Painting and calligraphy inscription feature reconstruction and facsimile guidance system and method based on AI vision
By using an AI vision-based system to perform multi-scale similarity comparison and feature acquisition of calligraphy, painting, and rubbings, the problem of insufficient accuracy in calligraphy, painting, and rubbings copying guidance has been solved, and highly accurate copying guidance has been achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-05
- Publication Date
- 2026-03-10
AI Technical Summary
Existing calligraphy and painting rubbings copying guides lack precision. Human guidance is greatly affected by subjective factors, making it difficult to fully and meticulously capture micro-level characteristics and in-depth information, thus failing to meet the demand for high-quality copying guidance.
The system employs an AI vision-based approach, using a multi-scale similarity comparison module to compare original calligraphy and painting rubbings with standard image data. It also uses multimodal imaging equipment to acquire brush and ink textures and compositional features, searches for creative techniques and style characteristics, and provides guidance and prompts for copying.
It improves the accuracy of copying guidance and achieves objective and consistent high-quality copying guidance.
Smart Images
Figure CN121640499A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of computer vision, and particularly relates to an AI vision-based calligraphy and painting tablet feature reconstruction and copying guidance system and method. BACKGROUND
[0002] Accurate copying of calligraphy and painting tablets is a core link in the inheritance and learning of traditional art, and the key lies in meticulous grasp of the classic model ink technique, layout and style charm. At present, in the field of calligraphy and painting tablet copying guidance, traditional manual guidance methods are mainly relied on, that is, experienced calligraphers or professionals observe the original works and copying works by naked eyes, and give copying suggestions and guidance according to their own experience. However, the manual guidance method has many limitations. On the one hand, manual observation is greatly affected by subjective factors. Different guides may give different guidance opinions due to differences in experience and aesthetic concepts, and lack of objectivity and unity. On the other hand, manual observation is difficult to fully and carefully capture micro-level features and deep-level information, resulting in limited accuracy of copying guidance and failing to meet the needs of learners for high-quality copying guidance.
[0003] At present, in the related art, the calligraphy and painting tablet copying guidance has the technical problem of insufficient accuracy. SUMMARY
[0004] The present application provides an AI vision-based calligraphy and painting tablet feature reconstruction and copying guidance system and method. By using an AI vision acquisition device to scan the target calligraphy and painting tablet original work, comparing it with standard image data in multiple scales, using a multi-modal imaging device to obtain the original work ink texture and layout features, searching for original work creation technique parameters and style feature parameters, and giving copying guidance prompts based on the above information, the technical problem of insufficient accuracy of existing calligraphy and painting tablet copying guidance is solved, and the technical effect of improving the accuracy of copying guidance is achieved.
[0005] The application provides an AI vision-based calligraphy and painting feature reconstruction and copying guidance system, comprising: a multi-scale similarity comparison module, configured to use an AI vision acquisition device to scan a target calligraphy and painting original work, and perform multi-scale similarity comparison on high-definition image data obtained by the AI vision acquisition device and standard image data stored by a historical classic model; a feature synchronous acquisition module, configured to use a multi-modal imaging device to synchronously acquire first features corresponding to ink texture of the target calligraphy and painting original work and second features corresponding to chapter layout; a creation technique parameter search module, configured to search for creation technique parameters corresponding to the target calligraphy and painting original work; a style feature parameter search module, configured to search for style feature parameters corresponding to the target calligraphy and painting original work; and a copying guidance prompt module, configured to perform copying guidance prompt based on multi-scale similarity comparison information, the first features corresponding to ink texture and the second features corresponding to chapter layout, in combination with the creation technique parameters and the style feature parameters.
[0006] In a possible implementation, the multi-scale similarity comparison module comprises: an image pyramid processing module, configured to perform image pyramid processing on the high-definition image data and the standard image data to generate a plurality of resolution image layers; an image feature descriptor determination module, configured to determine image feature descriptors under each resolution image layer in the plurality of resolution image layers, the image feature descriptors comprising at least one of ink stroke direction, line thickness variation and ink color gradient; and a similarity comparison module, configured to perform multi-scale similarity comparison of the image feature descriptors at different scales.
[0007] In a possible implementation, the similarity comparison module comprises: a comparison index setting module, configured to set a comparison index, and the multi-scale similarity comparison index comprises any one or more of stroke form consistency, line curvature matching degree and ink color distribution similarity.
[0008] In a possible implementation, the feature synchronous acquisition module comprises: a first feature acquisition module, configured to capture a detail magnified image of the target calligraphy and painting original work, and extract stroke pressure variation features, ink color gradation features and line turning features as the first features corresponding to ink texture; and a second feature acquisition module, configured to capture an overall layout image of the target calligraphy and painting original work, and extract word spacing and line spacing features, composition balance features and sparse distribution features as the second features corresponding to chapter layout.
[0009] In a possible implementation, the copying guidance prompt module includes: a technique and style state space construction module, used to construct a technique and style state space through creation technique parameters and style feature parameters; a copying deviation probability derivation module, used to deduce the copying deviation probability in the technique and style state space to obtain the copying deviation type and deviation probability value; and a judgment module, used to determine whether to trigger the copying guidance prompt based on the copying deviation type and deviation probability value.
[0010] In a possible implementation, the technique style state space construction module includes: a creative technique parameter determination module, used to determine the creative technique parameters corresponding to the target calligraphy and painting original work, including brushwork characteristics, ink rules, and character structure paradigms; and a style feature parameter determination module, used to determine the style feature parameters corresponding to the target calligraphy and painting original work, including brushstroke strength characteristics, composition and flow characteristics, and individual style identifiers.
[0011] In a possible implementation, the copying guidance prompt module includes: a preliminary deviation range acquisition module, used to obtain a preliminary deviation range through the multi-scale similarity comparison information if the copying guidance prompt is triggered; and a boundary correction module, used to correct the boundary of the preliminary deviation range according to the first feature corresponding to the ink texture, and to correct the boundary of the preliminary deviation range according to the second feature corresponding to the composition layout.
[0012] In a possible implementation, the boundary correction module includes: a semantic segmentation module, used to perform semantic segmentation on the magnified detail image within the initially defined deviation range, dividing the image into multiple semantic blocks with similar brushstroke features; a semantic block feature vector construction module, used to extract brushstroke trajectory features, ink density features, and texture roughness features of the multiple semantic blocks, and construct semantic block feature vectors; and an initial deviation range correction module, used to correct the boundary of the initially defined deviation range using the semantic block feature vectors and brushstroke anomaly features of the deviation area.
[0013] In a possible implementation, the preliminary deviation range correction module includes: a similarity threshold definition module, used to define a similarity threshold based on the cosine similarity between the semantic block feature vector and the standard pen stroke feature vector; a deviation region marking module, used to traverse the multiple semantic blocks, and if the cosine similarity is less than the similarity threshold, to mark the deviation region and extract the edge features of the deviation region; a direction vector field generation module, used to generate a direction vector field of edge pixels based on the brush stroke breakpoints and morphological abrupt change points associated with the edge features of the deviation region; and a brush stroke abnormality feature acquisition module, used to determine the angular distribution with respect to the standard brush stroke direction based on the direction vector field of the edge pixels, thereby obtaining the brush stroke abnormality features.
[0014] This application also provides a method for feature reconstruction and copying guidance of calligraphy and painting rubbings based on AI vision, including: scanning the original target calligraphy and painting rubbings using an AI vision acquisition device; performing multi-scale similarity comparison between the high-definition image data obtained by the AI vision acquisition device and the standard image data stored in historical classic models; using a multi-modal imaging device to simultaneously acquire the first feature corresponding to the brush and ink texture and the second feature corresponding to the composition layout of the original target calligraphy and painting rubbings; searching for the creative technique parameters corresponding to the original target calligraphy and painting rubbings; searching for the style feature parameters corresponding to the original target calligraphy and painting rubbings; and providing copying guidance based on the multi-scale similarity comparison information, the first feature corresponding to the brush and ink texture, the second feature corresponding to the composition layout, and the creative technique parameters and style feature parameters.
[0015] This application proposes a system and method for reconstructing and guiding the copying of calligraphy and painting rubbings based on AI vision. The system utilizes a multi-scale similarity comparison module to scan the original calligraphy and painting rubbings using an AI vision acquisition device. The high-resolution image data obtained from the AI vision acquisition device is then compared with standard image data stored in historical classic models at multiple scales. A feature synchronization acquisition module uses a multi-modal imaging device to simultaneously acquire the first feature corresponding to the brushstroke texture and the second feature corresponding to the composition layout of the original calligraphy and painting rubbings. A creative technique parameter search module searches for the corresponding creative technique parameters of the original calligraphy and painting rubbings, and a style feature parameter search module searches for the corresponding style feature parameters. Finally, a copying guidance module provides copying guidance based on the multi-scale similarity comparison information, the first feature corresponding to the brushstroke texture, the second feature corresponding to the composition layout, and the creative technique parameters and style feature parameters. This achieves the technical effect of improving the accuracy of copying guidance. Attached Figure Description
[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.
[0017] Figure 1 This is a schematic diagram of the structure of the AI vision-based calligraphy and painting rubbing feature reconstruction and copying guidance system provided in the embodiments of this application.
[0018] Figure 2 This is a flowchart illustrating the AI vision-based method for reconstructing and guiding the copying of calligraphy and painting rubbings, provided in an embodiment of this application.
[0019] Figure labeling: Multi-scale similarity comparison module 10, feature synchronous acquisition module 20, creative technique parameter search module 30, style feature parameter search module 40, and imitation guidance and prompts module 50. Detailed Implementation
[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.
[0021] This application provides an AI-based vision-based system for reconstructing and guiding the copying of calligraphy and painting rubbings, such as... Figure 1 As shown, the system includes: The multi-scale similarity comparison module 10 is used to scan the original target calligraphy and painting rubbings using an AI vision acquisition device, and to perform multi-scale similarity comparison between the high-definition image data obtained by the AI vision acquisition device and the standard image data stored in historical classic models.
[0022] Specifically, a high-resolution linear array CCD scanner is used to scan the original target calligraphy and painting rubbings, acquiring digital images with a resolution of no less than 600 DPI. The images are then subjected to grayscale conversion, Gaussian filtering for noise reduction, and affine transformation-based geometric correction to eliminate errors caused by shooting angle or paper distortion. The high-resolution scanned image data is compared with standard image data stored as historical classic models using a multi-scale similarity comparison. A Gaussian pyramid algorithm is used to downsample both the preprocessed high-resolution and standard images. For example, starting with the original image as layer 0, each subsequent layer reduces the image's width and height to half of the next layer's, and convolution is performed using a Gaussian kernel to eliminate aliasing effects caused by downsampling, ultimately generating an image pyramid containing 4 to 6 resolution layers, from the original resolution to extremely low resolution. The multi-scale feature matching process begins at the top of the pyramid, i.e., the lowest resolution, and proceeds layer by layer downwards. At the top layer, rapid template matching or global feature matching is performed to quickly locate the approximate corresponding regions of the two images, providing a rough initial position. As the number of layers decreases, the resolution increases. Then, within the coarse area defined by the previous layer, more refined features are used for matching, gradually refining the process.
[0023] In one possible implementation, the multi-scale similarity comparison module 10 includes: an image pyramid processing module, used to perform image pyramid processing on the high-definition image data and standard image data to generate multiple resolution image layers; an image feature descriptor determination module, used to determine the image feature descriptors under each resolution image layer in the multiple resolution image layers, the image feature descriptors including at least one of brush stroke direction, line thickness variation, and ink density gradient; and a similarity comparison module, used to perform multi-scale similarity comparison at different scales using the image feature descriptors.
[0024] Specifically, the image pyramid processing module is responsible for constructing the pyramid, which involves recursively performing Gaussian blur and downsampling. Gaussian blur applies a fixed-size two-dimensional Gaussian convolution kernel (e.g., 5×5 pixels) to the current layer image to smooth the image and reduce high-frequency information. Downsampling involves sampling the Gaussian-blurred image every other row and column, i.e., taking one sample every two pixels, resulting in an image with the original dimensions. Figure 1 A new image of 2 / 2 is created as the next layer of the pyramid. This process is repeated until the preset number of layers is reached.
[0025] The image feature descriptor determination module calculates mathematical vectors, i.e., image feature descriptors, that quantify the ink stroke features on each layer of the pyramid image. Specifically, this includes: extracting ink stroke direction features using a Gabor filter bank. For example, 32 Gabor filters (4 scales, 8 directions, total) are used to filter the image. Each filter produces a strong response in its specific direction. For each pixel in the image, the direction of the filter with the strongest response is taken as the principal direction of that point. A direction histogram is statistically analyzed within a certain neighborhood of each pixel to represent the features of that local region. Line thickness variation features are extracted using a combination of image skeletonization and distance transform. First, a thinning algorithm is applied to the binarized stroke image to obtain a single-pixel-wide stroke skeleton. Then, a distance transform is applied to the original stroke region. The distance transform value of each point on the skeleton line reflects the distance information from that point to the stroke boundary. By statistically analyzing the distribution of these distance values on the skeleton line, such as the mean and variance, thickness variation features are constructed. For ink density gradient features, the Sobel operator is used on the grayscale image to calculate the brightness gradient of each pixel in the X and Y directions, thus obtaining the gradient magnitude and direction. Among them, by statistically analyzing the distribution of gradient amplitude, the degree of intensity of ink color change can be reflected, with a large gradient indicating a clear boundary between light and dark areas.
[0026] The similarity comparison module is responsible for performing the specific comparison operations, calculating the differences between the feature descriptors of the copied image and the standard image at each pyramid level. Specifically, each of the above features, including orientation histogram, coarse-thickness distribution, gradient histogram, etc., is normalized and concatenated into a high-dimensional feature vector. The similarity between the two feature vectors is calculated using cosine similarity or Euclidean distance. Each pyramid level yields a similarity score, and the final total similarity is a weighted average of these level scores, with higher weights given to the middle levels because they offer the best balance between feature stability and discriminative power.
[0027] In one possible implementation, the similarity comparison module includes: a comparison index setting module, used to set comparison indexes, wherein the multi-scale similarity comparison indexes include any one or more of the following: brushstroke shape consistency, line curvature matching degree, and ink color distribution similarity.
[0028] Specifically, the comparison index setting module is used to set specific, quantifiable index thresholds for similarity calculation. Among them, the brushstroke shape consistency index is obtained by comparing the Bach coefficients of the Gabor orientation histograms of two images. The closer the coefficient is to 1, the more consistent the orientation distribution.
[0029] A continuous sequence of points is extracted from the stroke skeleton, fitted with a spline curve, and then the curvature of the fitted curve is calculated. The curvature sequences of the two skeleton curves are compared using a dynamic time warping algorithm, and the cumulative distance D of the warped path is calculated. The line curvature matching degree M = 1 / (1+D). The closer the M value is to 1, the higher the line curvature matching degree.
[0030] Calculate the joint histogram of the two images in the V lightness component of the HSV color space, and then calculate their chi-square distance. The smaller the distance, the more similar the ink color distribution.
[0031] The feature synchronous acquisition module 20 is used to simultaneously acquire the first feature corresponding to the brush and ink texture and the second feature corresponding to the composition layout of the target calligraphy and painting original work using a multimodal imaging device.
[0032] Specifically, the system integrates a high-resolution RGB color camera and a high-precision 3D contour scanner. The two devices are connected via a hardware synchronization line, and a single trigger signal is sent from the computer to ensure that 2D color information and 3D shape information of the same area are captured simultaneously. Because the two devices have different perspectives, the captured data requires rigid body transformation registration. By placing a standard calibration plate on the scanning platform and simultaneously capturing data from both sensors, the coordinate transformation matrix between the RGB camera and the 3D contour scanner is calculated. Afterward, each 3D point cloud can find its corresponding RGB pixel value.
[0033] In one possible implementation, the feature synchronous acquisition module 20 includes: a first feature acquisition module, used to capture a magnified image of the details of the original calligraphy and painting work, and extract brush pressure change features, ink color diffusion features and line turning features as first features corresponding to the brush and ink texture; and a second feature acquisition module, used to capture an overall layout image of the original calligraphy and painting work, and extract character spacing and line spacing features, composition balance features and density distribution features as second features corresponding to the composition layout.
[0034] Specifically, the first feature acquisition module extracts the profile height data of the center line of the stroke trajectory from the 3D point cloud data. The deeper the indentation in the paper, the greater the pressure applied during the stroke. The feature of stroke pressure variation is quantified into a pressure variation curve along the stroke. In the stroke edge region of the RGB image, the image is divided into small blocks using a superpixel segmentation algorithm. Then, the attenuation pattern of ink brightness value from the inside of the stroke to the outside of the paper in each small block is analyzed and fitted into an exponential decay function. The decay coefficient of the function is the quantized value of the ink bleed feature. Frechet distance calculation or corner point detection is performed on the skeleton line of the stroke to identify points on the skeleton line where the curvature suddenly increases. These points are the turning points. The sharpness of the turning point is represented by the curvature value of that point.
[0035] The second feature acquisition module uses connected component analysis to locate and label each character in the image, and then calculates the bounding rectangle of each connected component. Character spacing is the horizontal distance between adjacent rectangles in the same row; line spacing is the vertical distance between adjacent rectangles in the same row. Character spacing and line spacing features are described by statistically analyzing the mean and variance of the entire text's character spacing and line spacing. For compositional balance features, the entire work is treated as a mechanical system, calculating the centroid coordinates of all stroke pixels, and then calculating the offset of this centroid from the geometric center of the canvas. The smaller the offset, the more balanced the composition. Simultaneously, the stroke pixel density on both sides of the horizontal and vertical central axes of the image is calculated to assess whether the visual weight is balanced. For density distribution features, the image is divided into a regular grid, such as 10×10, and the area ratio of stroke pixels within each grid is calculated. Then, a two-dimensional discrete Fourier transform is performed on this density matrix to analyze the energy distribution of its low-frequency and high-frequency components to quantify the overall density rhythm.
[0036] The creative technique parameter search module 30 is used to search for the creative technique parameters corresponding to the original work of the target calligraphy, painting and inscription.
[0037] Specifically, a database containing numerous classic calligraphic works and their known creative technique labels is pre-established, such as the use of the central brushstroke, the side brushstroke for elegance, and the dry ink flying white technique. For each work in the database, a feature vector is extracted using the feature synchronization acquisition module 20. When the feature vector of the target calligraphy or painting original is input, a K-nearest neighbor search is performed in the technique database to find the K samples whose feature vectors are closest. Combining the technique labels of these K nearest neighbor samples, the most likely creative technique parameters used in the target calligraphy or painting original are determined through a voting method or a distance-weighted voting method.
[0038] The style feature parameter search module 40 is used to search for the style feature parameters corresponding to the original calligraphy and painting works.
[0039] Specifically, a VGG or ResNet network pre-trained on a large art image dataset is used, with its final classification layer removed, leaving the preceding layers as a fixed feature extractor. The original image of the target calligraphy or painting is input into this network, and a high-dimensional feature vector is output from an intermediate layer, such as the fc1 layer of VGG-19. This vector encodes the image's style information. Similarly, a nearest neighbor search is performed in a style database containing known artist style labels to find the sample with the best matching style features, thereby determining the attribution of style feature parameters such as brushstroke strength, composition, and flow.
[0040] The copying guidance module 50 is used to provide copying guidance based on multi-scale similarity comparison information, the first feature corresponding to the brush and ink texture, the second feature corresponding to the composition layout, and in combination with creative technique parameters and style feature parameters.
[0041] Specifically, the copying guidance module 50 is the system's decision center, used to integrate all information and determine, through a rule engine and probability model, whether and how guidance is needed.
[0042] In one possible implementation, the copying guidance prompt module 50 includes: a technique and style state space construction module, used to construct a technique and style state space through creation technique parameters and style feature parameters; a copying deviation probability derivation module, used to deduce the copying deviation probability in the technique and style state space to obtain the copying deviation type and deviation probability value; and a judgment module, used to determine whether to trigger the copying guidance prompt based on the copying deviation type and deviation probability value.
[0043] Specifically, the technique and style state space construction module transforms abstract artistic features of calligraphy and painting into a computable, structured model. This module uses the searched creative technique parameters and style feature parameters as a set of orthogonal feature dimensions, collectively spanning a high-dimensional technique and style state space. In this space, every calligraphy or painting work, whether a historical classic, the original target calligraphy or painting rubbing, or a user's copy, can be precisely mapped to a feature point. For intuitive visualization, manifold dimensionality reduction algorithms such as t-SNE can be used to project this high-dimensional space into two or three dimensions, thus visually presenting the relative positional relationship between the original target calligraphy or painting rubbing and the user's copy in the artistic feature space.
[0044] The copying deviation probability derivation module performs quantitative analysis within the technique and style state space. This module locates the feature vector obtained after feature extraction from the user's copied work within the aforementioned technique and style state space and calculates the multi-dimensional distance between it and the feature points of the original calligraphy or painting. Specifically, it calculates the differences between the copied work and the original's feature descriptors in various dimensions such as brushstroke shape, line curvature, ink distribution, and composition. For example, it uses Bach distance to measure the difference in direction histograms and dynamic time warping distance to measure the difference in curvature sequences. The system pre-sets a reasonable deviation range model based on a large amount of professional copying data. The module compares the calculated deviation values of each dimension with this probability model to determine the significance of the current deviation, i.e., the deviation probability value. For example, if the deviation value of the copied work in the dimension of brushstroke direction consistency falls outside the reasonable distribution range, the system will assign it a high deviation probability, indicating that the problem is very significant. Based on the significant deviation probability values in each dimension, the module summarizes the specific types of copying deviations. For example, high-probability deviations in the consistency of brushstroke direction and the matching degree of line curvature are summarized as brushstroke shape distortion; high-probability deviations in character spacing and line spacing features and deviations in compositional balance features are summarized as improper layout.
[0045] The judgment module includes a threshold judgment logic, which presets a probability threshold for each type of deviation. When the calculated probability of a certain deviation type exceeds its threshold, the system triggers a copying guidance prompt for that type.
[0046] In one possible implementation, the technique style state space construction module includes: a creative technique parameter determination module, used to determine the creative technique parameters corresponding to the target calligraphy and painting original work, including brushwork characteristics, ink rules, and character structure paradigms; and a style feature parameter determination module, used to determine the style feature parameters corresponding to the target calligraphy and painting original work, including brushstroke strength characteristics, composition and flow characteristics, and individual style identifiers.
[0047] Specifically, the creative technique parameter determination module transforms abstract, experiential creative techniques into a series of quantifiable, computer-processable data parameters. It directly measures these parameters from the artwork image using specific image processing and signal analysis algorithms. Specifically, brushstroke characteristic parameters include the ratio of central to side strokes, the encoding of starting and ending stroke forms, and the quantification of brushstroke speed. First, the strokes are skeletonized to extract the center line, then a distance transformation is performed to obtain the stroke width map. When using a central stroke, the ink is concentrated in the center of the stroke, and the center line highly overlaps with the core ink area; when using a side stroke, the ink is biased to one side, and the center line deviates from the core ink area. The module calculates the center line offset rate for each point on the skeleton lines of all strokes in the entire artwork, where the center line offset rate = offset distance / stroke half-width. A threshold is set; points with offset rates below the threshold are considered central stroke points, and those above are considered side stroke points. The final output is the proportion of central stroke points as a core parameter. Using histograms of oriented gradients combined with shape context descriptors, features are extracted from the starting and ending points of each stroke. These features are then categorized into different forms such as concealed tip, exposed tip, and returning tip using a pre-trained classifier, such as Support Vector Machine (SVM). The final output is a probability distribution vector for the starting / ending stroke form, for example: Concealed tip: 65%, Exposed tip: 25%, Returning tip: 10%. Brush speed is directly related to the dryness and wetness of the ink, and the effect of "flying white" (or "splashing white"), which is indirectly quantified by calculating the roughness and contrast of the internal texture of the stroke. Areas with more "flying white" and drier ink have higher texture roughness. Specifically, a gray-level co-occurrence matrix (GLCM) is used to extract the contrast and entropy values of the stroke area. High contrast and high entropy values correspond to faster brushstrokes and more "flying white". The final output is the average texture contrast of the stroke area as a parameter for brush speed.
[0048] The ink technique parameters include the distribution of dark, light, dry, wet, and burnt ink areas, as well as the layering of splashed / accumulated ink. The L, a, and b components are extracted from the Lab color space, the S component from the HSV color space, and the local variance of the V component is calculated. These five features are combined into a five-dimensional feature vector, and then the image pixels are clustered in this five-dimensional feature space using a K-means clustering algorithm. Each cluster corresponds to a specific ink color state. The final output is a histogram showing the area proportion of each ink color pixel in the entire artwork and their spatial distribution. The layering of splashed / accumulated ink is obtained by analyzing the depth information of the ink areas. Multiple applications of ink accumulation or splashing will cause changes in the ink layer thickness and texture. The surface roughness and height variance of the ink areas are calculated using depth data acquired by the 3D contour scanner in the feature synchronization acquisition module 20. Areas with rich layering will have higher entropy values for depth texture than flat areas.
[0049] The character structure parameters include the proportions of the character's structure and the regularity of stroke spacing. The proportions of the character's structure are obtained through bounding rectangle analysis of each character. This involves calculating the character's aspect ratio, its centroid position relative to its geometric center, and the pixel density distribution of strokes within each grid after dividing the bounding rectangle into a nine-square grid. This density distribution vector represents the character's structural parameters. Within each character, the regularity of stroke spacing is obtained by calculating the average distance and variance between parallel strokes, such as horizontal and vertical strokes. A smaller variance indicates a more uniform stroke arrangement and stronger regularity. The average stroke spacing and spacing variance are output as the character structure's normative parameters.
[0050] The style feature parameter determination module extracts macroscopic features that transcend individual techniques and collectively define the uniqueness of a calligrapher, painter, or school of thought. It relies on high-order information captured through statistical models and deep learning. Specifically, the physical force of the brushstrokes is strongly correlated with the three-dimensional depth information and stroke width acquired by the feature synchronization acquisition module 20. A heavier stroke results in a deeper indentation in the paper and, simultaneously, more ink accumulation, leading to a wider stroke. Statistical analysis of all stroke points in the artwork yields a two-dimensional Gaussian distribution. The mean vector of this distribution (average depth, average stroke width) represents the average force level, and the covariance matrix represents the synergistic relationship between force and stroke width variation. This set of distribution parameters constitutes the brushstroke force feature.
[0051] The compositional flow and rhythm characteristics include the line axis fluctuation vector and the inter-character correspondence and correlation. For each line of text, the centroid of each column of pixels is calculated, and then these centroids are connected to form a curve, which is the line axis. A Fourier transform is performed on this line axis to extract the amplitude and phase of its low-frequency components. A large low-frequency amplitude indicates a large flow and rhythm fluctuation, while the phase information reflects the fluctuation pattern. This low-frequency Fourier descriptor is the flow and rhythm feature vector. The relative position vectors between the bounding rectangles of adjacent characters are calculated, and the autocorrelation of these vectors between multiple consecutive characters is calculated. The autocorrelation coefficient is the correlation parameter.
[0052] A pre-trained deep convolutional neural network is used to input the entire artwork or a large block of it into the network. Activation values are extracted from the middle layers of the network to form a high-dimensional feature vector. This vector, in the form of a dense vector, encodes all the subtle combinations of the artist's habits in brushstrokes, composition, and ink application, serving as a unique digital fingerprint of the artist's style. This high-dimensional style vector is the final identifier of the artist's individual style.
[0053] In one possible implementation, the copying guidance prompt module 50 includes: a preliminary deviation range acquisition module, used to obtain a preliminary deviation range through the multi-scale similarity comparison information if the copying guidance prompt is triggered; and a boundary correction module, used to correct the boundary of the preliminary deviation range according to the first feature corresponding to the ink texture, and to correct the boundary of the preliminary deviation range according to the second feature corresponding to the composition layout.
[0054] Specifically, when the judgment module triggers the prompt, the preliminary deviation range acquisition module extracts the image regions with the lowest similarity from the output of the multi-scale similarity comparison module 10. The specific operation is as follows: a similarity heatmap is generated on the highest resolution pyramid layer, and then a binary mask is obtained by image binarization, i.e., setting a low similarity threshold, where the white area is the preliminary deviation range.
[0055] The boundary correction module optimizes the coarse deviation boundary using finer local features. Specifically, it uses the initial deviation range as the possible foreground region for the GrabCut image segmentation algorithm, and the remaining portion as the background. Based on the color and texture information within the region, the GrabCut algorithm iteratively estimates the Gaussian mixture model of the foreground and background, and reassigns pixel labels, ultimately outputting a more accurate deviation region that more closely matches the brushstroke boundary. The color and texture information within the region comes from the brushstroke texture and composition layout features acquired by the feature synchronization acquisition module 20.
[0056] In one possible implementation, the boundary correction module includes: a semantic segmentation module, used to perform semantic segmentation on the magnified image of details within the initially defined deviation range, dividing the image into multiple semantic blocks with similar brushstroke features; a semantic block feature vector construction module, used to extract brushstroke trajectory features, ink density features, and texture roughness features of the multiple semantic blocks, and construct semantic block feature vectors; and an initial deviation range correction module, used to correct the boundary of the initially defined deviation range using the semantic block feature vectors and brushstroke anomaly features of the deviation area.
[0057] Specifically, the semantic segmentation module can use a pre-trained U-Net convolutional neural network. The network's training data consists of a large number of manually labeled calligraphy and painting images with different ink category labels, such as dark ink cores, light ink edges, dry brush strokes, and paper backgrounds. The network input is an image patch of the deviation region, and the output is an image of the same size, where each pixel is labeled as one of the preset semantic categories, thereby dividing the image into multiple semantic blocks.
[0058] The semantic block feature vector construction module calculates the internal features of each segmented semantic block. Specifically, for semantic blocks belonging to the ink category, its skeleton lines are extracted, and the shape of the skeleton is represented by a Fourier descriptor to ensure it is unaffected by translation, rotation, and scaling, thus obtaining the brushstroke trajectory features. The average brightness value of all pixels within the semantic block in the grayscale image is calculated to obtain the ink density features. The local binary pattern histogram of the semantic block is calculated to obtain the texture coarseness features. These features are then concatenated into a comprehensive feature vector.
[0059] The initial deviation range correction module first filters out abnormal semantic blocks using cosine similarity, then locates stroke breakpoints and morphological abrupt change points within these abnormal blocks, and finally confirms the most critical deviation pixels that need to be highlighted based on the abnormal stroke direction.
[0060] In one possible implementation, the preliminary deviation range correction module includes: a similarity threshold definition module, used to define a similarity threshold based on the cosine similarity between the semantic block feature vector and the standard pen stroke feature vector; a deviation region marking module, used to traverse the multiple semantic blocks, and if the cosine similarity is less than the similarity threshold, to mark the deviation region and extract the edge features of the deviation region; a direction vector field generation module, used to generate a direction vector field of edge pixels based on the brush stroke breakpoints and morphological abrupt change points associated with the edge features of the deviation region; and a brush stroke abnormality feature acquisition module, used to determine the angular distribution with respect to the standard brush stroke direction based on the direction vector field of the edge pixels, thereby obtaining the brush stroke abnormality features.
[0061] Specifically, the similarity threshold definition module is used to define the similarity threshold. This threshold is not a fixed value but is determined through statistical process control. During the learning phase, the system calculates the cosine similarity between the semantic block feature vectors of a large number of qualified imitation samples and the standard vector, and then calculates the mean and standard deviation of these similarity values. In practical applications, the threshold is set to the mean minus 3 times the standard deviation; values below this are considered low-probability anomalies.
[0062] The deviation region labeling module iterates through each semantic block, calculates the cosine similarity between its feature vector and the feature vector of the corresponding region in the standard template, and compares it with the aforementioned dynamic threshold. If the similarity is lower than the threshold, all pixel coordinates of that semantic block are recorded in the fine deviation pixel list.
[0063] The direction vector field generation module applies the Canny edge detection operator to areas with fine deviations to obtain edges that are one pixel wide. The Sobel operator is used to calculate the gradient of the edge image in the X and Y directions. The direction vector of each edge pixel is the gradient of the edge image in the X and Y directions. This vector field represents the direction of the fastest brightness change at each edge point in the image; in stroke regions, this direction is perpendicular to the stroke direction.
[0064] The brushstroke anomaly feature acquisition module calculates the edge direction vector field of the corresponding region in the standard template image using the same method. For each edge pixel in the deviation region, the angle between its direction vector and the direction vector of the corresponding pixel in the standard template is calculated. All angles are statistically analyzed; if a large number of angles exceed a preset value, it indicates a brushstroke direction error in that region, such as incorrect rubbing direction or deviation in the brushstroke path. The statistical distribution of these angles constitutes the final brushstroke anomaly feature. The system highlights these abnormal pixels as the final and most accurate guidance for the user.
[0065] This application's embodiments use AI visual acquisition devices to scan the original target calligraphy and painting rubbings, compare them with standard image data at multiple scales, use multimodal imaging devices to obtain the original's brushstroke texture and composition features, search for the original's creative technique parameters and style feature parameters, and provide copying guidance based on the above information. These technical means solve the technical problem of insufficient accuracy in existing calligraphy and painting rubbing copying guidance, and achieve the technical effect of improving the accuracy of copying guidance.
[0066] In the above text, refer to Figure 1 This paper describes in detail an AI vision-based system for reconstructing and guiding the copying of calligraphy and painting rubbings according to an embodiment of the present invention. Next, we will refer to... Figure 2 This invention describes a method for reconstructing and guiding the copying of calligraphy and painting rubbings based on AI vision, according to an embodiment of the present invention.
[0067] The AI-based vision-based method for reconstructing and guiding the copying of calligraphy and painting rubbings according to embodiments of the present invention addresses the technical problem of insufficient accuracy in existing calligraphy and painting rubbings copying guidance, thereby improving the accuracy of copying guidance.
[0068] A method for reconstructing and guiding the copying of calligraphy and painting rubbings based on AI vision includes: scanning the original target calligraphy and painting rubbings using an AI vision acquisition device; performing multi-scale similarity comparison between the high-definition image data obtained by the AI vision acquisition device and the standard image data stored in historical classic models; simultaneously acquiring the first feature corresponding to the brush and ink texture and the second feature corresponding to the composition layout of the original target calligraphy and painting rubbings using a multi-modal imaging device; searching for the creative technique parameters corresponding to the original target calligraphy and painting rubbings; searching for the style feature parameters corresponding to the original target calligraphy and painting rubbings; and providing copying guidance based on the multi-scale similarity comparison information, the first feature corresponding to the brush and ink texture, the second feature corresponding to the composition layout, and the creative technique parameters and style feature parameters.
[0069] The process of performing multi-scale similarity comparison between the high-definition image data obtained by the AI visual acquisition device and the standard image data stored in historical classic templates may further include: performing image pyramid processing on the high-definition image data and the standard image data to generate multiple resolution image layers; determining image feature descriptors for each resolution image layer in the multiple resolution image layers, wherein the image feature descriptors include at least one of brushstroke direction, line thickness variation, and ink density gradient; and performing multi-scale similarity comparison at different scales using the image feature descriptors.
[0070] The multi-scale similarity comparison using the image feature descriptor at different scales may further include: multi-scale similarity comparison indicators including any one or more of the following: brushstroke shape consistency, line curvature matching degree, and ink color distribution similarity.
[0071] The process of simultaneously acquiring the first feature corresponding to the brush and ink texture and the second feature corresponding to the composition layout of the original target calligraphy, painting, and rubbing can further include: capturing a magnified image of the details of the original target calligraphy, painting, and rubbing, and extracting the features of brush pressure change, ink color diffusion, and line turning as the first feature corresponding to the brush and ink texture; capturing the overall layout image of the original target calligraphy, painting, and rubbing, and extracting the features of character spacing and line spacing, compositional balance, and density distribution as the second feature corresponding to the composition layout.
[0072] The method of combining creative technique parameters and style characteristic parameters to provide guidance and prompts for imitation can further include: constructing a technique and style state space using creative technique parameters and style characteristic parameters; deriving the imitation deviation probability in the technique and style state space to obtain the imitation deviation type and deviation probability value; and determining whether to trigger the imitation guidance prompt based on the imitation deviation type and deviation probability value.
[0073] The process of constructing a technique and style state space by means of creative technique parameters and style characteristic parameters may further include: determining the creative technique parameters corresponding to the target original calligraphy and painting inscription, including brushwork characteristics, ink rules, and character structure paradigms; and determining the style characteristic parameters corresponding to the target original calligraphy and painting inscription, including brushstroke strength characteristics, composition and flow characteristics, and individual style identifiers.
[0074] The process of determining whether the copying guidance prompt is triggered may further include: if the copying guidance prompt is triggered, obtaining a preliminary definition of the deviation range through the multi-scale similarity comparison information; correcting the boundary of the preliminary definition of the deviation range according to the first feature corresponding to the ink texture; and correcting the boundary of the preliminary definition of the deviation range according to the second feature corresponding to the composition layout.
[0075] The step of correcting the boundary of the initially defined deviation range based on the first feature corresponding to the ink texture may further include: performing semantic segmentation on the magnified detail image within the initially defined deviation range to divide the image into multiple semantic blocks with similar ink features; extracting the brushstroke trajectory features, ink density features, and texture roughness features of the multiple semantic blocks to construct semantic block feature vectors; and correcting the boundary of the initially defined deviation range using the semantic block feature vectors and the abnormal brushstroke features of the deviation area.
[0076] The method may further include: defining a similarity threshold based on the cosine similarity between the semantic block feature vector and the standard brushstroke feature vector; traversing the multiple semantic blocks, and if the cosine similarity is less than the similarity threshold, marking the deviation region and extracting the edge features of the deviation region; generating a direction vector field of edge pixels based on the brushstroke breakpoints and morphological abrupt change points mapped by the edge features of the deviation region; and determining the angular distribution with respect to the standard brushstroke direction based on the direction vector field of the edge pixels to obtain the brushstroke abnormal features.
[0077] The AI-based vision-based calligraphy and painting stele feature reconstruction and copying guidance system provided in the embodiments of the present invention can execute the AI-based vision-based calligraphy and painting stele feature reconstruction and copying guidance method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.
[0078] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.
[0079] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.
Claims
1. An AI vision-based calligraphy and painting tablet feature reconstruction and copying guidance system, characterized in that, The method comprises the following steps: A multi-scale similarity comparison module is used to scan the target calligraphy and painting original work using an AI vision acquisition device, and to perform multi-scale similarity comparison on high-definition image data scanned by the AI vision acquisition device and standard image data stored by a historical classic model; A feature synchronous acquisition module is used to synchronously acquire first features corresponding to ink texture and second features corresponding to layout of the target calligraphy and painting original work using a multi-modal imaging device; An artistic technique parameter search module is used to search for artistic technique parameters corresponding to the target calligraphy and painting original work; A style feature parameter search module is used to search for style feature parameters corresponding to the target calligraphy and painting original work; A copying guidance prompt module is used to perform copying guidance and prompt based on multi-scale similarity comparison information, first features corresponding to ink texture and second features corresponding to layout, in combination with artistic technique parameters and style feature parameters. 2.The AI vision-based calligraphy and painting tablet feature reconstruction and copying guidance system of claim 1, wherein, The multi-scale similarity comparison module comprises: An image pyramid processing module is used to perform image pyramid processing on the high-definition image data and the standard image data to generate a plurality of resolution image layers; An image feature descriptor determination module is used to determine image feature descriptors under each resolution image layer in the plurality of resolution image layers, wherein the image feature descriptors comprise at least one of ink stroke direction, line thickness variation and ink color gradient; A similarity comparison module is used to perform multi-scale similarity comparison on the image feature descriptors at different scales. 3.The AI vision-based calligraphy and painting tablet feature reconstruction and copying guidance system of claim 2, wherein, The similarity comparison module comprises: A comparison index setting module is used to set comparison indexes, and the multi-scale similarity comparison indexes comprise any one or more of stroke form consistency, line curvature matching degree and ink color distribution similarity. 4.The AI vision-based calligraphy and painting tablet feature reconstruction and copying guidance system of claim 1, wherein, The feature synchronous acquisition module comprises: A first feature acquisition module is used to capture a detail enlarged image of the target calligraphy and painting original work, and to extract stroke pressure variation features, ink color gradation features and line turning features as first features corresponding to ink texture; A second feature acquisition module is used to capture an overall layout image of the target calligraphy and painting original work, and to extract word spacing and line spacing features, composition balance features and density distribution features as second features corresponding to layout. 5.The AI vision-based calligraphy and painting tablet feature reconstruction and copying guidance system according to claim 1, wherein, The copying guidance prompt module comprises: A technique and style state space construction module is used to construct a technique and style state space through artistic technique parameters and style feature parameters; A copying deviation probability derivation module is used to derive copying deviation types and deviation probability values in the technique and style state space; A judgment module is used to judge whether to trigger the copying guidance prompt according to the copying deviation types and deviation probability values. 6.The AI vision-based calligraphy and painting tablet feature reconstruction and copying guidance system according to claim 5, wherein, The technique and style state space construction module comprises: An artistic technique parameter determination module is used to determine artistic technique parameters of the target calligraphy and painting original work, including stroke characteristics, ink rules and character connection patterns; A style feature parameter determination module is used to determine style feature parameters of the target calligraphy and painting original work, including stroke force characteristics, layout and qi characteristics and individual style identifiers. 7.The AI vision-based calligraphy and painting tablet feature reconstruction and copying guidance system according to claim 6, characterized in that, The copying guidance prompt module comprises: The preliminary deviation range determination module is configured to, if the copying guidance prompt is triggered, determine a preliminary deviation range based on the multi-scale similarity comparison information. The boundary correction module is configured to correct the boundary of the preliminary deviation range based on the first feature corresponding to the ink texture and correct the boundary of the preliminary deviation range based on the second feature corresponding to the layout. 8.The AI vision-based calligraphy and painting tablet feature reconstruction and copying guidance system of claim 7, wherein, The boundary correction module includes: The semantic segmentation module is configured to perform semantic segmentation on the detail-enlarged image in the preliminary deviation range, and divide the image into a plurality of semantic blocks with similar ink features. The semantic block feature vector construction module is configured to extract the brush stroke trajectory feature, ink color density feature, and texture roughness feature of the plurality of semantic blocks, and construct a semantic block feature vector. The preliminary deviation range correction module is configured to correct the boundary of the preliminary deviation range based on the semantic block feature vector and the brush stroke abnormality feature of the deviation region. 9.The AI vision-based calligraphy and painting tablet feature reconstruction and copying guidance system of claim 8, wherein, The preliminary deviation range correction module includes: The similarity threshold definition module is configured to define a similarity threshold based on the cosine similarity between the semantic block feature vector and the standard ink feature vector. The deviation region marking module is configured to traverse the plurality of semantic blocks, and if the cosine similarity is less than the similarity threshold, mark the deviation region and extract the edge feature of the deviation region. The direction vector field generation module is configured to generate a direction vector field of edge pixels based on the edge feature of the deviation region, the brush stroke breakpoint, and the morphological mutation point associated with the mapping. The brush stroke abnormality feature acquisition module is configured to determine the angle distribution with the standard brush stroke direction based on the direction vector field of the edge pixels, and obtain the brush stroke abnormality feature. 10.A method for AI vision-based feature reconstruction and copying guidance of calligraphy and painting tablets, characterized in that, The method is implemented by the AI vision-based calligraphy and painting characteristic reconstruction and copying guidance system of any one of claims 1-9, and the method includes: scanning a target calligraphy and painting original work using an AI vision acquisition device, performing multi-scale similarity comparison on high-definition image data scanned by the AI vision acquisition device and standard image data stored by a historical classic model; synchronously acquiring the first feature corresponding to the ink texture and the second feature corresponding to the layout of the target calligraphy and painting original work using a multi-modal imaging device; searching for the creation technique parameters corresponding to the target calligraphy and painting original work; searching for the style feature parameters corresponding to the target calligraphy and painting original work; performing copying guidance based on the multi-scale similarity comparison information, the first feature corresponding to the ink texture, the second feature corresponding to the layout, the creation technique parameters, and the style feature parameters.