangSong), design font libraries (such as ZCOOL Fonts and ZaoZiFang), and handwritten databases (such as CASIA-HWDB and SCUT-HCCDoc) to construct a preliminary diverse sample library; (1.2) Determine the type of perturbation: geometric perturbation, including scaling, rotation, translation, and shearing; stroke detail perturbation, including broken strokes, burrs, and uneven thickness; stylistic perturbation, including intensity variation. (1.3) Develop a quantitative mapping table for “perturbation degree - score decrease”, wherein the perturbation includes scaling perturbation, rotation, translation, truncating strokes, and thickness fluctuation; (1.4) Construct a readable structured index table. For each descent gradient, manually bind the single-class perturbation parameter and the multi-class perturbation superposition parameter under that gradient to ensure the accuracy and consistency of the score descent. (1.5) Using standard calligraphy fonts as the high aesthetic benchmark, set the gradient step size, apply gradient geometric structure perturbation, stroke detail perturbation and stylization perturbation to generate continuous aesthetic pseudo-labels and form high-density samples; (1.6) Real sample collection: Collect handwritten Chinese characters of different writing levels and manually annotate them.
[0006] Step (2) specifically includes: (2.1) Convert the handwritten Chinese character images to grayscale and generate multi-channel representations such as grayscale image, binary mask, skeleton heatmap, and DistanceTransform channel; (2.2) Scale each image to a uniform size; (2.3) Perform data cleaning, including but not limited to removing blank or overly dark / bright samples, checking skeleton connectivity and abnormal stroke structure, verifying character labels with OCR or handwriting recognition models, removing erroneous characters and severely blurred samples, and adjusting the distribution of character category samples by resampling or weighted loss.
[0007] In step (3), the multi-channel feature fusion layer is constructed by the same handwritten Chinese character sample into a multi-channel input tensor and aligned to the same size. The multi-channel features include, but are not limited to, grayscale image channels, binary mask channels, skeleton heatmap channels and distance transformation channels. The multi-channel feature fusion layer performs channel dimension splicing or weighted summation on each channel and outputs fused features. The backbone feature extraction module inputs the fused features into the backbone network for stepwise feature extraction. The backbone network is selected from ResNet, EfficientNet, ConvNeXt, or CNN+Transformer. The backbone feature extraction module is composed of several feature extraction units connected in series. Each feature extraction unit includes a feature extraction layer and a normalization layer, and in addition to the first layer, it further includes a downsampling layer to obtain feature representations at different levels. The attention enhancement module applies channel attention and spatial attention weights to the features output by the backbone feature extraction module in sequence. Channel attention is used to adjust the weights of different channel features, while spatial attention is used to highlight key strokes or component areas. The output layer inputs the attention-enhanced features into the regression output head, which includes a global pooling layer and a fully connected layer, and outputs an aesthetic score of 0–100 through a linear mapping.
[0008] Step (4) specifically includes: (4.1) Divide the synthesized continuous pseudo-label dataset into training set / validation set / test set in a hierarchical ratio of 8:1:1. When hierarchizing, the dataset should be divided into partitions of 0-20, 20-40, 40-60, 60-80, and 80-100. (4.2) The pseudo-label supervised pre-trained target network has the following loss function: (1-1) Where N is the batch sample size. These are pseudo-labels (0-100). Here, T represents the network prediction value, and T is the temperature coefficient. (4.3) Self-distillation optimization: A "teacher-student" distillation framework was constructed. The teacher model used a large-capacity backbone (EfficientNet-B4 / ConvNeXt-B) and was trained to convergence using a mixed dataset of pseudo-labels and real samples. The student model was the target evaluation network (ResNet18 / EfficientNet-B2), with the loss function being... (1-2) (1-3) (1-4) in, Output for the teacher model, Output for the student model. These are pseudo-tags, set to 0-100. The distillation temperature. = 0.7, = 0.3, N is the batch sample size, K is dimensionality; (4.4) Lightweight Fine-tuning with Real Samples: A small number of real, manually labeled samples are used to fine-tune the training and validation sets in a 9:1 ratio. The loss function is: (1-5) (1-6) (1-7) in, , These are pseudo-labels (0-100). , This is the network prediction value. Marginal value, This is an indicator function (1 if the condition is met, 0 otherwise). = 0.6, where N is the batch sample size; (4.5) Training monitoring and optimization: Real-time monitoring of Spearman coefficients, Pearson coefficients, MAE, and JS divergence. Monitoring module gradients and adjusting the learning rate when anomalies occur.
[0009] Step (5) specifically includes: (5.1) To evaluate the accuracy of the model in ranking relative aesthetic qualities, calculate the Spearman rank correlation coefficient (ρ), which satisfies the formula: (1-8) Where n is the total number of test samples, Let be the difference between the "true aesthetic level" and the "model predicted aesthetic level" of the i-th sample, where the level = the ranking of the sample by aesthetic score; (5.2) To measure the relevance of the model's absolute score predictions, calculate the Pearson linear correlation coefficient (r), which satisfies the formula... (1-9) in, These are genuine and fake labels. The true mean score. To predict false labels, To predict the mean score; (5.3) Verify the accuracy of the model's predictions in high-segment segments, and calculate the mean absolute error (MAE) in high-segment segments. ), satisfying the formula: (1-10) in, For the sample of true scores of 80-100, For the high-segment sample size, These are genuine and fake labels. To predict false labels; (5.4) Verify the consistency between the model output scores and the true distribution, and calculate the score distribution consistency index—JS divergence (JS divergence). ), satisfying the formula (1-11) (1-12) (1-13) Where P is the distribution of the actual aesthetic scores, Q is the distribution of the model-predicted aesthetic scores, and M is the "average distribution" of P and Q. This represents the proportion of the true samples in the k-th score interval. The model predicts the proportion of samples for the k-th score interval; A second aspect of the present invention relates to a handwritten Chinese character aesthetic evaluation system, comprising a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement a handwritten Chinese character aesthetic evaluation method of the present invention.
[0010] A third aspect of the invention relates to a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements a method for evaluating the aesthetic appeal of handwritten Chinese characters according to the invention.
[0011] The beneficial effects of this invention are that it constructs diverse training datasets through multi-source font libraries and handwritten samples, generates continuous aesthetic labels through perturbation mapping, improves feature expression and character robustness by employing multi-channel feature fusion and attention modules, and optimizes the model's ranking ability and absolute scoring accuracy through phased training, thereby generating a high-precision and scalable handwritten Chinese character aesthetic evaluation model. The results are better than existing methods and have universality and scalability. Attached Figure Description
[0012] Figure 1 This is a schematic diagram of the overall process of the method for evaluating the aesthetic appeal of handwritten Chinese characters according to the present invention. The diagram shows five main steps: constructing a continuous pseudo-label database, constructing a multi-channel feature map, building a fusion attention network, multi-stage training and fine-tuning, and comprehensive evaluation of multi-dimensional indicators.
[0013] Figure 2 shows an example of a continuous aesthetic pseudo-label sample generated based on gradient perturbation in this invention; Figure 2(a) shows a standard font benchmark sample without perturbation (aesthetic score: 100 points); Figure 2(b) shows a sample after applying slight scaling and rotation perturbation (aesthetic score: 80 points); Figure 2(c) shows a sample after applying moderate geometric transformation and stroke noise (aesthetic score: 60 points); Figure 2(d) shows a sample after applying severe skewing and stroke truncation (aesthetic score: 40 points); Figure 2(e) shows a sample under extreme degradation mode (aesthetic score: 20 points).
[0014] Figure 3 is a schematic diagram of the multi-channel feature representation generated after preprocessing the handwritten Chinese character image in this invention; Figure 3(a) shows channel one: a unified grayscale image that preserves texture and stroke details; Figure 3(b) shows channel two: a binary mask that extracts contours based on Otsu adaptive thresholding; Figure 3(c) shows channel three: a skeleton topological heatmap that represents the connectivity and structure of strokes; Figure 3(d) shows channel four: a distance transformation map that represents the thickness of strokes and the center measurement.
[0015] Figure 4 This is a graph showing the change of the mean absolute error (MAE) of the model in the high-scoring range (80-100 points) of the validation set with the number of training rounds. The background color areas in the graph correspond to the three stages of the training process: stage one (pseudo-label pre-training), stage two (self-distillation optimization), and stage three (real sample fine-tuning). Detailed Implementation
[0016] The technical solution of the present invention will be further described below with reference to the accompanying drawings.
[0017] Example 1
[0018] The core technology of this invention is to use an aesthetics network to evaluate the aesthetics score of handwritten Chinese characters.
[0019] This invention proposes a method and system for evaluating the aesthetic appeal of handwritten Chinese characters, comprising the following steps: (1) Dataset expansion (1.1) Obtain original fonts and handwritten samples from publicly available standard font libraries (such as KaiTi, XingKai, and FangSong), design font libraries (such as ZCOOL Fonts and ZaoZiFang), and handwritten databases (such as CASIA-HWDB and SCUT-HCCDoc) to construct a preliminary diverse sample library; (1.2) Determine the type of perturbation: geometric perturbation, including scaling, rotation, translation, and shearing; stroke detail perturbation, including broken strokes, burrs, and uneven thickness; stylistic perturbation, including intensity variation. (1.3) Develop a quantitative mapping table for “perturbation degree - score decrease” (Table 1), wherein the perturbation includes scaling perturbation, rotation, translation, truncating strokes, and thickness fluctuation; Table 1
[0020] (1.4) Construct a readable structured index table (Table 2). For each descent gradient, manually bind the single-class perturbation parameter and the multi-class perturbation superposition parameter under that gradient to ensure the accuracy and consistency of the score descent. Table 2
[0021] (1.5) Using standard calligraphy font as the high aesthetic benchmark (100 points), set a gradient step size of 0.5 points, apply gradient geometric structure perturbation, stroke detail perturbation and stylization perturbation to generate continuous aesthetic pseudo-labels (Figure 2) and form high-density samples; (1.6) Real sample collection: Collect handwritten Chinese characters of different writing levels and manually annotate them.
[0022] (2) Dataset processing: Image processing is performed on the dataset expanded in (1).
[0023] (2.1) The handwritten Chinese character images are uniformly grayscaled, and multi-channel representations such as grayscale image, binary mask, skeleton heatmap, and DistanceTransform channel are generated (Figure 3). (2.2) Scale each image to a uniform size; (2.3) Perform data cleaning, including but not limited to removing blank or overly dark / bright samples, checking skeleton connectivity and abnormal stroke structure, verifying character labels with OCR or handwriting recognition models, removing erroneous characters and severely blurred samples, and adjusting the distribution of character category samples by resampling or weighted loss.
[0024] (3) Model selection and construction: Based on the dataset processed in (2), construct a suitable training model.
[0025] (3.1) Establish a multi-channel feature fusion layer, construct the same handwritten Chinese character sample into a multi-channel input tensor and align it to the same size, wherein the multi-channel features include but are not limited to grayscale image channels, binary mask channels, skeleton heatmap channels and distance transformation channels; the multi-channel feature fusion layer performs channel dimension splicing or weighted summation on each channel and outputs fused features; (3.2) Establish a backbone feature extraction module, input the fused features into the backbone network for stepwise feature extraction, and select EfficientNet as the backbone network; the backbone feature extraction module is composed of several feature extraction units connected in series, each feature extraction unit includes a feature extraction layer and a normalization layer, and further includes a downsampling layer in addition to the first level, so as to obtain feature representations at different levels. (3.3) Establish an attention enhancement module, and apply channel attention and spatial attention weights to the features output by the backbone feature extraction module in sequence. Channel attention is used to adjust the weights of different channel features, and spatial attention is used to highlight key strokes or component areas. (3.4) Establish an output layer, input the attention-enhanced features into the regression output head, the regression output head includes a global pooling layer and a fully connected layer, and outputs an aesthetic score of 0–100 through linear mapping.
[0026] (4) Training a model to evaluate the aesthetic appeal of handwritten Chinese characters
[0027] (4.1) Divide the synthesized continuous pseudo-label dataset into training set / validation set / test set in a hierarchical ratio of 8:1:1. When hierarchizing, the dataset should be divided into partitions of 0-20, 20-40, 40-60, 60-80, and 80-100. (4.2) The pseudo-label supervised pre-trained target network has the following loss function: (1-1) Where N is the batch sample size. These are pseudo-labels (0-100). Here, T represents the network prediction value, and T is the temperature coefficient. (4.3) Self-distillation optimization: A "teacher-student" distillation framework was constructed. The teacher model used a large-capacity backbone (EfficientNet-B4 / ConvNeXt-B) and was trained to convergence using a mixed dataset of pseudo-labels and real samples. The student model was the target evaluation network (ResNet18 / EfficientNet-B2), with the loss function being... (1-2) (1-3) (1-4) in, Output for the teacher model, Output for the student model. These are pseudo-tags, set to 0-100. The distillation temperature. = 0.7, = 0.3, N is the batch sample size, K is dimensionality; (4.4) Lightweight Fine-tuning with Real Samples: A small number of real, manually labeled samples are used to fine-tune the training and validation sets in a 9:1 ratio. The loss function is... (1-5) (1-6) (1-7) in, , These are pseudo-labels (0-100). , This is the network prediction value. Marginal value, This is an indicator function (1 if the condition is met, 0 otherwise). = 0.6, where N is the batch sample size; (4.5) Training monitoring and optimization: Real-time monitoring of Spearman coefficients, Pearson coefficients, and MAE ( Figure 4 ), JS divergence. Monitor module gradients and adjust the learning rate when anomalies occur.
[0028] (5) Evaluation Model
[0029] (5.1) To evaluate the accuracy of the model in ranking relative aesthetic qualities, calculate the Spearman rank correlation coefficient (ρ), which satisfies the formula: (1-8) Where n is the total number of test samples, Let be the difference between the "true aesthetic level" and the "model predicted aesthetic level" of the i-th sample, where the level = the ranking of the sample by aesthetic score; (5.2) To measure the relevance of the model's absolute score predictions, calculate the Pearson linear correlation coefficient (r), which satisfies the formula... (1-9) in, These are genuine and fake labels. The true mean score. To predict false labels, To predict the mean score; (5.3) Verify the accuracy of the model's predictions in high-segment segments, and calculate the mean absolute error (MAE) in high-segment segments. ), satisfying the formula: (1-10) in, For the sample of true scores of 80-100, For the high-segment sample size, These are genuine and fake labels. To predict false labels; (5.4) Verify the consistency between the model output scores and the true distribution, and calculate the score distribution consistency index—JS divergence (JS divergence). ), satisfying the formula (1-11) (1-12) (1-13) Where P is the distribution of the actual aesthetic scores, Q is the distribution of the model-predicted aesthetic scores, and M is the "average distribution" of P and Q. This represents the proportion of the true samples in the k-th score interval. The model predicts the proportion of samples for the k-th score interval; Example 2 This embodiment relates to a handwritten Chinese character aesthetic evaluation system, including a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they are used to implement a handwritten Chinese character aesthetic evaluation method of Embodiment 1.
[0030] Example 3
[0031] The present invention relates to a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements a method for evaluating the aesthetic appeal of handwritten Chinese characters as described in Embodiment 1.
[0032] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.
Claims
1. A method for evaluating the aesthetic appeal of handwritten Chinese characters, characterized in that, Includes the following steps: (1) Obtain diverse samples through public font libraries and handwritten databases. Use standard fonts as high aesthetic benchmarks and apply geometric structure perturbation, stroke detail perturbation and stylization perturbation according to the target score difference. Combine with manually labeled real samples to construct continuous aesthetic pseudo-labels for subsequent model training. (2) Clean, normalize and construct multi-channel images of Chinese characters, including grayscale images, binary masks, skeleton heatmaps and distance transformation maps, to provide high-quality input and structural features for the model; (3) Construct an aesthetic evaluation network; the structure of the aesthetic evaluation network is as follows: starting from the input layer, it is connected in sequence to the multi-channel feature fusion layer, the backbone feature extraction module, the attention enhancement module and the output layer; The multi-channel feature fusion layer is used to fuse multi-channel input features; the backbone feature extraction module is used to extract deep features of handwritten Chinese characters and has an optional structure. The backbone feature extraction module forms different levels of feature representation through hierarchical feature extraction; the attention enhancement module is used to perform weighted enhancement on the features corresponding to key strokes or component regions; and the output layer is used to output an aesthetic score of 0–100. (4) A phased training process is adopted: In the first phase, pseudo-labels are used for basic supervised training to obtain the initial model; in the second phase, the initial model is used as the teacher to perform self-distillation, and the scoring stability is improved by aligning soft labels with features; in the third phase, a light regression is performed on the basis of the distillation results to make the model output more consistent with the continuous aesthetic distribution and obtain the final model. (5) The final model is evaluated by multiple indicators, including: Rank Correlation, Score Correlation (Pearson / Spearman), High Score Absolute Error (MAE), and Distribution Consistency (KL / JS distance), in order to comprehensively measure the model’s aesthetic prediction quality and label fitting ability. (6) Input the handwritten Chinese characters into the final model and evaluate its aesthetic appeal.
2. The method according to claim 1, characterized in that, Step (1) specifically includes: (1.1) Obtain original fonts and handwritten samples from publicly available standard font libraries (such as KaiTi, XingKai, and FangSong), design font libraries (such as ZCOOL Fonts and ZaoZiFang), and handwritten databases (such as CASIA-HWDB and SCUT-HCCDoc) to construct a preliminary diverse sample library; (1.2) Determine the type of perturbation: geometric perturbation, including scaling, rotation, translation, and shearing; stroke detail perturbation, including broken strokes, burrs, and uneven thickness; stylistic perturbation, including intensity variation. (1.3) Develop a quantitative mapping table for "perturbation degree - score decrease", wherein the perturbation includes scaling perturbation, rotation, translation, truncating strokes, and thickness fluctuation; (1.4) Construct a readable structured index table. For each descent gradient, manually bind the single-class perturbation parameter and the multi-class perturbation superposition parameter under that gradient to ensure the accuracy and consistency of the score descent. (1.5) Using standard calligraphy fonts as the high aesthetic benchmark, set the gradient step size, apply gradient geometric structure perturbation, stroke detail perturbation and stylization perturbation to generate continuous aesthetic pseudo-labels and form high-density samples; (1.6) Real sample collection: Collect handwritten Chinese characters of different writing levels and manually annotate them.
3. The method according to claim 1, characterized in that, Step (2) specifically includes: (2.1) Convert the handwritten Chinese character images to grayscale and generate multi-channel representations such as grayscale image, binary mask, skeleton heatmap, and DistanceTransform channel; (2.2) Scale each image to a uniform size; (2.3) Perform data cleaning, including but not limited to removing blank or overly dark / bright samples, checking skeleton connectivity and abnormal stroke structure, verifying character labels with OCR or handwriting recognition models, removing erroneous characters and severely blurred samples, and adjusting the distribution of character category samples by resampling or weighted loss.
4. The method according to claim 1, characterized in that, Step (3) describes the establishment of a multi-channel feature fusion layer, which constructs the same handwritten Chinese character sample into a multi-channel input tensor and aligns it to the same size. The multi-channel features include, but are not limited to, grayscale image channels, binary mask channels, skeleton heatmap channels, and distance transformation channels. The multi-channel feature fusion layer performs channel dimension splicing or weighted summation on each channel to output fused features. The backbone feature extraction module inputs the fused features into the backbone network for stepwise feature extraction. The backbone feature extraction module is selected from ResNet, EfficientNet, ConvNeXt, or CNN+Transformer. The backbone feature extraction module is composed of several feature extraction units connected in series. Each feature extraction unit includes a feature extraction layer and a normalization layer, and further includes a downsampling layer in addition to the first layer to obtain feature representations at different levels. The attention enhancement module applies channel attention and spatial attention weights to the features output by the backbone feature extraction module in sequence. Channel attention is used to adjust the weights of different channel features, while spatial attention is used to highlight key strokes or component areas. The output layer inputs the attention-enhanced features into the regression output head, which includes a global pooling layer and a fully connected layer, and outputs an aesthetic score of 0–100 through a linear mapping.
5. The method according to claim 1, characterized in that, Step (4) specifically includes: (4.1) Divide the synthesized continuous pseudo-label dataset into training set / validation set / test set in a hierarchical ratio of 8:1:
1. When hierarchizing, the dataset should be divided into partitions of 0-20, 20-40, 40-60, 60-80, and 80-100. (4.2) The pseudo-label supervised pre-trained target network has the following loss function: （1-1） Where N is the batch sample size. These are pseudo-labels (0-100). Here, T represents the network prediction value, and T is the temperature coefficient. (4.3) Self-distillation optimization: A "teacher-student" distillation framework was constructed. The teacher model used a large-capacity backbone (EfficientNet-B4 / ConvNeXt-B) and was trained to convergence using a mixed dataset of pseudo-labels and real samples. The student model was the target evaluation network (ResNet18 / EfficientNet-B2), and the loss function was: （1-2） （1-3） （1-4） in, Output for the teacher model, Output for the student model. These are pseudo-tags, set to 0-100. The distillation temperature. = 0.7, = 0.3, N is the batch sample size, K is dimensionality; (4.4) Lightweight Fine-tuning with Real Samples: A small number of real, manually labeled samples are used to fine-tune the training and validation sets in a 9:1 ratio. The loss function is: （1-5） （1-6） （1-7） in, , These are pseudo-labels (0-100). , This is the network prediction value. Marginal value, This is an indicator function; it returns 1 if the condition is met, and 0 otherwise. = 0.6, where N is the batch sample size; (4.5) Training monitoring and optimization: Real-time monitoring of Spearman coefficients, Pearson coefficients, MAE, and JS divergence. Monitoring module gradients and adjusting the learning rate when anomalies occur.
6. The method according to claim 1, characterized in that, Step (5) specifically includes: (5.1) To evaluate the accuracy of the model in ranking relative aesthetic qualities, calculate the Spearman rank correlation coefficient (ρ), which satisfies the formula: （1-8） Where n is the total number of test samples, Let be the difference between the "true aesthetic level" and the "model predicted aesthetic level" of the i-th sample, where the level = the ranking of the sample by aesthetic score; (5.2) To measure the relevance of the model's absolute score predictions, calculate the Pearson linear correlation coefficient (r), which satisfies the formula... （1-9） in, These are genuine and fake labels. The true mean score. To predict false labels, To predict the mean score; (5.3) Verify the accuracy of the model's predictions in high-segment segments, and calculate the mean absolute error (MAE) in high-segment segments. ), satisfying the formula: （1-10） in, For the sample of true scores of 80-100, For the high-segment sample size, These are genuine and fake labels. To predict false labels; (5.4) Verify the consistency between the model output scores and the true distribution, and calculate the score distribution consistency index—JS divergence (JS divergence). ), satisfying the formula （1-11） （1-12） （1-13） Where P is the distribution of the actual aesthetic scores, Q is the distribution of the model-predicted aesthetic scores, and M is the "average distribution" of P and Q. This represents the proportion of the true samples in the k-th score interval. The model predicts the proportion of samples for the k-th score interval.
7. A system for evaluating the aesthetic appeal of handwritten Chinese characters, characterized in that, The device includes a memory and one or more processors, wherein the memory stores executable code, and the one or more processors execute the executable code to implement a method for evaluating the aesthetic appeal of handwritten Chinese characters according to any one of claims 1-7.
8. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements a method for evaluating the aesthetic appeal of handwritten Chinese characters as described in any one of claims 1-6.