A cultural symbol propagation method based on Sichuan opera mask data correlation analysis

Abstract

The application discloses a cultural symbol propagation method based on Sichuan opera face pattern data correlation analysis, and relates to the field of Sichuan opera face pattern data analysis and propagation.The application constructs a Sichuan opera face pattern multi-modal standardized data set containing spatial information, generates a structured feature correlation matrix by combining clustering and edge detection algorithms, calculates the symbol binding strength between visual features and cultural semantics by using an improved correlation rule mining model introducing a visual advantage degree weight, constructs a semantic network based on a graph theory iteration algorithm to quantize the cultural propagation potential energy of core symbols, simultaneously performs vectorization component reconstruction and metadata encapsulation on the core schema, and finally calculates the cultural resonance degree by combining audience preferences and propagation potential energy, so that the whole-link intelligent processing of Sichuan opera face pattern cultural symbols from deep semantic mining, high-quality digital production to accurate propagation is realized, and the accuracy and effectiveness of intangible cultural heritage propagation are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of Sichuan Opera facial makeup data analysis and dissemination, specifically a method for disseminating cultural symbols based on the correlation analysis of Sichuan Opera facial makeup data. Background Technology

[0002] Sichuan Opera facial makeup serves as a crucial visual symbol in Sichuan Opera performance art, used to portray character traits and distinguish between good and evil, loyalty and treachery. Through exaggerated color combinations (such as "red for loyalty, white for treachery") and varied line and stripe patterns (such as "three tiles" and "whole face" patterns), it constructs a rigorous system of visual expression. In the digital age, effectively extracting and disseminating the cultural meaning behind these facial makeup designs, and digitally deconstructing and recombining their "form" and "meaning," has become a key issue in the digital protection and inheritance of intangible cultural heritage.

[0003] However, existing research on the digitization of Sichuan Opera facial makeup largely focuses on simple image acquisition and archiving or shallow classification and recognition based on deep learning. For example, current technologies often utilize convolutional neural networks to classify facial makeup for different roles or simply establish static image databases, failing to delve into the deep-seated connections between the visual features (colors, patterns) and cultural semantics (personality, allusions) of facial makeup. Traditional association rule mining methods, when applied to this type of cultural image data, often calculate support and confidence based solely on statistical frequency, ignoring the decisive weight of "visual dominance" such as color coverage area and pattern complexity in the facial makeup on the semantic expression of the symbols. This results in rules lacking domain explanatory power and failing to distinguish the weight differences in meaning between large areas of primary color and small decorative colors. Furthermore, existing dissemination and distribution methods are mostly based on simple user visual geometric feature preferences for recommendations, lacking a quantitative assessment of the global dissemination value of core cultural symbols. This makes it difficult to achieve an effective balance between cultural value guidance and personalized user needs, easily leading to the loss of core semantics or homogenization of recommended content during the dissemination of cultural symbols. Summary of the Invention

[0004] This invention proposes a cultural symbol dissemination method based on Sichuan Opera facial makeup data association analysis, aiming to address the technical problems of existing technologies failing to effectively integrate the physical attributes of visual features for deep semantic association mining, and lacking a precise distribution mechanism based on cultural dissemination potential.

[0005] One method for the dissemination of cultural symbols based on the correlation analysis of Sichuan Opera facial makeup data includes the following steps:

[0006] S1. Collect images of Sichuan opera facial makeup and corresponding text description data, segment the key facial regions of the images and extract multimodal features, and construct a standardized facial makeup dataset containing spatial information;

[0007] Specifically, step S1 involves establishing a dedicated original database for Sichuan Opera facial makeup and using a high-precision scanner or high-definition camera to collect original images containing full-face and close-up details. Simultaneously, expert-annotated character traits, role classifications, and historical anecdotes are entered. Image annotation tools are used to perform anatomical region-of-interest segmentation on each facial makeup image, dividing it into the main forehead and eyebrow / eye structure area, and the chin and mouth auxiliary area. A unique spatial coordinate mask is generated for each area. Natural language processing technology is used to clean the text data to establish a dedicated Sichuan Opera dictionary and normalize synonyms. The processed images are input into a convolutional neural network to extract high-dimensional visual feature tensors covering texture and shape. The standardized text is input into a language model to extract semantic feature vectors. Based on a unique sample ID, the visual tensor and semantic vector are concatenated to construct a standardized facial makeup dataset that includes both pixel spatial distribution and abstract cultural meaning.

[0008] S2. Read the output face normalization dataset, perform quantization calculation on the image data in the face normalization dataset, generate the distribution ratio of primary and secondary tones and geometric feature vectors of texture lines, perform index alignment and numerical normalization processing on the text semantic index vectors in the face normalization dataset, and construct a structured feature association matrix.

[0009] Specifically, step S2 reads the standardized facial dataset and performs color quantization processing on the visual feature tensor. A clustering algorithm is used to divide the color space of the facial image into several primary color clusters and secondary color clusters, and the pixel coverage ratio of each color cluster in the effective area of ​​the face is calculated to generate the distribution ratio of primary and secondary colors, thereby quantifying the hue features of the facial image. Geometric morphology quantization is performed on the image data, and the skeleton of the facial pattern is extracted using the edge detection operator to calculate the mean curvature, distribution density, and gradient direction histogram of the pattern skeleton, thereby generating geometric feature vectors of the pattern lines to quantify the brushwork features of the facial image. The semantic feature vectors in the dataset are parsed, and the discretized character personality label index and pattern pattern classification index are extracted. The generated color ratio, geometric feature vector, and semantic index vector are aligned row-wise according to the sample ID, and the continuous numerical features are normalized to eliminate the difference in units. The resulting matrix is ​​then constructed by splicing the data to form a structured feature association matrix with rows representing samples and columns representing multi-dimensional attributes.

[0010] S3. Convert the structured feature association matrix generated in step S2 into a transaction dataset, input it into the association rule mining model based on the Apriori algorithm, and calculate the symbol binding strength between visual features and cultural semantic labels by introducing visual dominance as a weight factor, and construct a strong association rule library for facial recognition.

[0011] Specifically, step S3 discretizes the structured feature association matrix to map continuous color proportions and geometric feature values ​​to interval categories and converts them together with discrete semantic labels into a Boolean transaction dataset. An association rule mining model based on the Apriori algorithm is constructed, and minimum support and minimum confidence thresholds are set to traverse the transaction dataset to filter out frequently occurring co-occurrence combinations of visual feature items and cultural semantic label items. The prior probability of semantic labels and the posterior probability under specific visual feature occurrence conditions are calculated respectively. Visual dominance is introduced as a key weighting factor, and the original values ​​in the structured feature association matrix are backtracked to calculate the average coverage area or average line intensity of visual features in all supporting samples for each candidate set, thereby generating visual dominance. The ratio of posterior probability to prior probability, i.e., the basic lift, is weighted and corrected using visual dominance to calculate the symbol binding strength between visual features and cultural semantic labels after visual saliency enhancement. Combinations with binding strength below the preset threshold are eliminated, and strong association rules are retained to construct a strong association rule library for facial recognition.

[0012] S4. Construct a semantic network based on the calculated symbol binding strength, use a graph iterative algorithm to calculate the cultural dissemination potential of each node in the network, identify and output core cultural symbols; reconstruct the visual form into vectorized components based on the core cultural symbols, and encapsulate the calculated symbol binding strength and cultural dissemination potential into metadata to generate a digital dissemination material library.

[0013] Specifically, step S4 parses the strong association rule base of facial makeup and instantiates the visual feature items in the rules into facial makeup visual nodes and the cultural semantic label items into facial makeup semantic nodes. It uses the calculated symbol binding strength as the edge weight connecting the visual nodes and semantic nodes to construct a heterogeneous facial makeup semantic network. It defines the cultural dissemination potential energy index and runs a graph theory-based iterative algorithm on the semantic network to simulate the process of energy flowing from high semantic value nodes to high visual association nodes. It repeatedly updates the weight values ​​of each node until the network reaches a convergence state, thereby identifying the core facial makeup patterns that have both visual representativeness and semantic richness in the entire Sichuan Opera cultural system. Based on the index of the core facial makeup patterns, it backtracks the standardized dataset and extracts the corresponding high-resolution local image patches. It uses computer graphics algorithms to denoise and binarize the image patches and uses Bézier curve fitting technology to vectorize and reconstruct the pattern outlines and fill the regions based on the main color data to generate lossless scalable vectorized facial makeup components. It writes the calculated symbol binding strength and cultural dissemination potential energy into the metadata attributes of the vector file to generate a plug-and-play digital dissemination material library with self-explanatory data.

[0014] The beneficial effects of the invention are:

[0015] This invention constructs a standardized multimodal dataset of Sichuan Opera facial makeup containing spatial information, generates a structured feature association matrix by combining clustering and edge detection algorithms, calculates the symbol binding strength between visual features and cultural semantics using an improved association rule mining model that introduces visual dominance weights, and constructs a semantic network based on graph theory iterative algorithms to quantify the cultural dissemination potential of core symbols. At the same time, it reconstructs core graphs into vectorized components and encapsulates metadata. Finally, it calculates cultural resonance by combining audience preferences and dissemination potential, thereby realizing intelligent processing of Sichuan Opera facial makeup cultural symbols across the entire chain from deep semantic mining and high-quality digital production to precise dissemination, significantly improving the accuracy and effectiveness of intangible cultural heritage dissemination. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating a method for disseminating cultural symbols based on the correlation analysis of Sichuan Opera facial makeup data, according to an embodiment of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, and not all of them. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0019] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention. It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0020] Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or machine that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or machine. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or machine that includes said element.

[0021] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0022] Example 1

[0023] Among them, such as Figure 1 A method for the dissemination of cultural symbols based on the correlation analysis of Sichuan Opera facial makeup data includes the following steps:

[0024] S1. Collect images of Sichuan opera facial makeup and corresponding text description data, segment the key facial regions of the images and extract multimodal features, and construct a standardized facial makeup dataset containing spatial information;

[0025] S2. Read the output face normalization dataset, perform quantization calculation on the image data in the face normalization dataset, generate the distribution ratio of primary and secondary tones and geometric feature vectors of texture lines, perform index alignment and numerical normalization processing on the text semantic index vectors in the face normalization dataset, and construct a structured feature association matrix.

[0026] S3. Convert the structured feature association matrix generated in step S2 into a transaction dataset, input it into the association rule mining model based on the Apriori algorithm, and calculate the symbol binding strength between visual features and cultural semantic labels by introducing visual dominance as a weight factor, and construct a strong association rule library for facial recognition.

[0027] S4. Construct a semantic network based on the calculated symbol binding strength, use a graph iterative algorithm to calculate the cultural dissemination potential of each node in the network, identify and output core cultural symbols; reconstruct the visual form into vectorized components based on the core cultural symbols, and encapsulate the calculated symbol binding strength and cultural dissemination potential as metadata to generate a digital dissemination material library.

[0028] Furthermore, step S1 specifically includes the following sub-steps:

[0029] S101. Establish a multimodal raw database of Sichuan Opera facial makeup, collect a set of raw facial makeup images including front view, side view and close-up details, and collect a set of raw text descriptions corresponding to each image.

[0030] S102. Read the original set of facial images, use image annotation tools to divide the forehead, eye sockets and jaw area into regions of interest, and generate region-annotated image data; and perform synonym merging and noise reduction on the original text description set to generate a standardized text sequence.

[0031] S103. Input the generated region-labeled image data into a pre-trained convolutional neural network to extract the high-dimensional feature map of the image, output the visual feature tensor, and input the generated standardized text sequence into a natural language processing model to perform word embedding encoding and output the semantic feature vector.

[0032] S104. The visual feature tensor and semantic feature vector are concatenated and formatted according to the unique sample ID index to construct a standardized Facebook dataset containing image features, text features and corresponding labels.

[0033] For example, the implementation process of the above steps is shown below: A relational database system such as MySQL or MongoDB is deployed, and a 4K high-resolution industrial camera is used to take full-face and close-up images of the face at a Sichuan Opera museum or troupe. Simultaneously, OCR optical character recognition technology is used to input textual descriptions of character traits and historical allusions from printed scripts or classic texts. The original facial makeup image set is read and loaded into LabelImg professional image annotation software. The forehead, eye socket, and jaw area are precisely selected manually or semi-automatically to generate an XML or JSON format region annotation file containing precise vertex coordinates. Python's Jieba word segmentation library, combined with a pre-built Sichuan Opera professional dictionary, is used to segment the original text. A predefined stop word list is used to filter meaningless function words, and homonymous patterns and colors such as "peach red" and "cinnabar" are separated. Color terminology is uniformly mapped to standard terminology; the generated region-annotated image data is input into a ResNet50 convolutional neural network model pre-trained on the ImageNet dataset. Forward propagation is performed by removing fully connected layers and freezing shallow parameters, thereby outputting a high-dimensional visual feature tensor that preserves spatial structure. The standardized text sequence is input into a BERT pre-trained language model, and contextual semantic analysis is performed through a multi-layer Transformer encoder to output a fixed-length semantic feature vector. Using a unique sample ID as the primary key index, the corresponding visual feature tensor and semantic feature vector are retrieved. Both are encapsulated and stored using efficient binary data formats such as HDF5 or TFRecord, and the values ​​are Z-score standardized before storage, thus constructing a standardized face dataset with a unified structure that supports efficient subsequent reading.

[0034] Furthermore, regarding the generation and calculation logic of "region-labeled image data" in the above implementation scheme, in the prior art, the background of facial images is complex (such as costumes or stage backgrounds), and directly inputting them into the neural network would introduce a lot of noise, leading to machine association errors. However, this embodiment uses an XML file generated by LabelImg to record...<xmin, ymin, xmax, ymax> Coordinates. Technically, it's not simply about cropping the image; instead, a Region of Interest (ROI) mask is generated. This is a matrix of the same size as the original image, with values ​​of 1 within the ROI and 0 outside. A matrix multiplication is performed between the original image and the mask, causing pixel values ​​in non-critical areas to be set to zero. Therefore, when the convolutional neural network extracts features, its activation map only responds to the textures of key areas such as the forehead and eye sockets.

[0035] Furthermore, in the oral and personal transmission of Sichuan Opera, the same pattern has different names. For example, the "tiger head pattern" is called "Wang character forehead" in some scripts. If not unified in data mining, it will lead to scattered association rules and unable to mine strong rules. As a preferred implementation, before cleaning, a mapping dictionary based on the Key-Value structure needs to be preset. Exemplarily, define the dictionary Dict = {"Chi": "red", "Zhu": "red", "Dan": "red", "碎花脸": "碎脸"}. Traverse the text recognized by OCR. Once the vocabulary in the Key is matched, it is immediately replaced with the standard term in the Value. Remove the function words without actual semantics such as "的", "了", "以及" through the stop-word list, and only retain nouns (colors, patterns) and adjectives (characters) to reduce interference when calculating semantic vectors.

[0036] Furthermore, for the data alignment method in the standardization process, this embodiment is implemented by establishing a Group indexed by Sample_ID in the file, that is:

[0037] Group_001 / Visual_Data: Store a floating-point matrix of 7×7×2048;

[0038] Group_001 / Text_Data: Store a BERT semantic vector of 1×768;

[0039] Group_001 / Meta: Store the original label.

[0040] Furthermore, the step S2 specifically includes the following sub-steps:

[0041] S201. Read the image data in the facial makeup standardization dataset, perform color segmentation on pixel points through the K-Means clustering algorithm, and count the proportion of the number of pixels belonging to the same clustering cluster in the total area of the facial makeup to generate the main and auxiliary color distribution ratios;

[0042] S202. Read the image data in the facial makeup standardization dataset, extract the pattern skeleton through the Canny edge detection operator, and calculate the average curvature, distribution density and gradient direction of the line based on the skeleton pixel points to generate the geometric feature vector of the pattern line;

[0043] S203. Read the semantic features in the facial makeup standardization dataset, parse and extract the corresponding character personality label index and pattern spectrum classification index to generate the text semantic index vector;

[0044] S204. The generated primary and secondary color distribution ratios, geometric feature vectors of decorative lines, and text semantic index vectors are indexed and aligned according to the unique sample IDs, and the numerical features are normalized. Finally, they are spliced ​​together to construct a structured feature association matrix.

[0045] Specifically, the image data in steps S201-S203, which includes the visual feature tensor and semantic feature vector described in steps S101-S104, is essentially a collection of unstructured raw pixels and text descriptions transformed into computer-computable mathematical objects. The principle is as follows: the three-dimensional visual feature tensor (with dimensions exemplarily H×W×C) is reshaped and flattened to construct a feature point matrix. K color centroids are initialized using the K-Means++ clustering algorithm. Each feature point in the tensor is assigned to the nearest color cluster through Euclidean distance iteration. After convergence, the number of feature points in each color cluster is counted and divided by the total number of feature points in the tensor, thus obtaining the main and secondary color distribution ratio vector of the "main color" and "secondary color" proportions in the facial makeup. Further, a convolution operation is performed on the visual feature tensor using Gaussian difference edge detection operators such as Canny. Preferably, the Sobel method can also be used to perform the above operation, calculating the gradient magnitude and direction at each position, and refining the wide edges using non-maximum suppression. A single-pixel-width decorative skeleton is used. Based on the skeleton trajectory, the mean curvature is obtained by calculating the rate of change of the tangent angle between three adjacent skeleton points. The distribution density is obtained by statistically analyzing the proportion of non-zero pixels in the binarized skeleton image. The direction vector is obtained by statistically analyzing the gradient direction histogram of the skeleton points, thereby constructing a geometric feature vector of decorative lines describing the complexity of the facial makeup pattern. Further, a high-dimensional semantic feature vector is received. Through a pre-trained classifier or cosine similarity retrieval technology, continuous floating-point word embedding vectors are mapped to a discrete Sichuan opera body knowledge base to lock the character personality category ID and spectrum classification ID corresponding to the sample, generating a discrete semantic index vector. Further, using the unique sample ID as the primary key, the above color ratio vector, geometric feature vector and semantic index vector are horizontally concatenated. For the continuous numerical features, the maximum and minimum values ​​of each column in the whole set are calculated. Substituting them into the normalization formula, all values ​​are compressed into the closed interval [0,1] to eliminate dimensional differences, constructing a structured feature association matrix where rows represent samples and columns represent multi-dimensional attributes.

[0046] Furthermore, the calculation method for the distribution ratio of primary and secondary hues is as follows:

[0047] ;

[0048] Specifically, the distribution ratio of primary and secondary hues, as the core feature for quantifying the color dimension of facial makeup, is filled into the structured feature association matrix to complete the transformation from unstructured image to numerical matrix. In the data preprocessing stage of step S301, this value serves as the input basis for the discretization binning algorithm, mapping continuous floating-point proportion values ​​to discrete category labels such as "primary hue - red" or "secondary hue - black" to construct a Boolean transaction dataset adapted to the Apriori algorithm. In addition, in the intensity calculation process of steps S303 and S304, the original value is backtracked by the system to calculate the average visual dominance. That is, the magnitude of this value is used to directly measure the physical coverage area and visual salience of a specific color in the facial makeup visual composition, and it is used as a weight parameter in the weighted correction formula to nonlinearly amplify the basic lift, thereby ensuring that a higher symbol binding strength can be given to the primary hue in a large area when constructing the strong association rule base of facial makeup.

[0049] The calculation method for the density of the decorative lines is as follows:

[0050] ;

[0051] Specifically, the density of decorative lines, as a geometric quantitative indicator representing the complexity of facial makeup patterns and the style of brushwork, is simultaneously aligned to the structured feature association matrix. This is used to numerically distinguish between high-density complex patterns such as "broken faces" and low-density simple patterns such as "whole faces" or "three tiles." In the association rule mining model, this value is not only used to generate discrete feature terms describing the complexity of decorative patterns to uncover potential cultural logic such as "high-density decorative patterns indicate a volatile personality" or "low-density lines indicate a stable personality," but also serves as the basic data for calculating the visual advantage of decorative features. By quantifying the density of lines, the confidence weight of association rules is corrected, so that decorative patterns with complex line structures obtain a weighted coefficient proportional to visual complexity in the symbol binding strength calculation. Furthermore, in the precise distribution stage of step S5, the output results of the above two formulas together constitute the numerical basis for constructing the material attribute feature vector. This vector is used to calculate the cosine similarity with the preference weights for color vibrancy and pattern complexity in the target audience user profile, thereby determining the final content matching score and distribution priority.

[0052] Among them, the This represents the distribution ratio of the k-th color cluster in the image; the This represents image data from the Facebook normalized dataset; The Kronecker function is defined as having a value of 1 when the two parameters are equal, and 0 otherwise; Represents image coordinates The color cluster category to which the pixel belongs; the This represents the total number of pixels in the image data; This represents the distribution density value in the geometric feature vector of the decorative lines; Represents image data in coordinates The binarized result after edge detection is 1 if it is a texture edge point, and 0 otherwise.

[0053] Specifically, regarding the formula for calculating the distribution ratio of primary and secondary tones and its application in sub-step S201, the calculation of this formula is based on the K-Means clustering algorithm for face standardization dataset image data. The processing results, including parameters This is derived from the statistics of the total number of pixels in the non-zero mask region after ROI segmentation of the image, representing the total area of ​​the effective facial analysis region, while the parameters... The clustering algorithm uses pixel RGB or HSV numerical similarity as coordinates. The unique cluster label index value assigned to each pixel at a given location. (Kronecker function) The core logic-gated filter here is used to compare the cluster label of the current pixel pixel by pixel. The value is determined by whether the pixel matches the specific color cluster number k of the target color. A value of 1 is output only if the two are completely equal, indicating that the pixel belongs to the target color; otherwise, a value of 0 is output. The formula uses a summation symbol. Traversing image data The sum of the number of pixels belonging to the k-th color cluster is used to calculate the total number of pixels, and then divided by the total number of pixels. This yields the normalized distribution ratio of primary and secondary hues. This value is related to the visual coverage percentage of a primary red color or a secondary black and white color in the overall composition of a facial mask.

[0054] Furthermore, regarding the formula for calculating the density of decorative lines and its specific application in sub-step S202, the foundation of this formula lies in using the Canny or Sobel edge detection operator to analyze face-normalized dataset image data. The convolution operation and binarization processing are performed. The parameters... The binary feature matrix directly derived from the edge detection output contains coordinates. The value at a given point is determined by the gradient magnitude. If the gradient change at that point exceeds a threshold, it is marked as 1, which physically represents that the point is on the edge of the pattern or on the line skeleton. A value of 0 represents a flat color-filled area. Parameters Similarly, the normalized denominator represents the total number of pixels in the effective area of ​​the face. The formula uses a summation symbol. State values ​​of all pixels in the binarized matrix The summation essentially calculates the total number of lines and pixels that make up the facial makeup pattern, and this total number is related to the total area of ​​the image. The ratio is the density of the decorative lines. This value, as a dimensionless geometric feature, directly reflects that the more complex and detailed the design of the facial mask, the closer the value is to 1; conversely, the simpler the design and the more blank space, the closer the value is to 0.

[0055] Furthermore, step S3 specifically includes the following sub-steps:

[0056] S301. Convert the structured feature association matrix into a Boolean transaction dataset. The specific process is as follows: Discretize the distribution ratio of main and secondary colors and the geometric feature vector of the decorative lines in the continuous numerical type of the structured feature association matrix to generate visual feature items, and directly map the text semantic index vector in the structured feature association matrix to cultural semantic label items.

[0057] S302. Input the generated transaction dataset into the association rule mining model based on the Apriori algorithm, set the minimum support and minimum confidence thresholds, traverse and filter out the candidate set that meets the threshold conditions, and calculate the prior probability of cultural semantic label items in the candidate set, as well as the posterior probability of cultural semantic label items appearing under the condition of visual feature items.

[0058] S303. By backtracking the original normalized values ​​in the structured feature association matrix, calculate the distribution ratio of primary and secondary hues in the candidate set or the mean value of the geometric feature vector of the decorative lines in all supporting samples to generate visual dominance.

[0059] S304. The generated visual advantage is used as a weighting factor to adjust the ratio of the prior probability to the posterior probability of the pair, thereby obtaining the symbolic binding strength between the visual feature item and the cultural semantic label item. Combinations with symbolic binding strength greater than a preset threshold are retained to construct a strong association rule base for facial recognition.

[0060] Specifically, the execution flow of steps S301-S304 involves using the mature Apriori algorithm to identify basic associations and combining them with the unique visual features of facial recognition for secondary weighting. When executing step S301, the structured feature association matrix from step S2 is received, and the distribution ratio of primary and secondary colors and the geometric feature vector columns of decorative lines are identified. Through data discretization methods such as equal-frequency binning, the numerical range is converted into category labels with clear meanings (such as low proportion of primary color and medium density of decorative lines). For example, the numerical range is generally 0.0-0.2 or 0.2-0.4. These labels are then row-wise concatenated with the already discretized character personality index and pattern spectrum index in the matrix to generate a transaction dataset containing only Boolean values. Each row of this dataset represents a Sichuan opera mask sample, and each column represents a possible visual or semantic feature item, with a value of 1 indicating the feature exists in the sample and 0 indicating its absence. Further, during step S302, the system loads the generated Boolean transaction dataset into an association rule mining engine based on the Apriori algorithm. This engine scans the dataset, identifies frequently occurring single feature items, and gradually combines them into multiple sets until a match is found. The largest frequent itemset that satisfies (i.e., greater than or equal to) the minimum support threshold is used. For each candidate itemset that satisfies the minimum support threshold, the co-occurrence frequency of its visual feature item and cultural semantic label item is counted. This allows the calculation of the global occurrence probability of the semantic label item, i.e., the prior probability P, and the conditional probability of the semantic label item occurring given the occurrence of a visual feature item, i.e., the posterior probability. The initially generated rules are then filtered according to the set minimum confidence threshold. Furthermore, when executing step S303, the structured feature association matrix output in step S2 is backtracked, and for each sample that supports a specific candidate itemset selected in S302, the process is performed. The normalized values ​​of visual features (such as high proportion of main color and low density of patterns) contained in the candidate set are extracted from the original matrix. The arithmetic mean of these values ​​is calculated to obtain the average coverage intensity of the visual feature in the sample group supporting the rule, that is, the mean of visual dominance. It should be noted that the inherent artistic principle of the main color setting and the meaning expression of intricate patterns in Sichuan Opera facial makeup is that, in the limited facial composition space, the larger the physical space proportion (coverage) of a color area or the denser the line structure (density coverage) of a pattern, the greater the stimulus intensity produced in the viewer's visual cognition, and thus the greater the influence on the character's personality. The higher the weight of the definition in the shaping process, the higher the weight of the definition. Therefore, in the algorithm design, this scheme directly defines the normalized objective physical coverage value as the visual advantage at the subjective perception level, and directly transforms the physical quantity of how much space it occupies into a weight value with how much importance it has. This is to realize the use of calculable spatial breadth index to accurately characterize the visual influence index that cannot be directly measured. Furthermore, when executing step S304, the basic association strength index (e.g., the improvement obtained by dividing the posterior probability by the prior probability) is used as the benchmark, and the average visual advantage is introduced into the weight correction through the preset visual adjustment coefficient, according to the formula. The final symbol binding strength is calculated. Specifically, by assigning greater weight to visually more prominent features, the calculated binding strength can better reflect their importance in the Sichuan Opera facial makeup cultural symbol system. Based on the preset symbol binding strength threshold, rules with low strength values ​​are filtered out, and the strong association rules of visual features → cultural semantic labels that meet the requirements are entered into the facial makeup strong association rule library in the form of visual-semantic-binding strength.

[0061] Furthermore, the above implementation performs feature discretization preprocessing on the input structured feature association matrix, uses the MDLP algorithm based on information entropy to determine the optimal segmentation threshold, maps the continuous distribution ratio of primary and secondary hues and the floating-point values ​​such as the geometric feature vectors of decorative lines to discrete interval categories with clear semantic boundaries, and directly converts the discrete text semantic index vectors into nominal variables. After merging the two, a Boolean transaction dataset is constructed through One-Hot encoding, where each row represents a face sample and each column represents a feature item state. This dataset is then loaded into the association mining calculation engine. The engine performs frequent itemset search on the transaction dataset according to the preset minimum support threshold, uses an iterative mechanism to generate candidate itemsets layer by layer and prunes and removes combinations that do not meet the frequency requirements. For each candidate itemset that is retained and contains both visual feature antecedents and cultural semantic consequents, statistics are performed. The frequency of semantic tags in the entire sample set is used to determine the prior probability, and the frequency of their occurrence in a sub-sample set containing specific visual features is used to determine the posterior probability. To address the issue that traditional statistical probabilities cannot reflect the differences in visual impact of facial images, the index of the candidate set is used to look up the corresponding original normalized values ​​in the structured feature association matrix. The arithmetic mean of the values ​​of all samples supporting the rule on the corresponding visual feature dimension is calculated to obtain the visual dominance, which represents the physical salience of the visual element. A non-linear weighting function is constructed using this visual dominance to correct the ratio of the posterior probability to the prior probability, i.e., the base lift, and outputs a symbol binding strength index that can simultaneously reflect statistical significance and visual significance. By comparing with a preset strength threshold, weak association rules are filtered out, and the remaining high-strength combinations are serialized and stored as a strong association rule library for facial images that includes visual form, semantic connotation, and binding strength.

[0062] For example, the exemplary setup and usage process of the association rule mining model based on the Apriori algorithm involved in this solution is as follows:

[0063] During the model building phase, the model is built using the mlxtend library in the Python environment. Specifically, a transaction encoder is built, which converts the transaction data in list form prepared in step S301 into a sparse matrix that the model can recognize, i.e., a DataFrame format. The row index corresponds to the sample ID, the column index corresponds to all possible feature items (such as "main color_red", "pattern_bat", "personality_loyalty"), and the matrix elements are only True / False or 1 / 0.

[0064] During the model parameter configuration phase, `min_support` (minimum support) and `min_confidence` (minimum confidence) are set. In the Sichuan Opera facial makeup domain, it is preferable to set `min_support` between 0.05 and 0.1 to capture specific patterns that, while not frequently occurring, are highly representative; and to set `min_confidence` between 0.5 and 0.7 to ensure the high reliability of the mined rules. Furthermore, the `max_len` parameter needs to be configured to limit the maximum length of frequent itemsets (e.g., set to 3) to focus on the direct mapping between visual features and cultural semantics.

[0065] During the model execution and post-processing phase, the `apriori()` function is called to input a boolean dataset and output a list of frequent itemsets. Then, the `association_rules()` function is called, specifying `lift` as the evaluation metric, to generate a rule table containing antecedents, consequents, support, confidence, and lift. At this point, the system does not directly output this table; instead, it attaches a custom Callback Function to read the antecedent index from the rule table, backtracks to the original matrix from step S2 to calculate the visual advantage, and applies the weighting formula proposed in this implementation to update the weight column in the rule table. Finally, the original `lift` column is replaced with a new symbolic binding strength column, completing the model's specific application.

[0066] Furthermore, step S304 specifically includes the following sub-steps:

[0067] S3041. Based on the prior and posterior probabilities obtained from statistics, calculate the ratio of the posterior probability to the prior probability to generate the basic lift degree that characterizes the statistical correlation.

[0068] S3042. By introducing a preset visual weight adjustment coefficient, the basic lift is nonlinearly amplified or suppressed using a weighted correction formula to calculate the symbol binding strength; the weighted correction formula is:

[0069] ;

[0070] Among them, the Representing visual feature terms Cultural semantic tags The symbol binding strength between them; Indicated in visual feature terms Cultural semantic tags appearing under certain conditions The posterior probability of occurrence; Cultural semantic tags Prior probabilities in the transaction dataset; This represents the preset visual weight adjustment coefficient; the Representing visual feature terms The mean value of visual dominance;

[0071] S3043. Compare the calculated symbol binding strength with the preset symbol binding strength threshold, remove combinations with strength values ​​less than or equal to the threshold, and store the remaining combinations in the database according to the triple structure of visual feature item - cultural semantic label item - symbol binding strength to construct a strong association rule library for Facebook.

[0072] Specifically, the above implementation receives the probability statistics output from the association mining stage and performs a division operation on each candidate association pair of visual feature items and cultural semantic tag items to be evaluated. The posterior probability of the cultural semantic tag appearing under specific visual feature conditions is used as the numerator, and the prior probability of the cultural semantic tag appearing in global transactions is used as the denominator to calculate the base lift. This index mathematically eliminates false strong association misjudgments caused by the high frequency of semantic tags and characterizes the actual amplification factor of the appearance of visual features on the probability of semantic tag appearance. The calculation module then reads the average visual dominance obtained from backtracking calculation and constructs a nonlinear gain function in combination with a preset visual weight adjustment coefficient. This gain function is then applied to the base lift to calculate the base lift. The weighted correction formula is applied, which essentially transforms the prior knowledge that the larger the color coverage area or the more complex the pattern structure of Sichuan Opera facial makeup, the more significant its cultural meaning, into mathematical weights. This allows the generated symbol binding strength to simultaneously reflect the co-occurrence frequency at the data level and the physical salience at the visual level. All calculated candidate association pairs are traversed, and the obtained symbol binding strength values ​​are compared and verified with preset strength thresholds. Marginal rules that satisfy statistical association but have weak visual expressiveness are automatically filtered out. The remaining strong association pairs after filtering are serialized and encoded according to a standardized triple data structure and written into a graph database or relational database. Finally, the instantiation and storage of the facial makeup strong association rule library that can support the subsequent construction of a semantic network are completed.

[0073] Furthermore, regarding the symbol binding strength calculation model (i.e., the specific implementation environment of the formula) involved in this solution, its exemplary setup and usage process is as follows:

[0074] During the computing environment setup phase, this model is typically embedded in Python's data analysis pipeline as a post-processing module for association rule mining, utilizing NumPy or Pandas libraries for efficient vectorized computation. First, a basic rule DataFrame containing fields such as [antecedents, consequents, support, confidence, lift] needs to be constructed. The visual_dominance column calculated in step S303 is then merged into this DataFrame to ensure that each rule corresponds to the physical attribute value of its visual feature.

[0075] In the parameter configuration and formula definition phase, a visual weight adjustment coefficient is defined. This coefficient is a hyperparameter and is calibrated through expert system scoring or small-sample testing. For example, if a strong emphasis on the semantic weight of large color blocks (such as the entire face) is desired, the visual weight adjustment coefficient can be set to a larger value (1.0); if the authenticity of the statistical results is to be maintained with only minor adjustments, a smaller value (0.2) can be set. The calculation logic is encapsulated as a vectorized function that takes the `lift` column and the `visual_dominance` column as input arrays and executes... The broadcast operation is used to complete the strength correction of the rule.

[0076] During the threshold filtering and database insertion phase, a threshold_bs (symbolic binding strength threshold) is set. This threshold is determined based on the quantiles of the data distribution (e.g., taking the top 20% of high-scoring rules). Filtering is completed using the set filtering conditions. Subsequently, ORM technology is used to map the cleaned DataFrame row data to database entity objects. The field mapping relationship is: visual_feature -> antecedent, semantic_label -> consequent, binding_strength -> weight. Finally, a batch insertion operation is performed to complete the rule base construction.

[0077] Furthermore, step S4 specifically includes the following sub-steps:

[0078] S401. Instantiate the visual feature items in the strong association rule base of the face as face visual nodes, and instantiate the cultural semantic label items as face semantic nodes. Read the symbol binding strength as the edge weight connecting each node to construct the face semantic network.

[0079] S402. Define the propagation potential energy, and iteratively update the propagation potential energy value of each node by traversing the Facebook semantic network using a graph theory-based iterative algorithm until convergence.

[0080] S403. Sort the converged propagation potential energy values ​​in descending order and select the top-ranked facial visual nodes as the core facial pattern; backtrack the standardized facial dataset based on the index of the core facial pattern, extract the corresponding high-resolution local images, and use vectorization algorithms to perform contour extraction and geometric reconstruction to generate vectorized facial components.

[0081] S404. Encapsulate the symbol binding strength and the converged propagation potential energy value as metadata into the file attributes of the vectorized face component to generate a digital propagation material library containing self-explanatory data.

[0082] Specifically, the above implementation parses the strongly associated rule base of facial makeup stored in the database, and uses a graph database or in-memory graph computing framework to map the antecedent visual feature item of each rule to the source node in the graph and the consequent cultural semantic label item to the target node. At the same time, it reads the symbol binding strength value in the rule attributes and directly assigns it as the edge weight attribute connecting the source node and the target node, thereby constructing a weighted heterogeneous facial makeup semantic network that can intuitively represent the mapping relationship between the form and content of Sichuan opera facial makeup. The computing module then initializes the propagation potential energy state vector of all nodes in the network and starts a graph iterative algorithm based on the PageRank or HITS idea. In each iteration, according to the transition probability matrix defined by the edge weight, the energy of high semantic nodes is diffused and accumulated to strongly bound visual nodes. This process continues until the difference norm of the potential energy vector between the two iterations is less than the preset convergence threshold to ensure that the calculation result reflects the true meaning of the facial makeup. The system establishes a stable global network structure, sorts and extracts the potential energy of converged nodes, and locks the most representative core facial pattern visual nodes according to the Top-K strategy. It uses the sample index associated with the node to retrieve the original high-definition image data in the standardized dataset and extract the corresponding region of interest. It calls computer vision algorithms to perform adaptive binarization segmentation and morphological closing operations on the image of the region to obtain a clear pattern binary mask. It uses a Bézier curve fitting algorithm to track the pixel gradient changes at the mask edge and calculates the coordinates of control points to generate a mathematically described closed vector path. When generating vector files in SVG or EPS format, the system uses XML namespace extension technology to write the symbol binding strength calculated in the association analysis and the propagation potential energy value obtained in the graph calculation into the metadata header area of ​​the file, thereby generating a digital dissemination material library that contains both high-fidelity visual morphology and core cultural evaluation indicators.

[0083] Preferably, the graph iterative algorithm is implemented using Python's NetworkX library during the construction and execution phases. A directed weighted graph object, DiGraph, is constructed. The association rule base is traversed, and visual and semantic nodes are added using `add_node()`, while edges are added using `add_edge(weight=binding_strength)`. In the algorithm configuration, a damping coefficient β is defined, typically set to 0.85, along with the maximum number of iterations. An iterative function is written that performs matrix operations in each loop iteration... Where M is the transition matrix and E is the average distribution vector, updating the potential values ​​of all nodes. This is the updated potential energy vector. This is the potential energy vector from the previous round. When detected... The iteration terminates at a certain point, and a dictionary-formatted table of node potential energy is output, completing the quantization and selection of core symbols.

[0084] In the setup and execution phase of vectorization reconstruction technology, OpenCV and Potrace (or svgwrite library) are integrated. The ROI bitmap corresponding to the core symbols is read, converted to grayscale using cv2.cvtColor, and adaptively thresholded using cv2.adaptiveThreshold to handle uneven lighting, generating a clean binary black and white image. Further, noise is removed and edges are smoothed. The processed bitmap data stream is then transmitted to the tracking algorithm module. This module identifies the boundaries of pixel connected components, fits them to a series of cubic Bézier curves, converts these curve coordinates into SVG path data, and sets the fill attribute based on the dominant RGB value obtained from the rule library. Simultaneously, an XML parsing library is used to insert data tags such as Propagation_Potential into the SVG file. <metadata>Within the tag pair, the encapsulation of the intelligent vector component is completed.

[0085] Furthermore, step S402 specifically includes the following sub-steps:

[0086] S4021. Count the total number of nodes in the Facebook semantic network, set the initial value of the propagation potential of Facebook visual nodes and Facebook semantic nodes to the reciprocal of the total number of nodes, and construct the initial potential vector.

[0087] S4022. In the t-th iteration, traverse each target node, retrieve all source nodes pointing to the target node, extract the current propagation potential of the source node and the symbol binding strength connecting the two, and calculate the energy value input from the source node to the target node;

[0088] S4023. Introduce a damping coefficient, sum the obtained input energy values ​​and combine them with the damping coefficient in a weighted manner to obtain the new propagation potential energy value of the target node in the (t+1)th iteration;

[0089] S4024. Calculate the Euclidean distance between the potential energy vectors of the (t+1)th iteration and the tth iteration. If the distance is less than the preset convergence threshold, stop the iteration and output the final propagation potential energy value; otherwise, let t=t+1 and return to step S4022.

[0090] Specifically, the above implementation constructs a normalized initial potential vector by statistically analyzing the total number of visual and semantic nodes in the Facebook semantic network and setting the initial propagation potential of all nodes to the reciprocal of the total number of nodes. By setting the initial potential of all nodes to the reciprocal of the total number of nodes, the sum of the potentials of all nodes in the network is strictly equal to 1, thus conforming to the basic definition of the probability distribution vector of the Markov chain state transition matrix. This operation physically represents the unbiased assumption that the system holds absolute equal weight to all Facebook visual and semantic nodes at the initial moment, that is, it is assumed that before examining the network topology and symbol binding strength, each symbol node has completely equal cultural propagation potential. After entering the iterative calculation process, for each target node in the network, all directed propagation potential vectors are retrieved. The edge points to the source node of the target node and reads the symbol binding strength value stored in the edge attribute. The effective energy value transmitted from the source node to the target node is calculated by multiplying the current propagation potential energy of the source node by the normalized edge weight. A damping coefficient is introduced to simulate the natural decay and random jump probability in the cultural transmission process. The input energy of all source nodes is weighted and summed and then linearly combined with the damping term to obtain the updated potential energy value of the target node in the next time step. The Euclidean distance between the potential energy vectors generated by the two iterations is calculated and compared with the preset convergence threshold. If the distance difference does not meet the convergence requirement, the iteration round index is automatically increased and the energy calculation step is returned to continue execution until the overall potential energy distribution of the network tends to stabilize, and the final propagation potential energy value is output.

[0091] Furthermore, step S403 specifically includes the following sub-steps:

[0092] S4031. Obtain the original sample ID and region coordinate information corresponding to the core face pattern, and crop out the corresponding region of interest image patch from the high-resolution image source of the face normalization dataset;

[0093] S4032. Perform Gaussian filtering to denoise the image patch of the region of interest, and use an adaptive threshold segmentation algorithm to separate the pattern from the skin background to generate a binary pattern outline bitmap;

[0094] S4033. Track the edge pixels of the pattern outline bitmap, calculate the coordinates of the control points using the least squares method, and transform the pixel edges into mathematically described closed vector paths;

[0095] S4034. Read the main color value recorded in the structured feature association matrix of the face pattern, fill the internal area of ​​the closed vector path with color, and combine the filled path objects to generate a vectorized face component.

[0096] Specifically, the above implementation retrieves the metadata table in the face standardization dataset based on the unique index ID of the core face pattern to obtain the storage path of the original image and the precise pixel coordinates of the ROI region. It then uses the matrix slicing function of the image processing library to losslessly crop local image patches containing specific textures from the high-resolution source image. To address potential skin texture noise or uneven lighting in these image patches, the system performs convolution operations with the image matrix using a Gaussian kernel function to smooth high-frequency noise. Following this, a local adaptive thresholding algorithm is applied, dynamically determining the binarization threshold for each pixel by calculating the gray-level mean or weighted sum within the local neighborhood of the image. This allows for the precise extraction of complex textures from the varied skin background. The high-contrast binary texture outline bitmap is generated, and the calculation module then initiates an edge tracking program to traverse the connected domain boundary pixels in the binary image. It collects discrete edge coordinate point sets and inputs them into the least squares fitting model. By minimizing the sum of squared errors, the control point parameters of the cubic Bézier curve or B-spline curve are solved, thereby transforming the jagged pixel edges into smooth and continuous closed vector paths. Finally, the system backtracks the structured feature association matrix to query the RGB or CMYK value of the main color of the sample to which the texture belongs. The color parameter is assigned to the fill attribute of the vector path, and all independent paths constituting the same texture are encapsulated into a grouped vector object, completing the reverse engineering reconstruction from pixel bitmap to editable vector component.

[0097] Furthermore, regarding the image vectorization algorithm and geometric reconstruction technology involved in the above embodiments, an exemplary setup and usage process is as follows: In the algorithm environment setup and preprocessing stage, OpenCV (for image vision processing) and Potrace (for bitmap to vector conversion) or SciPy (for curve fitting) libraries are typically integrated. First, the system reads the cropped ROI image, executes cv2.cvtColor to convert it to a grayscale image, and applies cv2.GaussianBlur to remove noise. The key step is to use cv2.adaptiveThreshold in conjunction with the cv2.ADAPTIVE_THRESH_GAUSSIAN_C parameter, which can effectively handle the lighting gradients caused by the curved surface of the face, generating a clean black and white binary image. To further optimize the contour, a morphological opening operation (cv2.morphologyEx) is usually performed to break up subtle adhering noise.

[0098] In the contour tracking and vector fitting stage, cv2.findContours is used to extract the pixel set of all independent contours. For each contour point set, instead of directly connecting the pixels, the scipy.optimize.least_squares or cv2.approxPolyDP algorithm is used. The algorithm sets an approximation accuracy epsilon and calculates a set of key control points and tangent handles that can cover the original pixel path with minimal geometric error. For complex Sichuan opera patterns (such as flame patterns), the contour is segmented and fitted into multiple cubic Bézier curves to ensure the continuity of the tangents at the connection points.

[0099] During the component generation and encapsulation phase, the SVG file structure is built using the svgwrite or lxml library, creating a... <path>The tag converts the calculated control point coordinates into SVG standard path instructions, for example, M for movement, C for Bezier curve, L for straight line, and Z for closure; further, it reads the main color value corresponding to the pattern from the correlation matrix (for example, #FF000) and writes it. <path>The `fill` attribute of the tag wraps all paths. <g>Within the (Group) tag, add a unique id attribute and save as an .svg file to generate a vectorized face component that can be directly rendered on the web or edited by design software.

[0100] Furthermore, in step S402, the specific process for iteratively updating the propagation potential energy values ​​of each node is expressed as follows:

[0101] ;

[0102] Among them, the Cultural semantic tags The propagation potential energy value at the (t+1)th iteration; This indicates the preset damping coefficient; the This represents all cultural semantic tags in the Facebook semantic network. Visual features The set; the Representing visual feature terms The propagation potential energy value at the t-th iteration; Representing visual feature terms Cultural semantic tags The symbol binding strength between them; Representing facial visual nodes All cultural semantic tags The set. It should be noted that the above implementation constructs an energy flow model based on the weighted PageRank concept to simulate the attention allocation mechanism and semantic verification process of Sichuan Opera facial makeup culture in the audience's cognitive network. The left side of the formula... The output represents the global importance score accumulated throughout the semantic network for a specific cultural semantic label s (such as "loyalty" or "fierceness") at iteration t+1. The right side of the formula... Part of it constitutes the damping reset term, in which As a damping coefficient, a value of 0.85 is used to simulate the random walk probability in the cultural transmission process. This assumes that the audience has a certain probability of randomly associating the cultural semantics without relying on current visual guidance. This ensures that even isolated semantic nodes in the network can obtain a non-zero basic survival potential, thus avoiding the black hole effect during computation. This is achieved through the summation term... Iterate through all visual feature terms v (such as "bright red face" or "silkworm eyebrows") that can point to the semantic label s and measure their propagation potential at the previous time t. The weighted summation is performed using the input source as the basis for the calculation. The core unit constituting the energy transfer probability matrix is ​​the molecule. The symbol binding strength calculated in step S3 quantifies the direct associative strength of a specific visual feature v stimulating a specific semantic s. The denominator... Then calculate the visual feature v pointing to all possible semantic labels in the network. The sum of the strengths is used to achieve normalization. The physical meaning of this fraction is to evaluate the "specific contribution ratio" of visual feature v to semantic label s. That is, in the field of Sichuan Opera, if "black face" can represent both "uprightness" and "rudeness", but its binding strength with "uprightness" is much greater than that with "rudeness", then the potential energy accumulated by "black face" will flow more to the "uprightness" node proportionally through this fraction. The logic of the entire iterative operation is essentially to simulate a dynamic cultural consensus formation process. That is, in the initial stage, all nodes have equal potential energy. As the iteration progresses, semantic labels with high visual dominance and strong binding relationships with other high-potential visual nodes will continuously absorb energy in the network, and finally converge to obtain the potential energy value. It no longer merely reflects local correlation, but represents the core position and dominance of this cultural semantic in the complex expressive system of Sichuan Opera woven from countless visual symbols, thereby guiding subsequent steps to accurately select the most representative cultural symbols for digital reconstruction and dissemination.

[0103] Example 2

[0104] Furthermore, as a preferred embodiment of the above embodiments, step S5 is proposed: collecting the preference characteristics of the target audience, combining them with the calculated dissemination potential energy value, calculating the cultural resonance between the audience and the material through cosine similarity matching, and generating a distribution list based on the cultural resonance. Step S5 specifically includes the following sub-steps:

[0105] S501. Collect the target audience's historical browsing records and interaction behavior data, map them to the feature space of the Facebook standardized dataset, count the audience's preference frequency for visual feature items and cultural semantic tag items, and construct a multi-dimensional audience preference feature vector.

[0106] S502. Traverse the digital communication material library, read the visual feature items and cultural semantic tag items encapsulated in the metadata of each vectorized face component, construct the material attribute feature vector consistent with the dimension of the audience preference feature vector, and read the communication potential value in the metadata;

[0107] S503. Calculate the cosine similarity between the audience preference feature vector and the material attribute feature vector, and use the propagation potential as a weighting coefficient to non-linearly weight the cosine similarity, through: The cultural resonance between the audience and the material is calculated, wherein the... Represents audience preference feature vector Material attribute feature vector The degree of cultural resonance between them; This represents the propagation potential energy value of the material read from the metadata; the... This represents the preset potential energy influence factor; the... and These represent the eigenvalues ​​of the vector in the i-th dimension; n represents the total dimension of the eigenvectors; Represents the magnitude of the vector;

[0108] S504. Sort the materials in the digital communication material library in descending order of cultural resonance, and select the top-ranked materials to generate a distribution list for the target audience.

[0109] Specifically, the above implementation method collects historical browsing and clicking behavior data of different Sichuan Opera cultural content by accessing the terminal interaction log interface of the target audience. Using the standardized facial makeup dataset constructed in step S1 as a feature dictionary, this unstructured behavioral data is projected into a unified high-dimensional feature space. By accumulating the frequency of user attention to specific visual features such as "bright red main color" or "bat pattern" and cultural semantic tags such as "loyalty" or "fierceness," a multi-dimensional audience preference feature vector that accurately characterizes the user's aesthetic and cognitive tendencies is constructed. Each vectorized facial makeup component stored in the digital communication material library is traversed, and the attribute information encapsulated in its metadata header is parsed to determine the visual features carried by the component. The semantic tags are mapped to material attribute feature vectors that are completely consistent with the dimensions of the audience preference feature vectors to ensure mathematical comparability between the two in the vector space. Simultaneously, the propagation potential value obtained by the component in the semantic network iterative computation, recorded in the metadata, is extracted as an independent scalar to measure the cultural authority of the material. The computation module uses a vector space model to calculate the cosine value of the angle between the audience preference feature vector and the material attribute feature vector to quantify the basic similarity between content features and user interests. To correct the "information cocoon" effect that may be caused by simply recommending based on interests and to strengthen the dissemination of core cultural symbols, the system introduces propagation potential as a non-linear weighting factor to correct the cosine similarity, according to the formula... The calculation of cultural resonance between the audience and the materials uses the physical meaning of the formula to represent the core position of Sichuan Opera facial makeup patterns in the cultural network by using the potential energy of dissemination. The index is used as a hyperparameter to amplify or reduce the influence of cultural authority on the recommendation results. When the material itself has extremely high cultural dissemination potential energy, which represents the classic patterns of Sichuan Opera, even if the cosine similarity between the user's historical preferences and its geometric features is average, a higher final score can be obtained through the weighting of the potential energy term. This achieves a strategy upgrade from simply "catering to preferences" to "taking cultural guidance into account." Based on the calculated cultural resonance value, all candidate components in the material library are sorted in descending order. The Top-K truncation strategy is applied to select the top-ranked materials to generate a personalized distribution list for the specific audience, completing the technical closed loop from data mining to precise cultural dissemination.

[0110] Example 3

[0111] Furthermore, as a preferred embodiment of the above embodiments, a system for disseminating cultural symbols based on Sichuan Opera facial makeup data association analysis, as described in Embodiment 1, is proposed. The system includes a multimodal data standardization processing module, a feature quantization and matrix construction module, a visually enhanced association rule mining module, a semantic network analysis and digital reconstruction module, and a cultural resonance intelligent distribution module. The multimodal data standardization processing module is responsible for acquiring Sichuan Opera facial makeup images and text and performing facial key region segmentation. The module's built-in dual-stream feature extraction model architecture includes a deep convolutional neural network branch for processing image data to extract visual feature tensors that retain spatial information by removing fully connected layers from the convolutional structure, and a pre-trained natural language processing model branch for processing text data to generate semantic feature vectors through a multi-layer transform encoder structure. The feature quantization and matrix construction module uses clustering algorithms and edge detection operators to transform multimodal features into quantized color distribution ratios and pattern geometric feature vectors, and performs index alignment and numerical normalization to construct a structured feature association matrix. The visually enhanced association rule mining module converts the structured matrix into a transaction dataset and inputs it into the mining engine. This module employs an improved weighted association analysis model architecture. The bottom layer is a frequent itemset search layer based on the Apriori algorithm for calculating prior and posterior probabilities. An embedded visual dominance calculation layer backtracks to calculate the mean from the original values. The top layer deploys a non-linear weight correction unit to weight the base lift using visual dominance, thus outputting the symbol binding strength. The semantic network analysis and digital reconstruction module constructs a facial semantic network and generates digital materials. This module integrates a graph iterative calculation model and a vector reconstruction model. The graph iterative calculation model architecture is based on Markov chain ideas to construct a state transition matrix and energy distribution vector, and updates the node propagation potential energy through a damped iterative mechanism. The vector reconstruction model architecture uses an adaptive threshold segmentation layer to extract binary contours and a least-squares fitting layer to transform pixel edges into Bézier curve paths. The cultural resonance intelligent distribution module collects user behavior to construct preference vectors and combines them with material propagation potential energy for recommendation. This module employs a weighted vector space model architecture, including a basic cosine similarity calculation layer to measure feature matching and a potential energy weighting control layer to non-linearly adjust the similarity results using propagation potential energy values ​​to output cultural resonance.

[0112] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.< / g> < / path> < / path> < / metadata>

Claims

1. A method for the dissemination of cultural symbols based on the correlation analysis of Sichuan Opera facial makeup data, characterized in that, Includes the following steps: S1. Collect images of Sichuan opera facial makeup and corresponding text description data, segment the key facial regions of the images and extract multimodal features, and construct a standardized facial makeup dataset containing spatial information; S2. Read the output face normalization dataset, perform quantization calculation on the image data in the face normalization dataset, generate the distribution ratio of primary and secondary tones and geometric feature vectors of texture lines, perform index alignment and numerical normalization processing on the text semantic index vectors in the face normalization dataset, and construct a structured feature association matrix. S3. The generated structured feature association matrix is ​​converted into a transaction dataset and input into the association rule mining model based on the Apriori algorithm. By introducing visual dominance as a weighting factor, the symbol binding strength between visual features and cultural semantic labels is calculated to construct a strong association rule base for facial recognition. S4. Construct a semantic network based on the calculated symbol binding strength, use a graph iterative algorithm to calculate the cultural dissemination potential of each node in the network, identify and output core cultural symbols; reconstruct the visual form into vectorized components based on the core cultural symbols, and encapsulate the calculated symbol binding strength and cultural dissemination potential into metadata to generate a digital dissemination material library. Step S3 specifically includes the following sub-steps: S301. Convert the structured feature association matrix into a Boolean-type transaction dataset. Specifically, discretize the continuous numerical distribution ratios of primary and secondary hues and the geometric feature vectors of decorative lines in the structured feature association matrix to generate visual feature items. Map the text semantic index vectors in the structured feature association matrix directly to cultural semantic label items. S302. Input the generated transaction dataset into an association rule mining model based on the Apriori algorithm. Set minimum support and minimum confidence thresholds, iterate through and filter candidate sets that meet the threshold conditions, and calculate the prior probability of cultural semantic label items in the candidate sets, as well as the posterior probability of cultural semantic label items appearing under the condition that visual feature items appear. S303. By backtracking the original normalized values ​​in the structured feature association matrix, calculate the mean value of the distribution ratios of primary and secondary hues or the geometric feature vectors of decorative lines in the candidate sets across all supporting samples to generate visual dominance. S304. The generated visual advantage is used as a weighting factor to adjust the ratio of the prior probability to the posterior probability of the pair, thereby obtaining the symbolic binding strength between visual feature items and cultural semantic label items. Combinations with symbolic binding strength greater than a preset threshold are retained to construct a strong association rule base for facial recognition.

2. The method for disseminating cultural symbols based on Sichuan Opera facial makeup data association analysis according to claim 1, characterized in that, Step S1 specifically includes the following sub-steps: S101. Establish a multimodal raw database of Sichuan Opera facial makeup, collect a set of raw facial makeup images including front view, side view and close-up details, and collect a set of raw text descriptions corresponding to each image. S102. Read the original set of facial images, use image annotation tools to divide the forehead, eye sockets and jaw area into regions of interest, and generate region-annotated image data; and perform synonym merging and noise reduction on the original text description set to generate a standardized text sequence. S103. Input the generated region-labeled image data into a pre-trained convolutional neural network to extract the high-dimensional feature map of the image, output the visual feature tensor, and input the generated standardized text sequence into a natural language processing model to perform word embedding encoding and output the semantic feature vector. S104. The visual feature tensor and semantic feature vector are concatenated and formatted according to the unique sample ID index to construct a standardized Facebook dataset containing image features, text features and corresponding labels.

3. The method for disseminating cultural symbols based on Sichuan Opera facial makeup data association analysis as described in claim 1, characterized in that, Step S2 specifically includes the following sub-steps: S201. Read the image data from the standardized facial dataset, perform color segmentation on the pixels using the K-Means clustering algorithm, count the proportion of pixels belonging to the same cluster to the total area of ​​the facial, and generate the primary and secondary color distribution ratio. S202. Read the image data from the face standardization dataset, extract the texture skeleton using the Canny edge detection operator, calculate the average curvature, distribution density and gradient direction of the lines based on the skeleton pixels, and generate the geometric feature vector of the texture lines. S203. Read the semantic features in the standardized facial dataset, parse and extract the corresponding personality tag index and pattern classification index, and generate a text semantic index vector; S204. The generated primary and secondary color distribution ratios, geometric feature vectors of decorative lines, and text semantic index vectors are indexed and aligned according to the unique sample IDs, and the numerical features are normalized. Finally, they are spliced ​​together to construct a structured feature association matrix.

4. The method for disseminating cultural symbols based on Sichuan Opera facial makeup data association analysis as described in claim 3, characterized in that, The specific method for calculating the distribution ratio of primary and secondary colors is as follows: ； The specific method for calculating the distribution density value of the geometric feature vector of the decorative lines is as follows: ； Among them, the This represents the distribution ratio of the k-th color cluster in the image; the This represents image data from the Facebook normalized dataset; The Kronecker function is defined as having a value of 1 when the two parameters are equal, and 0 otherwise; Represents image coordinates The color cluster category to which the pixel belongs; the The total number of pixels in the image data; This represents the distribution density value in the geometric feature vector of the decorative lines; Represents image data in coordinates The binarized result after edge detection is 1 if it is a texture edge point, and 0 otherwise.

5. The method for disseminating cultural symbols based on Sichuan Opera facial makeup data association analysis as described in claim 4, characterized in that, Step S304 specifically includes the following sub-steps: S3041. Based on the prior and posterior probabilities obtained from statistics, calculate the ratio of the posterior probability to the prior probability to generate the basic lift degree that characterizes the statistical correlation. S3042. By introducing a preset visual weight adjustment coefficient, the basic lift is nonlinearly amplified or suppressed using a weighted correction formula to calculate the symbol binding strength; the weighted correction formula is: ； Among them, the Representing visual feature terms Cultural semantic tags The symbol binding strength between them; Indicated in visual feature terms Cultural semantic tags appearing under certain conditions The posterior probability of occurrence; Cultural semantic tags Prior probabilities in the transaction dataset; This represents the preset visual weight adjustment coefficient; the Representing visual feature terms The mean value of visual dominance; S3043. Compare the calculated symbol binding strength with the preset symbol binding strength threshold, remove combinations with strength values ​​less than or equal to the threshold, and store the remaining combinations in the database according to the triple structure of visual feature item - cultural semantic label item - symbol binding strength to construct a strong association rule library for Facebook.

6. The method for disseminating cultural symbols based on Sichuan Opera facial makeup data association analysis as described in claim 1, characterized in that, Step S4 specifically includes the following sub-steps: S401. Instantiate the visual feature items in the strong association rule base of the face as face visual nodes, and instantiate the cultural semantic label items as face semantic nodes. Read the symbol binding strength as the edge weight connecting each node to construct the face semantic network. S402. Define the propagation potential energy, and iteratively update the propagation potential energy value of each node by traversing the Facebook semantic network using a graph theory-based iterative algorithm until convergence. S403. Sort the converged propagation potential energy values ​​in descending order and select the top-ranked facial visual nodes as the core facial pattern; backtrack the standardized facial dataset based on the index of the core facial pattern, extract the corresponding high-resolution local images, and use vectorization algorithms to perform contour extraction and geometric reconstruction to generate vectorized facial components. S404. Encapsulate the symbol binding strength and the converged propagation potential energy value as metadata into the file attributes of the vectorized face component to generate a digital propagation material library containing self-explanatory data.

7. The method for disseminating cultural symbols based on Sichuan Opera facial makeup data association analysis as described in claim 6, characterized in that, Step S402 specifically includes the following sub-steps: S4021. Count the total number of nodes in the Facebook semantic network, set the initial value of the propagation potential of Facebook visual nodes and Facebook semantic nodes to the reciprocal of the total number of nodes, and construct the initial potential vector. S4022. In the t-th iteration, traverse each target node, retrieve all source nodes pointing to the target node, extract the current propagation potential energy of the source node and the symbolic binding strength connecting the two, and calculate the energy value input from the source node to the target node; S4023. Introduce a damping coefficient, sum the obtained input energy values ​​and combine them with the damping coefficient in a weighted manner to obtain the new propagation potential energy value of the target node in the (t+1)-th iteration; S4024. Calculate the Euclidean distance between the potential energy vectors of the (t+1)th iteration and the tth iteration. If the distance is less than the preset convergence threshold, stop the iteration and output the final propagation potential energy value; otherwise, let t=t+1 and return to step S4022.

8. A method for disseminating cultural symbols based on Sichuan Opera facial makeup data association analysis as described in claim 6, characterized in that, Step S403 specifically includes the following sub-steps: S4031. Obtain the original sample ID and region coordinate information corresponding to the core face pattern, and crop out the corresponding region of interest image patch from the high-resolution image source of the face normalization dataset; S4032. Perform Gaussian filtering to denoise the image patch of the region of interest, and use an adaptive threshold segmentation algorithm to separate the pattern from the skin background to generate a binary pattern outline bitmap; S4033. Track the edge pixels of the pattern outline bitmap, calculate the coordinates of the control points using the least squares method, and transform the pixel edges into mathematically described closed vector paths; S4034. Read the main color value recorded in the structured feature association matrix of the face pattern, fill the internal area of ​​the closed vector path with color, and combine the filled path objects to generate a vectorized face component.

9. A method for disseminating cultural symbols based on Sichuan Opera facial makeup data association analysis as described in claim 6, characterized in that, In step S402, the specific process for iteratively updating the propagation potential energy values ​​of each node is as follows: ； Among them, the Cultural semantic tags The propagation potential energy value at the (t+1)th iteration; This indicates the preset damping coefficient; the This represents all cultural semantic tags in the Facebook semantic network. Visual features The set; the Representing visual feature terms The propagation potential energy value at the t-th iteration; Representing visual feature terms Cultural semantic tags The symbol binding strength between them; Representing facial visual nodes All cultural semantic tags A set of.

10. A method for disseminating cultural symbols based on Sichuan Opera facial makeup data association analysis as described in claim 1, characterized in that, It also includes step S5: collecting the preference characteristics of the target audience, combining the calculated dissemination potential value, calculating the cultural resonance between the audience and the material through cosine similarity matching, and generating a distribution list based on the cultural resonance.

11. A method for disseminating cultural symbols based on Sichuan Opera facial makeup data association analysis as described in claim 10, characterized in that, Step S5 specifically includes the following sub-steps: S501. Collect the target audience's historical browsing records and interaction behavior data, map them to the feature space of the Facebook standardized dataset, count the audience's preference frequency for visual feature items and cultural semantic tag items, and construct a multi-dimensional audience preference feature vector. S502. Traverse the digital communication material library, read the visual feature items and cultural semantic tag items encapsulated in the metadata of each vectorized face component, construct the material attribute feature vector consistent with the dimension of the audience preference feature vector, and read the communication potential value in the metadata; S503. Calculate the cosine similarity between the audience preference feature vector and the material attribute feature vector, and use the propagation potential as a weighting coefficient to non-linearly weight the cosine similarity, through: The cultural resonance between the audience and the material is calculated, wherein the... Represents audience preference feature vector Material attribute feature vector The degree of cultural resonance between them; This represents the propagation potential energy value of the material read from the metadata; the... This represents the preset potential energy influence factor; the... and These represent the eigenvalues ​​of the vector in the i-th dimension; n represents the total dimension of the eigenvectors; Represents the magnitude of the vector; S504. Sort the materials in the digital communication material library in descending order of cultural resonance, and select the top-ranked materials to generate a distribution list for the target audience.

Images