Non-abandoned pattern deconstruction regeneration system and method based on multi-scale recurrent neural network
Through the deconstruction and regeneration of intangible cultural heritage patterns by multi-scale recurrent neural network system, the problem of inefficiency of traditional methods is solved, efficient deconstruction and diversified regeneration of intangible cultural heritage patterns is achieved, and the application of intangible cultural heritage culture in modern design is promoted.
Patent Information
- Application Number
- CN202510435962.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-08
- Publication Date
- 2025-08-15
AI Technical Summary
Traditional pattern analysis and creation are inefficient and greatly affected by subjective factors, and existing methods are difficult to effectively deconstruct the multi-scale characteristics and complex structural relationships of high-quality regeneration of intangible cultural heritage patterns.
A multi-scale recurrent neural network system is adopted, including data acquisition and preprocessing, multi-scale feature extraction, recurrent neural network, pattern deconstruction and pattern regeneration modules. Features are extracted through multi-scale convolutional neural networks, gated recurrent unit neural network captures dependencies, and variational autoencoder generates innovative patterns.
It realizes efficient deconstruction and diversified regeneration of intangible cultural heritage patterns, improves the accuracy and stability of model training, can generate innovative patterns that meet the needs of modern design, and promotes the inheritance and development of intangible cultural heritage culture.
Smart Images
Figure CN120495440A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intangible cultural heritage protection and inheritance, and specifically to an intangible cultural heritage pattern deconstruction and regeneration system and method based on a multi-scale recurrent neural network. Background Art
[0002] Traditional patterns within intangible cultural heritage contain rich historical and cultural information and are important vehicles for the inheritance of national culture. However, with the rapid changes in modern lifestyles, many intangible cultural heritage patterns face a difficult inheritance dilemma. Their traditional design concepts and artistic forms are difficult to directly apply to modern design scenarios, urgently requiring innovative methods for their preservation and regeneration.
[0003] Traditional pattern analysis and creation relies primarily on manual labor, which is not only inefficient but also subject to significant subjective influences, making it difficult to fully explore the complex structures and regularities inherent in patterns. With the development of computer technology, methods based on image processing and machine learning have begun to be applied to pattern research. However, existing methods have significant shortcomings when dealing with the multi-scale characteristics and complex structural relationships of intangible cultural heritage patterns. For example, while traditional convolutional neural networks excel at extracting image features, they have limited ability to capture the coherence and cyclical nature of pattern features at different scales, making them ineffective in accurately deconstructing and reproducing patterns with high quality. Summary of the Invention
[0004] The purpose of the present invention is to provide a system and method for deconstructing and regenerating intangible cultural heritage patterns based on a multi-scale recurrent neural network to solve the problems raised in the above background technology.
[0005] To achieve the above objectives, the present invention provides the following technical solutions: a system for deconstructing and regenerating intangible cultural heritage patterns based on a multi-scale recurrent neural network, comprising a data acquisition and preprocessing module, a multi-scale feature extraction module, a recurrent neural network module, a pattern deconstruction module, and a pattern regeneration module. The data acquisition and preprocessing module collects a wide range of intangible cultural heritage pattern image data, covering patterns from different regions, ethnic groups, and historical periods, performs grayscale conversion, noise reduction, and normalization preprocessing operations on the collected images, unifies the image size and format, and, at the same time, annotates the patterns to clarify the category, location, and mutual relationship information of each pattern element, thereby constructing an annotated intangible cultural heritage pattern dataset.
[0006] Multi-scale feature extraction module: adopts a multi-scale convolutional neural network structure, setting convolution kernels of different sizes at different layers of the network. Smaller convolution kernels are used to extract the detailed features of the pattern; larger convolution kernels are used to extract the global features of the pattern, and the detailed features and global features are fused to form a feature vector containing multi-scale information;
[0007] Recurrent Neural Network Module: This module introduces a gated recurrent unit (GRU) neural network structure, inputting the feature vector sequence output by the multi-scale feature extraction module into the GRU neural network. Leveraging the memory properties of the GRU neural network, it captures the dependencies and cyclic information between pattern features at different scales and structural levels. By learning from a large amount of intangible cultural heritage pattern data, the GRU neural network can understand the arrangement patterns and combinations of pattern elements, as well as the formation mechanism of the overall structure.
[0008] Pattern deconstruction module: Based on the pattern structure and feature information learned by the recurrent neural network, the module deconstructs the intangible cultural heritage patterns and breaks them down into basic element units, including lines, geometric shapes, and pattern sub-modules. The module also analyzes the connection, hierarchy, and spatial layout relationships between the element units. Through deconstruction, a structured representation of the pattern is obtained, providing a basis for subsequent pattern regeneration.
[0009] Pattern regeneration module: uses a variational autoencoder generative model to regenerate patterns. Specifically, the pattern structure information and random noise output by the deconstruction module are used as the input of the generative model. The generative model generates new pattern images based on the learned intangible cultural heritage pattern characteristics and structural rules. During the generation process, the style and complexity attributes of the regenerated pattern are controlled by adjusting the input random noise or adding a specific style control vector, thereby generating diverse and innovative patterns.
[0010] Preferably, the data acquisition and preprocessing module specifically includes the following contents:
[0011] Extensive collection of intangible cultural heritage pattern image data: This involves on-site visits to intangible cultural heritage studios, museums, and folk exhibition halls across the country to capture and photograph relevant patterns. Furthermore, using internet search engines, and in compliance with applicable laws and regulations, relevant pattern images are downloaded from professional intangible cultural heritage websites and academic databases. Pattern materials are also collected from intangible cultural heritage inheritors and folk handicraft enthusiasts to expand data sources.
[0012] Image preprocessing operation: First, the collected intangible cultural heritage pattern image data is uniformly converted into grayscale images, and the image size and pixel value are normalized, that is, the image size is uniformly adjusted to a specific size. At the same time, the collected intangible cultural heritage pattern image data is subjected to noise reduction using the Gaussian filtering algorithm to obtain preprocessed image data;
[0013] Pattern annotation and dataset construction: Each intangible cultural heritage pattern image after preprocessing is carefully annotated, including pattern elements, element positions, and their relationships. After the annotation is completed, it is reviewed. Images with inaccurate or incomplete annotations are returned to the annotators for correction. After multiple rounds of annotation and review, a dataset of intangible cultural heritage patterns with detailed annotations is constructed. This dataset will serve as important basic data for subsequent model training and research.
[0014] Preferably, the multi-scale feature extraction module specifically works as follows: first, the pre-processed image data obtained in the data acquisition and pre-processing module is input into the shallow layer of the convolutional neural network, and a 3×3 convolution kernel is used to perform a convolution operation to extract the texture of the lines and the local shape detail features of the pattern. After the convolution operation, an activation function is used to introduce nonlinear factors to enhance the expression ability of the network; then the above image enters the deep convolution layer, and the deep convolution layer is set to a 7×7 convolution kernel to extract the global features of the pattern; the extracted detail features and global features are up-sampled using bilinear interpolation, The small-sized feature map is spatially aligned with the large-sized feature map so that the detail feature map and the global feature map have the same spatial size. Then the detail feature map and the global feature map with the same size are spliced in the channel dimension and feature fusion is performed to obtain a new feature map containing multi-scale information. Finally, a global pooling operation is performed on the fused new feature map. The global pooling operation compresses the new feature map in the spatial dimension, averages or takes the maximum value of the feature value of each channel, and obtains a vector of fixed length. The vector obtained after global pooling is the feature vector containing multi-scale information.
[0015] Preferably, the specific working content of the recurrent neural network module is as follows:
[0016] A. Gated Recurrent Unit Neural Network Architecture Design
[0017] The number of hidden layer neurons in the GRU unit is set according to the complexity of the actual intangible cultural heritage project, and the number of GRU layers is also determined. The input format of the GRU module is then determined based on the feature vector sequence output by the multi-scale feature extraction module. The input format is: number of samples, time step, feature dimension, where the time step represents the length of the feature vector sequence and the feature dimension is the length of each feature vector. The output format is a vector related to the pattern features.
[0018] B. Data preparation and input
[0019] Obtain the feature vector sequence output by the multi-scale feature extraction module; divide the prepared feature vector sequence dataset into a training set, a validation set, and a test set, specifically in a ratio of 70%, 15%, and 15%; the training set is used for model parameter learning, the validation set is used to adjust the model's hyperparameters and monitor the model's training process to prevent overfitting, and the test set is used to evaluate the model's final performance;
[0020] C. Capturing dependency and loop information
[0021] The GRU module processes the input feature vector sequence in sequence according to time steps, starting from the beginning of the sequence and gradually updating the hidden state until the entire sequence is processed. In this process, the GRU unit uses its gating mechanism to capture the dependency and cyclic information between pattern features at different scales and structural levels. Due to the gating mechanism of the GRU module, the gated recurrent unit neural network is used to handle long-term dependency problems in the sequence;
[0022] D. Model training and optimization
[0023] During the training process, the data in the training set is continuously input into the GRU module, the value of the loss function is calculated, and the network parameters are updated through the optimizer. At the same time, the performance of the model is regularly evaluated on the validation set, and the training parameters are adjusted according to the validation results to prevent overfitting;
[0024] E. Model Evaluation and Application
[0025] The trained GRU module was evaluated using a test set, and relevant evaluation metrics were calculated to measure model performance. The evaluation results revealed the GRU model's ability to capture pattern feature dependencies and learn pattern structural mechanisms. The trained GRU module was then applied to actual intangible cultural heritage pattern processing tasks to analyze the arrangement and combination of pattern elements. In pattern generation tasks, the GRU module generated new intangible cultural heritage patterns based on the learned structural mechanisms.
[0026] Preferably, the specific working content of the pattern deconstruction module is:
[0027] Receive the output of the recurrent neural network: establish a data interface with the recurrent neural network module to ensure that the learned pattern structure and feature information can be accurately obtained; at the same time, verify the received data to check the integrity, accuracy and consistency of the data;
[0028] Basic element unit identification: Use the K-Means clustering algorithm to analyze and cluster the feature information output by the recurrent neural network; divide the patterns into different categories based on feature similarity, and each category corresponds to a basic element unit; based on the clustering results, classify and label each basic element unit, establish an element classification dictionary, and clarify the names and feature descriptions of different types of elements;
[0029] Element connection relationship analysis: Based on the spatial position information of the elements in the pattern image, the connection relationship between the elements is judged. Specifically, the connectivity is determined by calculating the distance between the elements and the overlapping area index. At the same time, combined with the feature information learned by the recurrent neural network, the feature association between the elements is analyzed, and the element connection relationship obtained by the analysis is annotated. The connection between the elements is represented by a graph structure, where the nodes in the graph represent the basic element units, the edges represent the connection relationship between the elements, and the edges are assigned weights to indicate the strength or importance of the connection;
[0030] Element hierarchical relationship analysis: Through edge detection and contour analysis, the occlusion and coverage of element units in the image are analyzed to determine the hierarchical relationship of the elements. That is, if one element unit partially or completely occludes another element unit, the former is in the upper layer. The analyzed element hierarchical relationship is then annotated, and hierarchical attributes are added to the nodes in the graph structure to indicate the hierarchical position of the elements in the pattern.
[0031] Element spatial layout relationship analysis: This describes the relative positions of elements by calculating the center of gravity and bounding box information of element units. The analyzed element spatial layout relationships are annotated, and layout attributes are added to edges in the graph structure to represent the spatial layout characteristics between elements.
[0032] Structured representation generation: Integrate the classification annotation, connection relationship annotation, hierarchical relationship annotation and spatial layout relationship annotation of basic element units to construct a structured representation of the pattern.
[0033] Preferably, the specific working contents of the pattern regeneration module are:
[0034] Data preparation and input: The structured representation information of the pattern is obtained from the pattern deconstruction module and converted into the format of VAE input. Then, a random noise vector with the same dimension as the latent space is sampled using a Gaussian distribution. The pattern structure information vector output by the deconstruction module and the random noise vector are combined as the input of the VAE encoder.
[0035] Pattern Regeneration and Style Control: Use the trained VAE to regenerate patterns. This involves inputting the pattern structure information and random noise output by the deconstruction module. The random noise value is adjusted to generate different patterns. The regenerated pattern image is then obtained through the encoder, reparameterization, and decoder process.
[0036] Pattern evaluation and screening: Select preset evaluation indicators to evaluate the quality of recycled patterns, and combine professional designers or intangible cultural heritage experts to evaluate the recycled patterns, specifically evaluating the cultural connotation and artistic value of the patterns; then, based on the evaluation results, screen out high-quality recycled patterns.
[0037] Preferably, the method for deconstructing and regenerating intangible cultural heritage patterns based on a multi-scale recurrent neural network comprises the following steps:
[0038] Step 1, data preparation stage: through the data acquisition and preprocessing module, a wealth of intangible cultural heritage pattern images are collected, divided into training set, validation set and test set according to a certain ratio, and data enhancement operations are performed on the training set images;
[0039] Step 2: Model training phase:
[0040] First, the multi-scale feature extraction module is trained, and the training set images are input into the multi-scale convolutional neural network to extract the multi-scale features of the pattern;
[0041] The feature vector sequence output by the multi-scale feature extraction module is input into the recurrent neural network module for training, and the parameters of the recurrent neural network are optimized so that it can learn the cyclic structure information of the pattern features;
[0042] The pattern deconstruction module is jointly trained with the multi-scale feature extraction module and the recurrent neural network module. By comparing the deconstruction results with the annotated pattern structure information, the network parameters are adjusted to enable the pattern deconstruction module to accurately deconstruct the pattern.
[0043] Finally, the VAE generative model is jointly trained with the previous modules. The patterns generated by the generative model are compared with the real patterns in terms of features and structure. Through variational inference, the parameters of the generative model are optimized so that it can generate high-quality reproduced patterns.
[0044] Step 3, pattern deconstruction stage: After preprocessing, the intangible cultural heritage pattern image to be deconstructed is sequentially input into the trained multi-scale feature extraction module, recurrent neural network module, and deconstruction module to obtain the pattern deconstruction result, that is, the structured representation of the basic elements and their relationships;
[0045] Step 4, pattern regeneration stage: According to the design requirements, the model is generated to input specific random noise or style control vectors, and the pattern structure information obtained by the deconstruction module is combined to generate new intangible cultural heritage patterns. The generated patterns are evaluated and screened, and the patterns that meet the requirements are selected as the final regeneration results.
[0046] Compared with existing technologies, the present invention offers the following benefits: Efficient data collection and preprocessing, building a high-quality dataset: Through multiple channels, including field visits, online downloads, and solicitation from intangible cultural heritage inheritors, the invention extensively collects intangible cultural heritage pattern image data, covering diverse regions, ethnic groups, and historical periods, significantly enriching the data source. The collected images undergo preprocessing, including grayscale conversion, noise reduction, and normalization, and are meticulously annotated with pattern elements, locations, and relationships. This meticulously annotated dataset, constructed after multiple rounds of review, provides a solid and reliable data foundation for subsequent model training, effectively improving its accuracy and stability.
[0047] Accurate multi-scale feature extraction comprehensively captures pattern information: A multi-scale convolutional neural network architecture is employed, utilizing convolution kernels of varying sizes to extract both detailed and global features of the pattern. This multi-scale information is then fused through upsampling, concatenation, and global pooling to form a feature vector rich in information. This approach accurately captures key information about patterns at different scales, providing a more comprehensive and detailed understanding of pattern characteristics than traditional methods, providing more accurate data support for subsequent deconstruction and regeneration.
[0048] Powerful recurrent neural networks learn pattern structure patterns: This module introduces a gated recurrent unit neural network structure, leveraging its memory properties to capture the dependencies and cyclical information between pattern features at different scales and structural levels. By learning from a large amount of intangible cultural heritage pattern data, this module can deeply understand the arrangement and combination patterns of pattern elements, as well as the formation mechanism of the overall structure. This allows for more accurate analysis of the arrangement and combination of pattern elements, providing a strong theoretical basis for pattern deconstruction and regeneration.
[0049] Scientific pattern deconstruction achieves structured representation: Based on information learned from recurrent neural networks, techniques such as the K-Means clustering algorithm are used to deconstruct intangible cultural heritage patterns into basic elemental units. The connections, hierarchies, and spatial layout relationships between these elements are then analyzed to ultimately construct a structured representation of the pattern. This scientific and rational deconstruction method clearly presents the inherent structure of the pattern, providing an intuitive and easy-to-use foundation for pattern regeneration.
[0050] Diverse pattern reproduction to meet innovative design needs: A variational autoencoder generates a model that uses the pattern structure information output by the deconstruction module and random noise as input for pattern reproduction. By adjusting the random noise or adding a style control vector, the style and complexity of the regenerated pattern can be flexibly controlled, generating a variety of innovative patterns. This not only helps integrate intangible cultural heritage patterns with modern design needs, but also provides more possibilities for innovative applications of intangible cultural heritage patterns, promoting the inheritance and development of intangible cultural heritage in modern society. BRIEF DESCRIPTION OF THE DRAWINGS
[0051] Figure 1 Schematic diagram of the system structure of the present invention;
[0052] Figure 2 Schematic diagram of the method of the present invention. DETAILED DESCRIPTION
[0053] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0054] See also Figure 1 The present invention provides a technical solution: a non-legacy pattern deconstruction and regeneration system based on a multi-scale recurrent neural network, comprising a data acquisition and preprocessing module, a multi-scale feature extraction module, a recurrent neural network module, a pattern deconstruction module, and a pattern regeneration module. The data acquisition and preprocessing module widely collects various types of non-legacy pattern image data, covering patterns from different regions, ethnic groups, and historical periods, performs grayscale conversion, noise reduction, and normalization preprocessing operations on the collected images, unifies the image size and format, and simultaneously annotates the patterns to clarify the category, position, and mutual relationship information of each pattern element, thereby constructing an annotated non-legacy pattern dataset.
[0055] The data acquisition and preprocessing module specifically includes the following contents:
[0056] Extensive collection of intangible cultural heritage pattern image data: This involves on-site visits to intangible cultural heritage studios, museums, and folk exhibition halls across the country to capture and photograph relevant patterns. Furthermore, using internet search engines, and in compliance with applicable laws and regulations, relevant pattern images are downloaded from professional intangible cultural heritage websites and academic databases. Pattern materials are also collected from intangible cultural heritage inheritors and folk handicraft enthusiasts to expand data sources.
[0057] Image preprocessing operation: First, the collected intangible cultural heritage pattern image data is uniformly converted into grayscale images, and the image size and pixel value are normalized, that is, the image size is uniformly adjusted to a specific size. At the same time, the collected intangible cultural heritage pattern image data is subjected to noise reduction using the Gaussian filtering algorithm to obtain preprocessed image data;
[0058] Pattern annotation and dataset construction: Each intangible cultural heritage pattern image after preprocessing is carefully annotated, including pattern elements, element positions, and their relationships. After the annotation is completed, it is reviewed. Images with inaccurate or incomplete annotations are returned to the annotators for correction. After multiple rounds of annotation and review, a dataset of intangible cultural heritage patterns with detailed annotations is constructed. This dataset will serve as important basic data for subsequent model training and research.
[0059] Multi-scale feature extraction module: adopts a multi-scale convolutional neural network structure, sets convolution kernels of different sizes at different levels of the network, and smaller convolution kernels are used to extract the detailed features of the pattern; larger convolution kernels are used to extract the global features of the pattern, and fuse the detailed features and global features to form a feature vector containing multi-scale information; the specific work content is as follows: first, the pre-processed image data obtained in the data acquisition and pre-processing module is input into the shallow layer of the convolutional neural network, and a 3×3 convolution kernel is used for convolution operation to extract the texture of the lines and the local shape details of the pattern. After the convolution operation, the activation function is used to introduce nonlinear factors to enhance the expression ability of the network; then the above image enters the deep convolution layer, and the deep convolution layer is set The convolution kernel is 7×7, which is used to extract the global features of the pattern; bilinear interpolation is used to upsample the extracted detail features and global features, so that the small-size feature map is spatially aligned with the large-size feature map, so that the detail feature map and the global feature map have the same spatial size, and then the detail feature map and the global feature map with the same size are spliced in the channel dimension, and feature fusion is performed to obtain a new feature map containing multi-scale information; finally, a global pooling operation is performed on the fused new feature map, which compresses the new feature map in the spatial dimension, averages or takes the maximum value of the feature value of each channel, and obtains a vector of fixed length. The vector obtained after global pooling is the feature vector containing multi-scale information.
[0060] Recurrent Neural Network Module: This module introduces a gated recurrent unit (GRU) neural network structure, inputting the feature vector sequence output by the multi-scale feature extraction module into the GRU neural network. Leveraging the memory properties of the GRU neural network, it captures the dependencies and cyclic information between pattern features at different scales and structural levels. By learning from a large amount of intangible cultural heritage pattern data, the GRU neural network can understand the arrangement patterns and combinations of pattern elements, as well as the formation mechanism of the overall structure.
[0061] The specific work content of the recurrent neural network module is as follows:
[0062] A. Gated Recurrent Unit Neural Network Architecture Design
[0063] The number of hidden layer neurons in the GRU unit is set according to the complexity of the actual intangible cultural heritage project, and the number of GRU layers is also determined. The input format of the GRU module is then determined based on the feature vector sequence output by the multi-scale feature extraction module. The input format is: number of samples, time step, feature dimension, where the time step represents the length of the feature vector sequence and the feature dimension is the length of each feature vector. The output format is a vector related to the pattern features.
[0064] B. Data preparation and input
[0065] Obtain the feature vector sequence output by the multi-scale feature extraction module; divide the prepared feature vector sequence dataset into a training set, a validation set, and a test set, specifically in a ratio of 70%, 15%, and 15%; the training set is used for model parameter learning, the validation set is used to adjust the model's hyperparameters and monitor the model's training process to prevent overfitting, and the test set is used to evaluate the model's final performance;
[0066] C. Capturing dependency and loop information
[0067] The GRU module processes the input feature vector sequence in sequence according to time steps, starting from the beginning of the sequence and gradually updating the hidden state until the entire sequence is processed. In this process, the GRU unit uses its gating mechanism to capture the dependency and cyclic information between pattern features at different scales and structural levels. Due to the gating mechanism of the GRU module, the gated recurrent unit neural network is used to handle long-term dependency problems in the sequence;
[0068] D. Model training and optimization
[0069] During the training process, the data in the training set is continuously input into the GRU module, the value of the loss function is calculated, and the network parameters are updated through the optimizer. At the same time, the performance of the model is regularly evaluated on the validation set, and the training parameters are adjusted according to the validation results to prevent overfitting;
[0070] E. Model Evaluation and Application
[0071] The trained GRU module was evaluated using the test set, and relevant evaluation metrics were calculated to measure the model's performance. The evaluation results revealed the GRU model's ability to capture pattern feature dependencies and learn pattern structural mechanisms. The trained GRU module was then applied to actual intangible cultural heritage pattern processing tasks to analyze the arrangement and combination of pattern elements. In the pattern generation task, the GRU module generated new intangible cultural heritage patterns based on the learned structural mechanisms.
[0072] Pattern deconstruction module: Based on the pattern structure and feature information learned by the recurrent neural network, the intangible cultural heritage patterns are deconstructed and decomposed into basic element units, including lines, geometric shapes, and pattern sub-modules. The connection, hierarchy, and spatial layout relationships between the element units are analyzed. Through deconstruction, a structured representation of the pattern is obtained, providing a basis for subsequent pattern regeneration. The specific work content is as follows:
[0073] Receive the output of the recurrent neural network: establish a data interface with the recurrent neural network module to ensure that the learned pattern structure and feature information can be accurately obtained; at the same time, verify the received data to check the integrity, accuracy and consistency of the data;
[0074] Basic element unit identification: Use the K-Means clustering algorithm to analyze and cluster the feature information output by the recurrent neural network; divide the patterns into different categories based on feature similarity, and each category corresponds to a basic element unit; based on the clustering results, classify and label each basic element unit, establish an element classification dictionary, and clarify the names and feature descriptions of different types of elements;
[0075] Element connection relationship analysis: Based on the spatial position information of the elements in the pattern image, the connection relationship between the elements is judged. Specifically, the connectivity is determined by calculating the distance between the elements and the overlapping area index. At the same time, combined with the feature information learned by the recurrent neural network, the feature association between the elements is analyzed, and the element connection relationship obtained by the analysis is annotated. The connection between the elements is represented by a graph structure, where the nodes in the graph represent the basic element units, the edges represent the connection relationship between the elements, and the edges are assigned weights to indicate the strength or importance of the connection;
[0076] Element hierarchical relationship analysis: Through edge detection and contour analysis, the occlusion and coverage of element units in the image are analyzed to determine the hierarchical relationship of the elements. That is, if one element unit partially or completely occludes another element unit, the former is in the upper layer. The analyzed element hierarchical relationship is then annotated, and hierarchical attributes are added to the nodes in the graph structure to indicate the hierarchical position of the elements in the pattern.
[0077] Element spatial layout relationship analysis: This describes the relative positions of elements by calculating the center of gravity and bounding box information of element units. The analyzed element spatial layout relationships are annotated, and layout attributes are added to edges in the graph structure to represent the spatial layout characteristics between elements.
[0078] Structured representation generation: Integrate the classification annotation, connection relationship annotation, hierarchical relationship annotation and spatial layout relationship annotation of basic element units to construct a structured representation of the pattern
[0079] Pattern regeneration module: uses a variational autoencoder generative model to regenerate patterns. Specifically, the pattern structure information and random noise output by the deconstruction module are used as the input of the generative model. The generative model generates new pattern images based on the learned intangible cultural heritage pattern characteristics and structural rules. During the generation process, the style and complexity attributes of the regenerated pattern are controlled by adjusting the input random noise or adding a specific style control vector, thereby generating diverse and innovative patterns.
[0080] Among them, the specific work content of the pattern regeneration module is:
[0081] Data preparation and input: The structured representation information of the pattern is obtained from the pattern deconstruction module and converted into the format of VAE input. Then, a random noise vector with the same dimension as the latent space is sampled using a Gaussian distribution. The pattern structure information vector output by the deconstruction module and the random noise vector are combined as the input of the VAE encoder.
[0082] Pattern Regeneration and Style Control: Use the trained VAE to regenerate patterns. This involves inputting the pattern structure information and random noise output by the deconstruction module. The random noise value is adjusted to generate different patterns. The regenerated pattern image is then obtained through the encoder, reparameterization, and decoder process.
[0083] Pattern evaluation and screening: Select preset evaluation indicators to evaluate the quality of recycled patterns, and combine professional designers or intangible cultural heritage experts to evaluate the recycled patterns, specifically evaluating the cultural connotation and artistic value of the patterns; then, based on the evaluation results, screen out high-quality recycled patterns.
[0084] Example 2
[0085] like Figure 2 The method for deconstructing and regenerating intangible cultural heritage patterns based on a multi-scale recurrent neural network includes the following steps:
[0086] Step 1, data preparation stage: through the data acquisition and preprocessing module, a wealth of intangible cultural heritage pattern images are collected, divided into training set, validation set and test set according to a certain ratio, and data enhancement operations are performed on the training set images;
[0087] Step 2: Model training phase:
[0088] First, the multi-scale feature extraction module is trained, and the training set images are input into the multi-scale convolutional neural network to extract the multi-scale features of the pattern;
[0089] The feature vector sequence output by the multi-scale feature extraction module is input into the recurrent neural network module for training, and the parameters of the recurrent neural network are optimized so that it can learn the cyclic structure information of the pattern features;
[0090] The pattern deconstruction module is jointly trained with the multi-scale feature extraction module and the recurrent neural network module. By comparing the deconstruction results with the annotated pattern structure information, the network parameters are adjusted to enable the pattern deconstruction module to accurately deconstruct the pattern.
[0091] Finally, the VAE generative model is jointly trained with the previous modules. The patterns generated by the generative model are compared with the real patterns in terms of features and structure. Through variational inference, the parameters of the generative model are optimized so that it can generate high-quality reproduced patterns.
[0092] Step 3, pattern deconstruction stage: After preprocessing, the intangible cultural heritage pattern image to be deconstructed is sequentially input into the trained multi-scale feature extraction module, recurrent neural network module, and deconstruction module to obtain the pattern deconstruction result, that is, the structured representation of the basic elements and their relationships;
[0093] Step 4, pattern regeneration stage: According to the design requirements, the model is generated to input specific random noise or style control vectors, and the pattern structure information obtained by the deconstruction module is combined to generate new intangible cultural heritage patterns. The generated patterns are evaluated and screened, and the patterns that meet the requirements are selected as the final regeneration results.
[0094] The present invention covers five modules: data acquisition and preprocessing, multi-scale feature extraction, recurrent neural network, pattern deconstruction, and pattern regeneration. The data acquisition and preprocessing module widely collects various types of intangible cultural heritage pattern images, and constructs a data set after grayscale conversion, noise reduction, normalization, and annotation. The multi-scale feature extraction module uses convolution kernels of different sizes to extract details and global features and fuse them. The recurrent neural network module introduces gated recurrent units to capture the dependencies and cyclic information between pattern features. The pattern deconstruction module decomposes the pattern into basic element units based on the learning results of the recurrent neural network and analyzes their relationships to obtain a structured representation. The pattern regeneration module uses a variational autoencoder to generate new patterns with deconstructed information and random noise as input, and can control its style and complexity, thereby overcoming the shortcomings of traditional methods and being able to efficiently process the multi-scale features and complex structural relationships of intangible cultural heritage patterns. It is of great significance in the protection and inheritance of intangible cultural heritage patterns. It can not only deeply explore the intrinsic value of patterns, but also promote the integration of intangible cultural heritage culture and modern design, provide strong support for the innovative development of intangible cultural heritage patterns in modern society, and promote the inheritance and innovation of intangible cultural heritage.
[0095] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.
Claims
1. A system for deconstructing and regenerating intangible cultural heritage patterns based on a multi-scale recurrent neural network, comprising a data acquisition and preprocessing module, a multi-scale feature extraction module, a recurrent neural network module, a pattern deconstruction module, and a pattern regeneration module, characterized by: The data acquisition and preprocessing module collects a wide range of intangible cultural heritage pattern image data, covering patterns from different regions, nationalities, and historical periods. It performs grayscale conversion, noise reduction, and normalization preprocessing on the collected images, unifies the image size and format, and annotates the patterns to clarify the category, location, and relationship information of each pattern element, thereby constructing an annotated intangible cultural heritage pattern dataset. Multi-scale feature extraction module: adopts a multi-scale convolutional neural network structure, setting convolution kernels of different sizes at different layers of the network. Smaller convolution kernels are used to extract the detailed features of the pattern; larger convolution kernels are used to extract the global features of the pattern, and the detailed features and global features are fused to form a feature vector containing multi-scale information; Recurrent Neural Network Module: This module introduces a gated recurrent unit (GRU) neural network structure, inputting the feature vector sequence output by the multi-scale feature extraction module into the GRU neural network. Leveraging the memory properties of the GRU neural network, it captures the dependencies and cyclic information between pattern features at different scales and structural levels. By learning from a large amount of intangible cultural heritage pattern data, the GRU neural network can understand the arrangement patterns and combinations of pattern elements, as well as the formation mechanism of the overall structure. Pattern deconstruction module: Based on the pattern structure and feature information learned by the recurrent neural network, the module deconstructs the intangible cultural heritage patterns and breaks them down into basic element units, including lines, geometric shapes, and pattern sub-modules. The module also analyzes the connection, hierarchy, and spatial layout relationships between the element units. Through deconstruction, a structured representation of the pattern is obtained, providing a basis for subsequent pattern regeneration. Pattern regeneration module: uses a variational autoencoder generative model to regenerate patterns. Specifically, the pattern structure information and random noise output by the deconstruction module are used as the input of the generative model. The generative model generates new pattern images based on the learned intangible cultural heritage pattern characteristics and structural rules. During the generation process, the style and complexity attributes of the regenerated pattern are controlled by adjusting the input random noise or adding a specific style control vector, thereby generating diverse and innovative patterns.
2. The intangible cultural heritage pattern deconstruction and regeneration system based on a multi-scale recurrent neural network according to claim 1 is characterized by: The data acquisition and preprocessing module specifically includes the following contents: Extensive collection of intangible cultural heritage pattern image data: This involves on-site visits to intangible cultural heritage studios, museums, and folk exhibition halls across the country to capture and photograph relevant patterns. Furthermore, using internet search engines, and in compliance with applicable laws and regulations, relevant pattern images are downloaded from professional intangible cultural heritage websites and academic databases. Pattern materials are also collected from intangible cultural heritage inheritors and folk handicraft enthusiasts to expand data sources. Image preprocessing operation: First, the collected intangible cultural heritage pattern image data is uniformly converted into grayscale images, and the image size and pixel value are normalized, that is, the image size is uniformly adjusted to a specific size. At the same time, the collected intangible cultural heritage pattern image data is subjected to noise reduction using the Gaussian filtering algorithm to obtain preprocessed image data; Pattern annotation and dataset construction: Each intangible cultural heritage pattern image after preprocessing is carefully annotated, including pattern elements, element positions, and their relationships. After the annotation is completed, it is reviewed. Images with inaccurate or incomplete annotations are returned to the annotators for correction. After multiple rounds of annotation and review, a dataset of intangible cultural heritage patterns with detailed annotations is constructed. This dataset will serve as important basic data for subsequent model training and research.
3. The intangible cultural heritage pattern deconstruction and regeneration system based on a multi-scale recurrent neural network according to claim 1 is characterized by: The specific working content of the multi-scale feature extraction module is as follows: first, the pre-processed image data obtained in the data acquisition and pre-processing module is input into the shallow layer of the convolutional neural network, and a 3×3 convolution kernel is used to perform a convolution operation to extract the texture of the lines and the local shape detail features of the pattern. After the convolution operation, an activation function is used to introduce nonlinear factors to enhance the expression ability of the network; then the above image enters the deep convolution layer, and the deep convolution layer is set to a 7×7 convolution kernel to extract the global features of the pattern; the extracted detail features and global features are up-sampled using bilinear interpolation to make the small The feature map of the small size is spatially aligned with the feature map of the large size to achieve the same spatial size of the detail feature map and the global feature map. Then the detail feature map and the global feature map with the same size are spliced in the channel dimension and feature fusion is performed to obtain a new feature map containing multi-scale information. Finally, the fused new feature map is subjected to a global pooling operation. The global pooling operation compresses the new feature map in the spatial dimension, averages or takes the maximum value of the feature value of each channel, and obtains a vector of fixed length. The vector obtained after global pooling is the feature vector containing multi-scale information.
4. The intangible cultural heritage pattern deconstruction and regeneration system based on a multi-scale recurrent neural network according to claim 1, characterized in that: The specific working content of the recurrent neural network module is as follows: A. Gated Recurrent Unit Neural Network Architecture Design The number of hidden layer neurons in the GRU unit is set according to the complexity of the actual intangible cultural heritage project, and the number of GRU layers is also determined. The input format of the GRU module is then determined based on the feature vector sequence output by the multi-scale feature extraction module. The input format is: number of samples, time step, feature dimension, where the time step represents the length of the feature vector sequence and the feature dimension is the length of each feature vector. The output format is a vector related to the pattern features. B. Data preparation and input Obtain the feature vector sequence output by the multi-scale feature extraction module; divide the prepared feature vector sequence dataset into a training set, a validation set, and a test set, specifically in a ratio of 70%, 15%, and 15%; the training set is used for model parameter learning, the validation set is used to adjust the model's hyperparameters and monitor the model's training process to prevent overfitting, and the test set is used to evaluate the model's final performance; C. Capturing dependency and loop information The GRU module processes the input feature vector sequence in sequence according to time steps, starting from the beginning of the sequence and gradually updating the hidden state until the entire sequence is processed. In this process, the GRU unit uses its gating mechanism to capture the dependency and cyclic information between pattern features at different scales and structural levels. Due to the gating mechanism of the GRU module, the gated recurrent unit neural network is used to handle long-term dependency problems in the sequence; D. Model training and optimization During the training process, the data in the training set is continuously input into the GRU module, the value of the loss function is calculated, and the network parameters are updated through the optimizer. At the same time, the performance of the model is regularly evaluated on the validation set, and the training parameters are adjusted according to the validation results to prevent overfitting; E. Model Evaluation and Application The trained GRU module is evaluated using the test set, and relevant evaluation indicators are calculated to measure the performance of the model. The evaluation results show the ability of the GRU model in capturing the dependency relationship of pattern features and learning the pattern structure mechanism; The trained GRU module is then applied to actual intangible cultural heritage pattern processing tasks to analyze the arrangement rules and combination methods of pattern elements. In the pattern generation task, the GRU module generates new intangible cultural heritage patterns based on the learned structural mechanism.
5. The intangible cultural heritage pattern deconstruction and regeneration system based on a multi-scale recurrent neural network according to claim 1 is characterized by: The specific work content of the pattern deconstruction module is: Receive the output of the recurrent neural network: establish a data interface with the recurrent neural network module to ensure that the learned pattern structure and feature information can be accurately obtained; at the same time, verify the received data to check the integrity, accuracy and consistency of the data; Basic element unit identification: Use the K-Means clustering algorithm to analyze and cluster the feature information output by the recurrent neural network; divide the patterns into different categories based on feature similarity, and each category corresponds to a basic element unit; based on the clustering results, classify and label each basic element unit, establish an element classification dictionary, and clarify the names and feature descriptions of different types of elements; Element connection relationship analysis: Based on the spatial position information of the elements in the pattern image, the connection relationship between the elements is judged. Specifically, the connectivity is determined by calculating the distance between the elements and the overlapping area index. At the same time, combined with the feature information learned by the recurrent neural network, the feature association between the elements is analyzed, and the element connection relationship obtained by the analysis is annotated. The connection between the elements is represented by a graph structure, where the nodes in the graph represent the basic element units, the edges represent the connection relationship between the elements, and the edges are assigned weights to indicate the strength or importance of the connection; Element hierarchical relationship analysis: Through edge detection and contour analysis, the occlusion and coverage of element units in the image are analyzed to determine the hierarchical relationship of the elements. That is, if one element unit partially or completely occludes another element unit, the former is in the upper layer. The analyzed element hierarchical relationship is then annotated, and hierarchical attributes are added to the nodes in the graph structure to indicate the hierarchical position of the elements in the pattern. Element spatial layout relationship analysis: This describes the relative positions of elements by calculating the center of gravity and bounding box information of element units. The analyzed element spatial layout relationships are annotated, and layout attributes are added to edges in the graph structure to represent the spatial layout characteristics between elements. Structured representation generation: Integrate the classification annotation, connection relationship annotation, hierarchical relationship annotation and spatial layout relationship annotation of basic element units to construct a structured representation of the pattern.
6. The intangible cultural heritage pattern deconstruction and regeneration system based on a multi-scale recurrent neural network according to claim 1 is characterized by: The specific working content of the pattern regeneration module is: Data preparation and input: The structured representation information of the pattern is obtained from the pattern deconstruction module and converted into the format of VAE input. Then, a random noise vector with the same dimension as the latent space is sampled using a Gaussian distribution. The pattern structure information vector output by the deconstruction module and the random noise vector are combined as the input of the VAE encoder. Pattern Regeneration and Style Control: Use the trained VAE to regenerate patterns. This involves inputting the pattern structure information and random noise output by the deconstruction module. The random noise value is adjusted to generate different patterns. The regenerated pattern image is then obtained through the encoder, reparameterization, and decoder process. Pattern evaluation and screening: Select preset evaluation indicators to evaluate the quality of recycled patterns, and combine professional designers or intangible cultural heritage experts to evaluate the recycled patterns, specifically evaluating the cultural connotation and artistic value of the patterns; then, based on the evaluation results, screen out high-quality recycled patterns.
7. The method for deconstructing and regenerating intangible cultural heritage patterns based on a multi-scale recurrent neural network according to any one of claims 1 to 6, characterized in that: The following steps are involved: Step 1, data preparation stage: through the data acquisition and preprocessing module, a wealth of intangible cultural heritage pattern images are collected, divided into training set, validation set and test set according to a certain ratio, and data enhancement operations are performed on the training set images; Step 2: Model training phase: First, the multi-scale feature extraction module is trained, and the training set images are input into the multi-scale convolutional neural network to extract the multi-scale features of the pattern; The feature vector sequence output by the multi-scale feature extraction module is input into the recurrent neural network module for training, and the parameters of the recurrent neural network are optimized so that it can learn the cyclic structure information of the pattern features; The pattern deconstruction module is jointly trained with the multi-scale feature extraction module and the recurrent neural network module. By comparing the deconstruction results with the annotated pattern structure information, the network parameters are adjusted to enable the pattern deconstruction module to accurately deconstruct the pattern. Finally, the VAE generative model is jointly trained with the previous modules. The patterns generated by the generative model are compared with the real patterns in terms of features and structure. Through variational inference, the parameters of the generative model are optimized so that it can generate high-quality reproduced patterns. Step 3, pattern deconstruction stage: After preprocessing, the intangible cultural heritage pattern image to be deconstructed is sequentially input into the trained multi-scale feature extraction module, recurrent neural network module, and deconstruction module to obtain the pattern deconstruction result, that is, the structured representation of the basic elements and their relationships; Step 4, pattern regeneration stage: According to the design requirements, the model is generated to input specific random noise or style control vectors, and the pattern structure information obtained by the deconstruction module is combined to generate new intangible cultural heritage patterns. The generated patterns are evaluated and screened, and the patterns that meet the requirements are selected as the final regeneration results.