Calligraphy work automatic identification and evaluation method and system

Abstract

The invention provides a calligraphy work automatic identification and evaluation method and system, and the method comprises the steps: collecting calligraphy pattern images of different calligraphers, fonts and styles, and forming a multi-dimensional labeling original data set; carrying out standardization processing on the acquired calligraphy images; extracting stroke-level manual statistical features by using edge detection and contour analysis based on the standardized sample, extracting high-dimensional deep features by using a convolutional neural network, and fusing to form a unified multi-dimensional style feature vector; and performing unsupervised clustering on the fusion feature vectors of all the samples, and establishing a high-dimensional style prototype library. Obtaining a multi-dimensional similarity vector of a newly input calligraphy image; performing semantic unification processing on the new sample; and performing real-time scoring output on the new sample by using the exclusive evaluation model. The method can efficiently and accurately identify and evaluate the calligraphy work, is high in adaptability, and can be widely applied to the fields of calligraphy teaching, calligraphy work evaluation, digital calligraphy management and the like.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to an automatic recognition and evaluation method and system for calligraphy works. Background Technology

[0002] In traditional calligraphy teaching and evaluation scenarios, manual evaluation mainly relies on calligraphy professionals or experienced enthusiasts. However, manual evaluation has obvious drawbacks: (1) manual evaluation of each character is time-consuming and laborious, making it difficult to meet the needs of large-scale teaching or competition evaluation; (2) manual evaluation is highly subjective, with differences in aesthetic standards and professional levels among different evaluators, resulting in a lack of consistency and objectivity in the evaluation results; (3) manual evaluation usually cannot provide detailed feedback in real time, which is not conducive to learners adjusting their writing habits in a timely manner. In recent years, deep learning has made significant progress in the field of image recognition. Convolutional neural networks (CNNs) such as VGG and ResNet have shown powerful capabilities in feature extraction, and people have also made relevant explorations in the automatic recognition of calligraphy. However, due to the wide variety of calligraphy styles, the existing methods for automatic recognition and evaluation of calligraphy are poorly adapted to complex calligraphy styles and are difficult to capture deep artistic features such as "spirit" and "brushstrokes". Traditional feature extraction methods are greatly affected by font type (regular script, running script, clerical script, etc.) and writing style, and their generalization ability is insufficient in the evaluation of different calligraphers' works or samples of different writing levels.

[0003] Therefore, it is necessary to improve existing methods for automatic recognition and evaluation of calligraphy works in order to overcome the shortcomings of existing technologies. Summary of the Invention

[0004] To overcome the problems existing in related technologies, one of the objectives of this invention is to provide an automatic identification and evaluation method for calligraphy works. This method can efficiently and accurately identify and evaluate calligraphy works, and it has high data utilization efficiency, good feature representation ability, and strong adaptability. It can be widely used in calligraphy teaching, calligraphy work evaluation, and digital management of calligraphy.

[0005] An automatic recognition and evaluation method for calligraphy works includes:

[0006] S1. Collect, deduplicate, attribute-label, and structure-store brush calligraphy image samples from different calligraphers, fonts, and styles to form a multidimensional labeled original dataset;

[0007] S2. Perform size normalization, grayscale conversion, various geometric transformations and noise perturbation on the collected calligraphy images, and automatically batch process and associate them with labels to obtain a standardized image set with multiple styles and variations.

[0008] S3. Based on standardized samples, edge detection and contour analysis are used to extract stroke-level manual statistical features, and convolutional neural networks are used to extract high-dimensional deep features, which are then fused to form a unified multi-dimensional style feature vector.

[0009] S4. Perform unsupervised clustering on the fused feature vectors of all samples to establish a high-dimensional style prototype library. Each style prototype contains a central feature vector, typical samples, and style label metadata, and supports incremental clustering to dynamically expand the prototype library.

[0010] S5. Perform standardized preprocessing and feature extraction on the newly input calligraphy image, fuse shallow and deep features, calculate the similarity distribution between it and each prototype in the prototype library, and obtain a multidimensional similarity vector.

[0011] S6. Dynamically select the optimal style prototype based on similarity distribution, and project the features of new samples to the target prototype space through a feature mapping algorithm to achieve semantic consistency processing.

[0012] S7. Call the dedicated evaluation sub-model corresponding to the target prototype in the standardized prototype space to output real-time scores for new samples.

[0013] In a preferred embodiment of the present invention, after real-time scoring output of new samples, the method further includes: S8, continuously monitoring the deviation between the evaluation results and manual annotation or user feedback; if a new style drift or a decrease in accuracy is detected, automatically triggering prototype library expansion and incremental training of sub-models to achieve adaptive optimization of the system's generalization ability.

[0014] In a preferred embodiment of the present invention, the data collection in S1 includes web crawling, API calls and manual collection, the deduplication uses image hash comparison and structural similarity detection, and the attribute annotation includes manual and semi-automatic annotation and expert consensus verification.

[0015] In a preferred embodiment of the present invention, the data enhancement in S2 includes random rotation from -15° to +15°, horizontal or vertical translation from -10 to +10 pixels, scaling ratio from 0.9 to 1.1, horizontal and vertical flipping, and perturbation processing of Gaussian noise (variance 0.001 to 0.01) and salt-and-pepper noise (ratio 0.001 to 0.01).

[0016] In a preferred embodiment of the present invention, the manual statistical features in S3 include stroke length, width, curvature, number of strokes, ratio of blank space to stroke area, and center of gravity position, etc. Global features and deep features are fused by normalized feature splicing or principal component analysis for dimensionality reduction.

[0017] In a preferred embodiment of the present invention, the unsupervised clustering in S4 adopts K-means, DBSCAN or Gaussian mixture model, and the style prototype library supports structured storage and multi-level index retrieval of central feature vectors, typical samples, main font categories and calligrapher label distribution.

[0018] In a preferred embodiment of the present invention, the similarity metric in S5 includes cosine similarity, Euclidean distance and Mahalanobis distance, and outputs a multidimensional similarity distribution vector of the new sample in the prototype space.

[0019] In a preferred embodiment of the present invention, the feature mapping algorithm in S6 includes linear projection, subspace alignment, or adaptive normalization mapping, which supports the consistent alignment of new sample features with the target prototype space.

[0020] In a preferred embodiment of the present invention, the dedicated evaluation sub-model in S7 is trained by a dedicated neural network or ensemble learning algorithm and can be independently incrementally optimized according to the style prototype, outputting a multi-dimensional score including stroke quality, structural rationality and aesthetics.

[0021] In a preferred embodiment of the present invention, the evaluation result monitoring in S8 employs error analysis, statistical process monitoring, and novelty detection to automatically identify new style drift and incrementally expand the prototype library and optimize sub-model parameters.

[0022] The second objective of this invention is to provide an automatic recognition and evaluation system for calligraphy works. This system is used to implement the automatic recognition and evaluation method for calligraphy works described above, and includes:

[0023] Data acquisition module: used to collect, deduplicate, attribute-label, and structure the image samples of calligraphy from different calligraphers, fonts, and styles, forming a multidimensional labeled raw dataset;

[0024] Data preprocessing module: used to perform size normalization, grayscale conversion, various geometric transformations and noise perturbation on the acquired calligraphy images, and automatically batch process and associate them with labels to obtain a standardized image set with multiple styles and variations;

[0025] Feature extraction module: Based on standardized samples, it uses edge detection and contour analysis to extract stroke-level manual statistical features, and uses convolutional neural networks to extract high-dimensional deep features, which are then fused to form a unified multi-dimensional style feature vector;

[0026] Prototype library building module: used to perform unsupervised clustering on the fused feature vectors of all samples to build a high-dimensional style prototype library. Each style prototype contains a central feature vector, typical samples and style label metadata, and supports incremental clustering to dynamically expand the prototype library.

[0027] Similarity calculation module: used to perform standardized preprocessing and feature extraction on new input calligraphy images, fuse shallow and deep features, calculate the similarity distribution between the image and each prototype in the prototype library, and obtain a multi-dimensional similarity vector.

[0028] Feature mapping module: used to dynamically select the optimal style prototype based on similarity distribution, and to project the features of new samples onto the target prototype space through feature mapping algorithm to achieve semantic consistency processing;

[0029] Evaluation module: Used to call the dedicated evaluation sub-model corresponding to the target prototype within the standardized prototype space, and output real-time scores for new samples;

[0030] Optimization module: Used to continuously monitor the deviation between the evaluation results and manual annotation or user feedback. If a new style drift or a decrease in accuracy is detected, it will automatically trigger the expansion of the prototype library and incremental training of sub-models to achieve adaptive optimization of the system's generalization ability.

[0031] The beneficial effects of this invention are as follows:

[0032] This invention provides an automatic identification and evaluation method for calligraphy works. The method includes: collecting, deduplicating, attribute-labeling, and structurally storing brush-written character image samples from different calligraphers, fonts, and styles to form a multi-dimensional labeled original dataset; performing data augmentation on the collected brush-written character images using size normalization, grayscale conversion, various geometric transformations, and noise perturbation, and automatically batch-processing and associating them with labels to obtain a standardized image set with multiple styles and variations; based on the standardized samples, using edge detection and contour analysis to extract stroke-level manual statistical features, and utilizing convolutional neural networks to extract high-dimensional deep features, fusing them to form a unified multi-dimensional style feature vector; performing unsupervised clustering on the fused feature vectors of all samples to establish a high-dimensional style prototype library, where each style prototype contains a central feature... The method includes metadata on eigenvectors, typical samples, and style labels, and supports incremental clustering to dynamically expand the prototype library. It performs standardized preprocessing and feature extraction on new input calligraphy images, fusing shallow and deep features to calculate the similarity distribution between the new image and each prototype in the library, obtaining a multi-dimensional similarity vector. Based on the similarity distribution, it dynamically selects the optimal style prototype and projects the features of the new sample onto the target prototype space using a feature mapping algorithm, achieving semantic consistency. Within the standardized prototype space, it calls a dedicated evaluation sub-model corresponding to the target prototype to provide real-time scoring for the new sample. It continuously monitors the deviation between the evaluation results and manual annotations or user feedback; if new style drift or a decrease in accuracy is detected, it automatically triggers prototype library expansion and incremental sub-model training, achieving adaptive optimization of the system's generalization ability. This method, through multi-dimensional labeling and storage of calligraphers, fonts, and styles, combined with data augmentation using geometric transformations and noise perturbations, enables the model to adapt to variations in different writing styles and avoid overfitting. Furthermore, it can adapt to the evaluation of calligraphy works of different styles and meet the needs of batch evaluation. Attached Figure Description

[0033] Figure 1 The structural diagram of the automatic recognition and evaluation method for calligraphy works provided in the embodiments of this application is shown in the figure. Detailed Implementation

[0034] Preferred embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While preferred embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0035] The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” as used in this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0036] It should be understood that although the terms "first," "second," "third," etc., may be used in this invention to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this invention, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Thus, features defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0037] In traditional calligraphy teaching and evaluation scenarios, manual evaluation mainly relies on calligraphy professionals or experienced enthusiasts. However, manual evaluation has obvious drawbacks: (1) manual evaluation of each character is time-consuming and laborious, making it difficult to meet the needs of large-scale teaching or competition evaluation; (2) manual evaluation is highly subjective, with differences in aesthetic standards and professional levels among different evaluators, resulting in a lack of consistency and objectivity in the evaluation results; (3) manual evaluation usually cannot provide detailed feedback in real time, which is not conducive to learners adjusting their writing habits in a timely manner. In recent years, deep learning has made significant progress in the field of image recognition. Convolutional neural networks (CNNs) such as VGG and ResNet have shown powerful capabilities in feature extraction, and people have also made relevant explorations in the automatic recognition of calligraphy. However, due to the wide variety of calligraphy styles, the existing methods for automatic recognition and evaluation of calligraphy are poorly adapted to complex calligraphy styles and are difficult to capture deep artistic features such as "spirit" and "brushstrokes". Traditional feature extraction methods are greatly affected by font type (regular script, running script, clerical script, etc.) and writing style, and their generalization ability is insufficient in the evaluation of different calligraphers' works or samples of different writing levels.

[0038] Based on this, this application provides an automatic identification and evaluation method for calligraphy works.

[0039] Example

[0040] like Figure 1 As shown, this embodiment provides an automatic recognition and evaluation method for calligraphy works, including:

[0041] S1. Collect, deduplicate, attribute-label, and structure-store brush calligraphy image samples from different calligraphers, fonts, and styles to form a multidimensional labeled original dataset;

[0042] S2. Perform size normalization, grayscale conversion, various geometric transformations and noise perturbation on the collected calligraphy images, and automatically batch process and associate them with labels to obtain a standardized image set with multiple styles and variations.

[0043] S3. Based on standardized samples, edge detection and contour analysis are used to extract stroke-level manual statistical features, and convolutional neural networks are used to extract high-dimensional deep features, which are then fused to form a unified multi-dimensional style feature vector.

[0044] S4. Perform unsupervised clustering on the fused feature vectors of all samples to establish a high-dimensional style prototype library. Each style prototype contains a central feature vector, typical samples, and style label metadata, and supports incremental clustering to dynamically expand the prototype library.

[0045] S5. Perform standardized preprocessing and feature extraction on the newly input calligraphy image, fuse shallow and deep features, calculate the similarity distribution between it and each prototype in the prototype library, and obtain a multidimensional similarity vector.

[0046] S6. Dynamically select the optimal style prototype based on similarity distribution, and project the features of new samples to the target prototype space through a feature mapping algorithm to achieve semantic consistency processing.

[0047] S7. Call the dedicated evaluation sub-model corresponding to the target prototype in the standardized prototype space to output real-time scores for new samples.

[0048] S8. Continuously monitor the deviation between the evaluation results and manual annotation or user feedback. If a new style drift or a decrease in accuracy is detected, automatically trigger the expansion of the prototype library and incremental training of sub-models to achieve adaptive optimization of the system's generalization ability.

[0049] Specifically, the method is implemented as follows:

[0050] A total of 5000 single-character images of Yan Zhenqing's regular script "Duobao Pagoda Stele," Han Dynasty clerical script "Cao Quan Stele," and Qi Gong's running script were collected using a high-definition scanner. After removing duplicate samples using an image hashing algorithm, 4200 images remained. Each image was labeled with attribute tags, including:

[0051] Calligrapher tags: Yan Zhenqing, anonymous Han Dynasty clerical script, Qi Gong, etc.; Font tags: regular script, clerical script, running script; Style feature tags: quantitative parameters such as stroke thickness ratio, structural compactness, and starting stroke arc; The labeled images are stored in a distributed database in a three-level directory structure of "Calligrapher-Font-Style" to form the original dataset.

[0052] To enhance and normalize the original image, perform the following operations:

[0053] Size normalization: uniformly adjusted to 224×224 pixels;

[0054] Grayscale conversion: Convert to an 8-bit grayscale image;

[0055] Geometric transformations: random rotation (±15°), translation (±5 pixels), scaling (0.8-1.2x);

[0056] Noise disturbance: Add Gaussian noise (mean 0, variance 0.05);

[0057] A total of 21,000 enhanced samples were generated and then organized according to the original labeling rules to form a standardized image set.

[0058] Multidimensional style feature vector extraction, including:

[0059] Manual feature extraction: Canny edge detection is used to extract stroke contours, and statistical features such as contour perimeter, area, and concavity are calculated. Hough transform is combined to extract stroke angle distribution.

[0060] Deep feature extraction: Using the ResNet-50 convolutional neural network, extract the 2048-dimensional feature vector output from the 4th convolutional block;

[0061] Feature fusion: The handcrafted features (128 dimensions) are concatenated with the deep features and reduced to 512 dimensions through a fully connected layer to form a unified style feature vector.

[0062] Building a high-dimensional style prototype library:

[0063] The DBSCAN algorithm was used to perform unsupervised clustering on the feature vectors of 4200 original samples, generating 30 style prototypes, each containing:

[0064] Central feature vector: A 512-dimensional vector representing the cluster center;

[0065] Typical sample: 10 representative images closest to the center;

[0066] Metadata tags: such as "Yan style regular script - robust type" and "Qi Gong running script - elegant type"; when 200 new samples of modern calligraphers' running script are added, 5 new prototypes are dynamically added through incremental clustering, expanding the prototype library to 35.

[0067] In steps S5-S6: New sample feature extraction and semantic consistency

[0068] Input an image of the Chinese character "yong" written in a brush style of unknown style. After standardized preprocessing, extract the feature vector, and calculate its cosine similarities with the prototypes of "Regular Script in the Yan Style", "Regular Script in the Liu Style", and "Running Script in the Qi Style" in the prototype library, which are 0.75, 0.62, and 0.41 respectively. Select the prototype of "Regular Script in the Yan Style" with the highest similarity as the target. Project the features of the new sample into the feature space of this prototype through the linear mapping algorithm to eliminate the feature offset caused by style differences.

[0069] S7: Real-time scoring output

[0070] Call the evaluation sub-model corresponding to the prototype of "Regular Script in the Yan Style" (trained based on the BP neural network, and the loss function is the mean square error), score from three dimensions of "stroke integrity" (weight 30%), "structural balance" (weight 40%), and "style consistency" (weight 30%), and output the comprehensive score of 82 points, with the evaluation: "The strokes are powerful, but the arc of the ending of the horizontal stroke slightly deviates from the characteristics of the Yan style."

[0071] S8: System adaptive optimization

[0072] When in 50 consecutive evaluations, the proportion of samples with a deviation between manual annotation and system scoring exceeding 15% reaches 20%, it is detected that the recognition accuracy of samples in the "Modern Wei Bei Style" has decreased. The system automatically triggers the expansion of the prototype library, collects 1000 modern Wei Bei samples for feature extraction, generates a new prototype of "Modern Wei Bei - Powerful Style" through incremental clustering, and performs incremental training on the corresponding sub-model. After optimization, the recognition accuracy of samples in this style has increased from 68% to 89%.

[0073] This method stores multi-dimensional labels such as calligraphers, fonts, and styles, and combines data augmentation of geometric transformation and noise perturbation, enabling the model to adapt to variations in different writing styles and avoid overfitting. Manual features such as edge detection can intuitively reflect the stroke morphology (such as the concavity and convexity of the contour), and the deep features extracted by the convolutional neural network capture the abstract style semantics (such as the compactness of the structure). The feature vector formed by the fusion of the two not only retains the professional semantics of calligraphy evaluation but also has the discrimination ability in the high-dimensional space. The style prototype library constructed by unsupervised clustering can automatically discover potential calligraphy style patterns, and the incremental clustering mechanism supports the system to continuously learn new styles (such as the new modern Wei Bei prototype in the embodiment). When style drift occurs, the expansion of the prototype library enables the system to significantly improve the recognition accuracy of the new style within 3 iteration cycles, avoiding the degradation of recognition ability caused by fixed training data. Matching exclusive evaluation sub-models for different style prototypes (such as the Yan style regular script sub-model focuses on the arc of the stroke, and the running script sub-model emphasizes the coherence of the stroke momentum), combined with the semantic consistency processing of feature mapping, makes the scoring more in line with the calligraphy professional standards. The full-process automated processing of this method from data preprocessing to scoring output can be integrated into scenarios such as calligraphy teaching APPs and intelligent copying systems, providing instant feedback for users and meeting the batch evaluation requirements.

[0074] In this embodiment, the data collection in S1 includes web crawling, API calls and manual collection; the deduplication uses image hash comparison and structural similarity detection; and the attribute annotation includes manual and semi-automatic annotation and expert consensus verification.

[0075] The specific steps are as follows:

[0076] S1.1 filters and accesses data sources from various public calligraphy resource platforms, digital museums, calligraphy textbooks, and social media channels. It uses methods such as web crawling, API calls, or manual collection to obtain brush calligraphy image samples in batches to ensure coverage of a wide range of calligraphers, fonts, and style types, and to obtain multi-source original image data.

[0077] S1.2 performs preliminary screening and deduplication on the collected calligraphy image samples, using algorithms such as image hash comparison and similarity detection to remove duplicate and low-quality samples, thereby improving the effectiveness and representativeness of the original dataset.

[0078] S1.3 Based on expert knowledge base or existing data, each calligraphy image is manually or semi-automatically annotated, and corresponding calligrapher identity, font category (such as regular script, running script, clerical script, etc.) and style feature tags are assigned to achieve accurate binding of image with multi-dimensional attribute information.

[0079] S1.4 The labeled calligraphy images and their related attribute information are organized and stored in a structured manner. A unified data format and naming convention are adopted to import the samples, labels and metadata into a database or distributed storage system, providing an efficient data management foundation for subsequent data preprocessing and analysis.

[0080] In this embodiment, the data enhancement in S2 includes random rotation from -15° to +15°, horizontal or vertical translation from -10 to +10 pixels, scaling ratio from 0.9 to 1.1, horizontal and vertical flipping, and perturbation processing of Gaussian noise (variance 0.001 to 0.01) and salt-and-pepper noise (ratio 0.001 to 0.01).

[0081] The specific steps of S2 are as follows: S2.1 Perform size normalization processing on the collected calligraphy image samples, and use image resampling algorithms such as bilinear interpolation to adjust all images to a uniform preset size (such as 256×256 pixels) to eliminate the interference of the original sample resolution and size difference on subsequent feature extraction and model training, and obtain image output of standardized size.

[0082] S2.2 Based on the size-normalized calligraphy image, apply a grayscale conversion algorithm (such as weighted average method or OpenCV grayscale conversion) to uniformly convert the color or background noise image into a grayscale image, so as to highlight the stroke shape and structural details, reduce the channel dimension, improve the feature extraction efficiency, and output a single-channel grayscale image.

[0083] S2.3 applies various geometric transformation enhancement methods to the standardized grayscale image, including random rotation (within the range of -15° to +15°), translation (randomly selecting a few pixels in the horizontal and vertical directions), scaling (enlarging or shrinking according to a set ratio), and horizontal and vertical flipping. These operations are implemented using image processing libraries such as OpenCV or Pillow to expand the sample space, improve the robustness of the model to different writing postures and spatial deformations, and output a diverse set of enhanced samples.

[0084] S2.4, based on geometric transformation, further performs noise perturbation processing on the standardized grayscale image, such as adding Gaussian noise, salt and pepper noise, etc. By controlling the noise intensity parameter, it simulates the imaging defects and writing material effects in actual acquisition, so as to improve the model's ability to identify low-quality samples in actual applications and obtain an expanded sample set containing noise perturbation.

[0085] S2.5 automatically batch organizes and labels all standardized and enhanced samples generated by the above processing according to a unified naming convention and metadata structure, and maps the original attributes (such as author, font, style tags) to the generated enhanced samples one by one, and stores them in a database or file system, providing high-quality, multi-style, and multi-variant standardized image set input for subsequent feature extraction and modeling stages.

[0086] An automated batch file naming and management method is adopted (parameters: unified naming template [img_id][aug_type][noise_type]_[parameters].png, naming fields strictly distinguish between the original and various enhanced variants) to achieve unique identification encoding for all standardized and enhanced samples, ensuring that the same original sample and all its enhanced derivative samples maintain a hierarchical and easily searchable archiving system in the file system or database.

[0087] Furthermore, through an automatic metadata synchronization and update algorithm (parameters: attribute mapping rules, variant parameter inheritance, and tag extension fields), the original image attributes (such as author_name, font_type, style_tags, etc.) are batch mapped to all enhanced samples. At the same time, extended tags related to the transformation method are automatically added to each enhanced sample (such as aug_type = rotate, noise_type = gauss, param = σ). 20.005), ensuring that all output samples have complete and traceable attribute information.

[0088] Furthermore, a batch association and consistency verification method (parameters: primary key img_id + enhanced parameter composite index, consistency check rules) is adopted to perform one-to-many mapping and binding for each group of original and enhanced variant samples. Database uniqueness constraints or distributed file system metadata verification are used to ensure that all variants under the same original sample strictly correspond in terms of information such as labels and file paths, without omissions or mismatches.

[0089] Furthermore, through efficient data storage and index building algorithms (parameters: relational database table structure definition or NoSQL storage schema, index fields are img_id, author_name, font_type, aug_type, noise_type, etc.), all standardized and enhanced samples along with multi-dimensional attribute labels are automatically imported into the database or distributed file system, and multi-level retrieval indexes are built for key attributes to achieve second-level query and high-concurrency access for large-scale samples.

[0090] Through the aforementioned automated batch processing and label association algorithms, all calligraphy image samples generated after previous normalization, grayscale conversion, and enhancement processing are transformed into a standardized image set with unique identifiers, complete attributes, multi-level enhancement states, and efficient searchable storage structure. This provides high-quality, multi-style, multi-variable, and structurally rigorous data input for subsequent feature extraction and modeling stages, enabling closed-loop management of the data link and full-process traceability.

[0091] Exemplarily, in an actual application scenario, for the 350,400 standardized and enhanced samples of Chinese calligraphy written with a writing brush output after being processed by S2.1 - S2.4, they are batch - named using the naming template "imgid12345_rotate_+8_flipH_gauss0.005.png", where imgid12345 is the primary key of the original image, rotate_+8 represents a rotation of +8°, flipH represents horizontal flipping, and gauss0.005 represents a Gaussian noise variance of 0.005. Through a Python batch - processing script, all enhanced variants inherit the original attributes such as author_name = "Wang Xizhi", font_type = "Regular script", style_tags = ["vigorous"], etc., and automatically add extended fields such as aug_type, noise_type, param, etc. All samples and attribute tags are batch - imported into the MySQL database through SQL, and a unique composite index is established for the combined field of img_id + aug_type + noise_type + param. After consistency verification, there are no duplicate names, missing labels, or mismatches. Finally, 350,400 records of standardized Chinese calligraphy images with full - volume labels and extended attributes are output. Each record can be retrieved in milliseconds through multi - dimensional conditions such as author, font, style, enhancement type, etc. This batch - sorting and label - association process significantly improves data management efficiency and traceability ability, providing a solid data foundation and full - process compliance guarantee for downstream feature extraction, model training, and style generalization evaluation.

[0092] In this embodiment, the manually - counted features in S3 include stroke length, width, curvature, number of strokes, ratio of blank area to stroke area, center - of - gravity position, etc. The global features and deep features are fused through feature splicing after normalization or dimensionality reduction by principal - component analysis.

[0093] The specific steps of S3 are as follows:

[0094] S3.1 Apply an edge - detection algorithm (such as Canny edge detection or Sobel operator) to the standardized Chinese - calligraphy - image samples to obtain the contour information of the strokes, and output a binary edge image for subsequent morphological analysis.

[0095] S3.2 Based on the obtained binary edge image, perform contour - tracking and segmentation operations, and use the image connected - component analysis method to separate and label the contours of each stroke within a single character, realizing the spatial deconstruction of the stroke structure and preparing for stroke - level statistical feature extraction.

[0096] S3.3 For each separated stroke contour, use geometric analysis algorithms to calculate the stroke's length (e.g., cumulative distance through contour points), width (e.g., maximum-minimum cross-sectional distance), curvature (e.g., curve fitting and second derivative solution), and output the stroke-level manual statistical feature vector corresponding to each sample.

[0097] The input data consists of a set of stroke contours for individual Chinese characters after contour tracking and segmentation. Each contour is independently separated, possessing a complete sequence of contour points (spatial coordinate points) and a unique stroke number. The Euclidean distance accumulation method is used to calculate the length of the continuous point sequence ${(x_i,y_i)}_{i=1}^N$ for each stroke contour. The specific formula is as follows:

[0098]

[0099] Where $N$ is the total number of contour sampling points, and $(x_i, y_i)$ is the coordinate of the $i$-th point. This method achieves accurate quantification of the overall direction of the stroke and outputs the actual length value of each stroke. Furthermore, using the maximum and minimum cross-sectional distance method, for each stroke contour, the shortest distance projection algorithm is used to traverse all contour point pairs, find all transverse segments perpendicular to the main direction, and calculate their maximum and minimum widths. The specific formula is as follows:

[0100] W max =max j,k∈S ||(x j ,y j )-(x k ,y k )||W min =min j，k∈S′ ||(x j ,y j )-(x k ,y k )||

[0101] Where $S$ represents the set of all transverse intersection pairs, and $S'$ is the subset of valid transverse intersection pairs. This processing method achieves a comprehensive characterization of the stroke thickness variation range, outputting indicators such as maximum width and minimum width. Furthermore, through curve fitting and the second derivative method, polynomial fitting or B-spline fitting is applied to each stroke contour to obtain smooth curve parameters. Based on the fitted curve...

[0102] Given $f(t) = (x(t), y(t))$, calculate its curvature value at each sampling point. The curvature formula is as follows:

[0103] \kappa(t)=\frac{||x′(t)y”"(t)-y'(t)x"(t)|}{(x′(t)^2+y'(t)^2)^{3 / 2}}

[0104] Where x'(t) and y'(t) are the first derivatives, and x'(t) and y'(t) are the second derivatives. By statistically analyzing the mean, extreme values, and range of curvature, precise quantification of morphological features such as bends and turns in strokes is achieved. Furthermore, statistical parameters such as the length, maximum width, minimum width, average width, mean curvature, extreme values ​​of curvature, and variance of curvature for each stroke are normalized and concatenated into a unified feature vector according to stroke number, ensuring consistency in feature dimensions across different samples. This outputs a manually statistical feature vector at the stroke level for each single-character image. Through the aforementioned chain of geometric analysis and statistical parameter extraction algorithms, the separated stroke contour spatial coordinate data is efficiently transformed into multi-dimensional, standardized stroke-level morphological feature indicators, achieving accurate representation of stroke structure in the quantified space. This provides a solid data foundation for subsequent global structural feature extraction and multi-level style analysis. For example, from 58,400 standardized 256×256 pixel calligraphy images in the database, connected component analysis was performed to separate an average of 9 main stroke contours from each sample. Each contour was sampled using a 128-point sequence, and the average length (range 29-156 pixels) was calculated using the Euclidean distance cumulative method. The maximum width ranged from 3.4 to 19.2 pixels, and the minimum width from 1.2 to 6.7 pixels. Cubic B-spline fitting was applied to each contour, with parameter t uniformly sampled within the [0,1] interval. The curvature values ​​at all sampling points were calculated using first- and second-order difference methods, with a statistical mean of 0.014 / pixel and a maximum of 0.098 / pixel. After Z-score normalization, all parameters were concatenated in stroke order to form a 72-dimensional feature vector (9 strokes per character × 8 parameters). In practical verification, the feature vector exhibits significant differences in distribution within the principal component space for different calligraphic styles (such as regular script, clerical script, and cursive script), which can be used for subsequent style prototype clustering and model training. The above embodiments fully verify the efficiency, accuracy, and discriminative power of the geometric analysis and statistical parameter extraction method in the structural quantification of large-scale, multi-variant calligraphy images, providing highly discriminative morphological features for multi-style adaptive evaluation systems.

[0105] S3.4 further analyzes global morphological features such as the total number of strokes, the ratio of blank area to stroke area, and the position of the center of gravity at the level of the whole character structure. Combined with stroke-level features, it outputs a comprehensive set of handmade morphological features for each sample.

[0106] Specifically, in one particular implementation, step S3.4 is carried out as follows: The input data is a set of stroke contours of a single character image obtained after contour tracking and segmentation processing. Each contour has a complete sequence of spatial coordinate points and a unique stroke number. A stroke connected component counting method (parameters: minimum effective contour area threshold area_min = 15 pixels, maximum effective contour area threshold area_max = 5000 pixels) is used to independently count all major stroke regions in the current sample, outputting an overall stroke count index. Further, using a method to calculate the area ratio of blank areas to stroke regions (parameters: binarization threshold T = 128), after binarizing the standardized grayscale image, the total area corresponding to pixel values ​​of 0 (background blank) and 255 (stroke region) is counted. The blank ratio formula is as follows:

[0107] \text{whitespace ratio}=\frac{A_{bg}}{A_all}]

[0108] Among them, A bg A is the area of ​​blank pixels in the background. all Let A be the total pixel area of ​​the entire image. Further, by using a stroke region area statistical method, the pixels within all separated stroke outlines are accumulated to obtain the total stroke area A. stroke The area ratio between the stroke region and the blank area is obtained by combining the blank area. Furthermore, the centroid coordinates of the image are solved using a weighted average method through a centroid position calculation method (parameters: the set of all stroke contour points (xi,yi)). The formula is as follows:

[0109]

[0110] Where N represents the total number of stroke pixels, and (xi,yi) represents the spatial coordinates of each point. Further, the centroid position is normalized, mapping its coordinates to the [0,1] interval to facilitate unified comparative analysis between samples of different sizes and structures. Through the above multi-index statistical analysis, the global morphological features such as the overall number of strokes at the whole-character structure level, the ratio of blank areas to stroke areas, and the normalized centroid position are combined with the stroke-level feature set obtained in the previous step to output a comprehensive manual morphological feature set for each calligraphy sample, achieving a comprehensive quantitative representation of character structure, spatial layout, and overall balance. For example, in a practical application scenario, the Open CVfindContours method effectively separates nine main stroke contours from 58,400 standardized 256×256 pixel calligraphy images in the database. After accumulating all pixels within the contour, the average area of ​​a single character stroke region was calculated to be 5120 pixels, the background blank area area to be 59536 pixels, and the total image area to be 65536 pixels. The blank area ratio was calculated to be 0.908 using the formula. The spatial mean of all stroke points was calculated using a coordinate-weighted averaging method, yielding the centroid coordinates (124.3, 130.7), which were then normalized to (0.486, 0.510). For different fonts, the centroid distribution of the regular script samples remained stable between (0.48 and 0.52), while the centroid of the clerical script samples fluctuated more significantly. The final output for each sample included more than 10 global morphological features such as the total number of strokes, blank area ratio, stroke area ratio, and normalized centroid, which were concatenated to the original stroke-level feature vector. This provided a highly discriminative comprehensive morphological foundation for subsequent deep feature fusion and style clustering modeling. Actual testing showed that after introducing the above global morphological features, the multi-style prototype clustering discrimination rate improved by 6%, and the system's robustness to structurally drifting fonts was significantly enhanced.

[0111] S3.5 inputs standardized grayscale calligraphy images into a pre-trained convolutional neural network (such as VGG or ResNet), uses the output of intermediate layers as a high-dimensional deep feature representation, extracts the deep semantic feature vector of each sample through forward inference, and outputs feature data containing complex information such as style and structure.

[0112] S3.6 normalizes the manual statistical feature vectors and deep semantic feature vectors, and then fuses them through feature concatenation or dimensionality reduction fusion algorithms (such as principal component analysis PCA or feature weighted concatenation) to obtain a multi-dimensional fused feature vector for each sample, which serves as the basic input for subsequent style clustering and evaluation modeling.

[0113] In this embodiment, the unsupervised clustering in S4 adopts K-means, DBSCAN or Gaussian mixture model, and the style prototype library supports structured storage and multi-level index retrieval of central feature vectors, typical samples, main font categories and calligrapher label distribution.

[0114] The specific steps of S4 are as follows: S4.1 Take the fused feature vector of all standardized calligraphy image samples as input, and use principal component analysis (PCA) or t-SNE and other dimensionality reduction algorithms to perform dimensionality reduction preprocessing on the high-dimensional features in order to reduce the computational complexity of subsequent clustering analysis and highlight the style differences between samples, thereby achieving effective simplification of the feature space.

[0115] S4.2 Based on the dimensionality-reduced fused feature vector, it applies an unsupervised clustering algorithm (such as K-means, DBSCAN or Gaussian mixture model) to automatically analyze and divide several style prototype groups. Each prototype group aggregates samples with highly similar style features and outputs the cluster center points and corresponding sample allocation results.

[0116] S4.3 For the sample set within each style prototype group, statistically analyze and save representative feature information, including the group's central feature vector (cluster center), typical sample index, main font categories, and calligrapher label distribution, to form discriminative style label metadata.

[0117] S4.4 unifies the data structures of all style prototype groups (central feature vector, typical sample set, style label, and representative metadata) into a high-dimensional style prototype library, enabling queryable and scalable management of attributes between prototypes and providing a standardized interface for subsequent new sample similarity retrieval and dynamic mapping.

[0118] S4.5 periodically performs incremental clustering analysis on newly added fused feature vectors. By dynamically updating the cluster centers and prototype group structure, it expands and optimizes the style prototype library in a timely manner, ensuring that it continuously reflects the full picture of the diversity of current calligraphy styles as data accumulates.

[0119] In this embodiment, the similarity measure in S5 includes cosine similarity, Euclidean distance and Mahalanobis distance, and outputs a multidimensional similarity distribution vector of the new sample in the prototype space.

[0120] The specific steps of S5 are as follows: S5.1 Perform preprocessing operations on the newly input calligraphy image samples, including size normalization (e.g., adjusting to 256×256 pixels) and grayscale processing. Use the same image preprocessing algorithm as historical samples to ensure the consistency of input features and provide standardized input images for subsequent feature extraction steps.

[0121] S5.2 Based on the standardized new sample images, it applies a preset edge detection algorithm (such as Canny edge detection) and contour analysis method to extract stroke-level morphological features (such as length, width, curvature, etc.) and statistically analyze global structural features (such as number of strokes, blank ratio, center of gravity position, etc.) to obtain shallow manual statistical feature vectors.

[0122] S5.3 inputs a standardized grayscale image into a pre-trained convolutional neural network model (such as ResNet or VGG), uses the output of intermediate layers as high-dimensional deep features, and automatically extracts feature vectors containing complex semantic information such as style and structure through forward inference, thereby realizing automated representation of deep features.

[0123] S5.4 normalizes the shallow manual statistical features and deep convolutional features of the new samples, and uses feature splicing or dimensionality reduction fusion algorithms (such as principal component analysis PCA or feature weighted fusion) to output a multi-dimensional fusion style feature vector in a unified format, in preparation for subsequent similarity calculation.

[0124] S5.5 takes the style feature vector of the fused new sample as input, calls a similarity measurement algorithm (such as cosine similarity, Euclidean distance or Mahalanobis distance), calculates the similarity value between the vector and each prototype node in the style prototype library, and outputs the multidimensional similarity distribution vector of the new sample in the prototype space.

[0125] In this embodiment, the feature mapping algorithm in S6 includes linear projection, subspace alignment, or adaptive normalization mapping, which supports the consistent alignment of new sample features with the target prototype space.

[0126] The specific steps of S6 are as follows:

[0127] S6.1 takes the similarity distribution vector of the new sample in the style prototype library as input, uses the maximum similarity search or soft allocation algorithm to sort all style prototype nodes, dynamically selects the best matching target style prototype or weighted combination prototype to determine the reference space for subsequent feature mapping, and outputs the target prototype identifier and its related attributes.

[0128] S6.2 Based on the new sample fusion feature vector and the central feature vector of the selected target style prototype, a feature mapping algorithm (such as linear projection, subspace alignment, kernel method or adaptive normalization mapping) is called to perform spatial transformation processing on the new sample features to project them to the target prototype space, realize the alignment and standardization of multiple style features in the same semantic space, and output the projected standardized feature vector.

[0129] S6.3 performs consistency checks on the standardized feature vectors of the mapped new samples, using methods such as distance metrics, reconstruction error detection, or distribution statistical analysis to evaluate their matching degree with the target prototype space. If the projection deviation is found to exceed the threshold, a local fine-tuning algorithm (such as adaptive parameter optimization or local retraining) is triggered to ensure the effectiveness and accuracy of cross-style semantic consistency processing, and finally outputs standardized feature data that can be used for downstream model evaluation.

[0130] Furthermore, the dedicated evaluation sub-model in S7 is trained through a dedicated neural network or ensemble learning algorithm and can be independently incrementally optimized according to the style prototype, outputting a multi-dimensional score including stroke quality, structural rationality, and aesthetics.

[0131] The specific steps of S7 are as follows:

[0132] S7.1 automatically retrieves and matches the unique identifier of the selected style prototype in the prototype library for the new sample feature vector after standardization mapping, and uses it as the entry point for calling the dedicated evaluation sub-model to ensure that the subsequent scoring model has relevance and style adaptability.

[0133] S7.2 Based on the selected style prototype identifier, the standardized feature vector of the new sample is input into the corresponding dedicated evaluation sub-model (such as a scoring sub-model based on a neural network or ensemble learning algorithm), and forward inference or feature mapping is performed to obtain a multi-dimensional score prediction output under the style prototype.

[0134] S7.3 performs post-processing and standardized mapping on the evaluation dimensions such as stroke quality, structural rationality, and overall aesthetics according to the system's preset scoring rules and indicator system for each of the original scoring results output by the dedicated evaluation sub-model, ensuring that the scoring results are comparable and uniformly interpretable across different style prototypes.

[0135] S7.4 will generate a standardized output data package containing detailed sub-item scores and comprehensive evaluations based on the final processed multi-dimensional scoring results and according to the interface specifications and display requirements. This data package will serve as the evaluation result for real-time feedback to users or downstream business modules (such as interactive interfaces and data analysis platforms).

[0136] In this embodiment, the evaluation result monitoring in S8 adopts error analysis, statistical process monitoring and novelty detection, automatically identifies new style drift and incrementally expands the prototype library and optimizes sub-model parameters.

[0137] The specific steps of S8 are as follows:

[0138] S8.1 compares the automatic evaluation results of newly input calligraphy samples with manual annotation scores or actual user feedback data in real time. It uses error analysis algorithms (such as mean absolute error MAE, mean square error MSE, etc.) to calculate the scoring deviation on each evaluation dimension in order to obtain the quantitative difference between the current output of the evaluation model and the manual benchmark.

[0139] S8.2 Based on the scoring deviation data obtained from the comparison, statistical process monitoring methods (such as sliding window mean, CUSUM control chart, etc.) are applied to dynamically analyze the recent evaluation accuracy and distribution trend, identify whether there are generalization failure signs such as new style drift (such as the model's scoring error on a certain type of new style sample is continuously abnormal) or overall accuracy decline, so as to realize anomaly detection and alarm.

[0140] When S8.3 detects a new style drift or a decrease in model accuracy, it aggregates and organizes the relevant abnormal samples, and uses a feature similarity clustering algorithm (such as K-means, DBSCAN, etc.) to analyze the style feature distribution of the abnormal samples in order to determine whether the style prototype library needs to be expanded and to determine the prototype types and representative sample sets that need to be added.

[0141] For style categories that need to be expanded, S8.4 integrates the label, attribute and feature data of representative abnormal samples obtained from clustering, and uses the original prototype library construction process to perform style prototype node addition and central feature vector update operations to achieve adaptive expansion and diversity enhancement of the style prototype library structure.

[0142] Based on the expanded style prototype library, S8.5 performs incremental training of dedicated sub-models on relevant new style sample sets and their historical scoring data. It uses online learning algorithms or fine-tuning mechanisms to optimize and update the weights of the target prototype sub-model, thereby improving its scoring accuracy for new style samples and outputting dedicated sub-model parameters with enhanced generalization ability.

[0143] S8.6 synchronizes the updated style prototype library and parameters of each dedicated evaluation sub-model to the system's main control module, and automatically resumes the normal real-time evaluation process for subsequent new samples, ensuring that the system can adaptively respond to continuously emerging new style types, and achieve dynamic improvement and stable output of generalization capabilities.

[0144] Example 2

[0145] This embodiment provides an automatic calligraphy work identification and evaluation system for implementing the calligraphy work automatic identification and evaluation method described above. The system includes:

[0146] Data acquisition module: used to collect, deduplicate, attribute-label, and structure the image samples of calligraphy from different calligraphers, fonts, and styles, forming a multidimensional labeled raw dataset;

[0147] Data preprocessing module: used to perform size normalization, grayscale conversion, various geometric transformations and noise perturbation on the acquired calligraphy images, and automatically batch process and associate them with labels to obtain a standardized image set with multiple styles and variations;

[0148] Feature extraction module: Based on standardized samples, it uses edge detection and contour analysis to extract stroke-level manual statistical features, and uses convolutional neural networks to extract high-dimensional deep features, which are then fused to form a unified multi-dimensional style feature vector;

[0149] Prototype library building module: used to perform unsupervised clustering on the fused feature vectors of all samples to build a high-dimensional style prototype library. Each style prototype contains a central feature vector, typical samples and style label metadata, and supports incremental clustering to dynamically expand the prototype library.

[0150] Similarity calculation module: used to perform standardized preprocessing and feature extraction on new input calligraphy images, fuse shallow and deep features, calculate the similarity distribution between the image and each prototype in the prototype library, and obtain a multi-dimensional similarity vector.

[0151] Feature mapping module: used to dynamically select the optimal style prototype based on similarity distribution, and to project the features of new samples onto the target prototype space through feature mapping algorithm to achieve semantic consistency processing;

[0152] Evaluation module: Used to call the dedicated evaluation sub-model corresponding to the target prototype within the standardized prototype space, and output real-time scores for new samples;

[0153] Optimization module: Used to continuously monitor the deviation between the evaluation results and manual annotation or user feedback. If a new style drift or a decrease in accuracy is detected, it will automatically trigger the expansion of the prototype library and incremental training of sub-models to achieve adaptive optimization of the system's generalization ability.

[0154] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for automatically identifying and evaluating calligraphy works, characterized in that, include: S1. Collect, deduplicate, attribute-label, and structure-store brush calligraphy image samples from different calligraphers, fonts, and styles to form a multidimensional labeled original dataset; S2. Perform size normalization, grayscale conversion, geometric transformation and noise perturbation on the collected calligraphy images, and automatically batch process and associate them with labels to obtain a standardized image set with multiple styles and variations; S3. Based on standardized samples, edge detection and contour analysis are used to extract stroke-level manual statistical features, and convolutional neural networks are used to extract high-dimensional deep features, which are then fused to form a unified multi-dimensional style feature vector. S4. Perform unsupervised clustering on the fused feature vectors of all samples to establish a high-dimensional style prototype library. Each style prototype contains a central feature vector, typical samples, and style label metadata, and supports incremental clustering to dynamically expand the prototype library. S5. Perform standardized preprocessing and feature extraction on the newly input calligraphy image, fuse shallow and deep features, calculate the similarity distribution between it and each prototype in the prototype library, and obtain a multidimensional similarity vector. S6. Dynamically select the optimal style prototype based on similarity distribution, and project the features of new samples to the target prototype space through a feature mapping algorithm to achieve semantic consistency processing. S7. Within the standardized prototype space, call the dedicated evaluation sub-model corresponding to the target prototype to output real-time scores for new samples.

2. The automatic recognition and evaluation method for calligraphy works as described in claim 1, characterized in that: After the real-time scoring output of the new sample, the method also includes: S8. Continuously monitor the deviation between the evaluation results and manual annotation or user feedback. If a new style drift or a decrease in accuracy is detected, automatically trigger the expansion of the prototype library and incremental training of sub-models to achieve adaptive optimization of the system's generalization ability.

3. The automatic recognition and evaluation method for calligraphy works as described in claim 1, characterized in that: The data collection in S1 includes web crawling, API calls, and manual collection. The deduplication uses image hash comparison and structural similarity detection. The attribute annotation includes manual and semi-automatic annotation and expert consensus verification.

4. The automatic recognition and evaluation method for calligraphy works as described in claim 1, characterized in that: The data enhancements in S2 include random rotation from -15° to +15°, horizontal or vertical translation from -10 to +10 pixels, scaling from 0.9 to 1.1, horizontal and vertical flipping, and perturbation processing of Gaussian noise (and salt-and-pepper noise).

5. The automatic recognition and evaluation method for calligraphy works as described in claim 2, characterized in that: The manual statistical features in S3 include stroke length, width, curvature, number of strokes, ratio of blank space to stroke area, and center of gravity position. Global features and deep features are fused by normalized feature splicing or principal component analysis for dimensionality reduction.

6. The automatic recognition and evaluation method for calligraphy works as described in claim 1, characterized in that: The unsupervised clustering in S4 uses K-means, DBSCAN, or Gaussian mixture models. The style prototype library supports structured storage and multi-level index retrieval of central feature vectors, typical samples, main font categories, and calligrapher label distributions.

7. The automatic recognition and evaluation method for calligraphy works as described in claim 6, characterized in that: The similarity metrics in S5 include cosine similarity, Euclidean distance, and Mahalanobis distance, and output a multidimensional similarity distribution vector of the new sample in the prototype space.

8. The automatic recognition and evaluation method for calligraphy works as described in claim 3, characterized in that: The feature mapping algorithm in S6 includes linear projection, subspace alignment, or adaptive normalization mapping, which supports the consistent alignment of new sample features with the target prototype space.

9. The automatic recognition and evaluation method for calligraphy works as described in claim 8, characterized in that: The dedicated evaluation sub-model in S7 is trained through a dedicated neural network or ensemble learning algorithm and can be independently incrementally optimized according to style prototype, outputting a multi-dimensional score including stroke quality, structural rationality and aesthetics.

10. The automatic recognition and evaluation method for calligraphy works as described in claim 2, characterized in that: The evaluation result monitoring in S8 employs error analysis, statistical process monitoring, and novelty detection to automatically identify new style drift and incrementally expand the prototype library and optimize sub-model parameters.

Images