A method for recommending creative materials for art design based on artificial intelligence

Abstract

The present application provides an art design creative material recommendation method based on artificial intelligence, comprising: obtaining materials from different fields, and using a deep learning algorithm to extract features from the materials to obtain feature vectors that reflect the essential attributes of the materials; designing a general rule expression method based on a Turing machine, wherein state nodes are used to characterize the static features of the materials, and state transfer functions are used to characterize the dynamic change features of the materials; for each application field, designing a dedicated rule module and parameter set, and generating calculation rules that conform to the characteristics of the materials in the field by configuring the state nodes and transfer functions; constructing a material calculation rule knowledge base, using a graph database structure to store the calculation rules, rule modules and parameter sets summarized and refined in various fields; designing a rule optimization mechanism based on reinforcement learning, adjusting the dedicated rule modules and parameter sets, taking the comprehensive quality of the generated materials as the optimization goal, and maximizing the field adaptability of the generated materials.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to an art design creative material recommendation method based on artificial intelligence. Background Art

[0002] In the field of art design, creative materials play an important role in guiding the creative thinking of designers. With the development of artificial intelligence, creative materials generated by artificial intelligence technology are increasingly used in art design. In the process of generating creative materials, the use of knowledge transfer technology can realize the recommendation and generation of cross-domain materials, and the key technology is the migration of material calculation rules. In the process of migrating material calculation rules, there is a key technical contradiction: how to design an adaptable and universal rule expression method between different fields. Due to the large differences in the characteristics, properties and internal laws of materials in different fields, directly applying the calculation rules of other fields may cause the generated materials to lack artistry and practicality. For example, the growth rules in biology emphasize organicity and randomness, while graphic design pays more attention to geometry and regularity. When migrating rules, it is necessary to deeply analyze the essential properties of materials in different fields, extract the common laws therein, and design an expression method that is compatible with these common laws; at the same time, it is also necessary to consider the dynamic change characteristics of materials in different fields during the generation process, such as growth rate, morphological changes, etc., and design a rule expression method that can characterize these dynamic characteristics; in addition, when realizing cross-domain rule migration, it is also necessary to balance the contradiction between generality and specificity; on the one hand, the rule expression method must be general enough to cover the characteristics of materials in different fields; on the other hand, special rule modules and parameters must be designed for the unique attributes of each field to ensure the diversity and uniqueness of the generated materials; how to find a balance between generality and specificity is another technical difficulty faced by material calculation rule migration. Summary of the invention

[0003] The present invention provides an art design creative material recommendation method based on artificial intelligence, which mainly includes:

[0004] Materials from different fields are obtained, and deep learning algorithms are used to extract features from the materials to obtain feature vectors that reflect the essential attributes of the materials; cluster analysis is performed on the feature vectors, and the common law representation of materials from different fields is obtained based on the clustering results, which includes the structural characteristics, color distribution, and texture characteristics of the materials; a general rule expression method based on a Turing machine is designed, which includes state nodes for describing the static characteristics of the materials and state transfer functions for describing the dynamic change characteristics of the materials; for each application field, a dedicated rule module and parameter set are designed, and calculation rules that meet the characteristics of the materials in the field are generated by configuring state nodes and transfer functions; a material calculation rule knowledge base is constructed, and a graph database structure is used to store the calculation rules, rule modules, and parameter sets summarized and refined in various fields; a transfer learning algorithm is used to transfer cross-domain rules, and the rule modules and parameters learned in one field are used as the prior knowledge of the target field, and the dedicated rule modules and parameter sets are fine-tuned to adapt to the characteristics of the materials in the target field; a rule optimization mechanism based on reinforcement learning is designed, and the dedicated rule modules and parameter sets are adjusted to optimize the comprehensive quality of the generated materials and maximize the field adaptability of the generated materials.

[0005] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0006] The present invention discloses a method for recommending creative materials for art design based on artificial intelligence. The method first performs feature extraction and cluster analysis on materials in different fields, and summarizes the common regularity representation of the materials. Then a general rule expression method based on a Turing machine is designed, including state nodes and state transfer functions, which are used to characterize the static and dynamic characteristics of the materials. For each application field, the present invention designs a dedicated rule module and parameter set, and constructs a material calculation rule knowledge base. Crucially, the present invention adopts a transfer learning algorithm to transfer cross-domain rules, and uses the rules learned in one field as prior knowledge in other fields, and adapts to the target field through fine-tuning. In addition, the present invention also designs a rule optimization mechanism based on reinforcement learning, with the comprehensive quality of generated materials as the goal, and continuously adjusts and optimizes the rules. This method of cross-domain migration and continuous optimization enables the present invention to efficiently generate materials that meet the characteristics of different fields, greatly improving the versatility and adaptability of material generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 The present invention is a flowchart of a method for recommending creative art design materials based on artificial intelligence.

[0008] Figure 2 A schematic diagram of an art design creative material recommendation method based on artificial intelligence of the present invention.

[0009] Figure 3This is another schematic diagram of an art design creative material recommendation method based on artificial intelligence of the present invention. DETAILED DESCRIPTION

[0010] The following will describe the technical solutions in the embodiments of the present invention in detail in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention.

[0011] like Figure 1-3 In this embodiment, a method for recommending creative art design materials based on artificial intelligence may specifically include:

[0012] Step S101, obtain materials from different fields, and use deep learning algorithms to extract features from the materials to obtain feature vectors that reflect the essential attributes of the materials.

[0013] A1. Obtain material data from different fields, distinguish and label the fields according to the source of the materials; A2. Preprocess the material data, including data cleaning and normalization operations; A3. Use a pre-trained deep learning model to extract features from the material data to obtain a high-dimensional feature vector that reflects the essential attributes of the material; A4. Cluster the extracted high-dimensional feature vectors using the K-means clustering algorithm to automatically discover the commonalities and differences between different materials; A5. Based on the clustering results, build a field feature library, which is used to store typical feature vectors of each field; A 6. When new material data is obtained, repeat steps A1-A3 to perform preprocessing operations on it and extract its feature vector; A7. Perform cosine similarity calculation on the feature vector of the new material and the typical vector in the field feature library to determine the field to which the new material belongs; A8. After determining the field to which the new material belongs, perform fine-grained attribute comparison on it with the materials in the field, and obtain the detailed attribute label of the new material through the attribute label matching algorithm; A9. Associate the field attributes and detailed attributes of the new material to form structured material attribute information, and store it in the material attribute database.

[0014] Specifically, first, a large amount of material data is obtained from different fields such as news, social media, e-commerce and other channels, and annotated according to the source. Then, the data is cleaned to remove invalid, repeated and abnormal data, and the data in different formats such as text and images are normalized and converted into a unified vector representation. Next, pre-trained deep learning models such as BERT and ResNet are used to extract features from material data in different fields to obtain a high-dimensional feature vector with a dimension of generally 128 to 1024. Subsequently, the K-means clustering algorithm is used to set a suitable K value (such as 100) to cluster the extracted feature vectors. By analyzing the center point features of each cluster, the commonalities and differences of different materials in terms of content, style, etc. can be found. Based on the clustering results, 3 to 5 representative typical feature vectors are selected for each field to construct a field feature library. When new material data arrives, it is also pre-processed and feature extracted to obtain its feature vector. Then, the cosine similarity between the vector and each typical vector in the field feature library is calculated, and the field corresponding to the maximum value is taken as the field label of the material. After determining the domain, the feature vector of the material is further compared with the materials in the corresponding domain in a fine-grained manner. By designing an attribute label matching algorithm, the similarity between materials is calculated from multiple dimensions such as semantics, emotions, and themes, and the most similar Top-N materials are extracted to inherit their attribute labels. Finally, the domain attributes and detailed attributes of the new material are associated, and structured material attribute information is generated according to the predefined data model and stored in the material attribute database. Through this series of processes, the automated processing, domain division, and attribute extraction of massive heterogeneous material data can be achieved, and a large-scale, high-quality material attribute knowledge base can be built, laying the foundation for subsequent content production, recommendation, and other applications.

[0015] Step S102, cluster analysis is performed on the feature vectors, and a common law representation of materials in different fields is obtained based on the clustering results. The representation includes multi-dimensional information such as structural features, color distribution, and texture features of the materials.

[0016] Acquire a material set, the material set including a plurality of materials to be classified; perform clustering processing on the material set using a K-means clustering algorithm to obtain a plurality of material clusters; extract the structural features, color distribution and texture features of the material cluster for each material cluster, and generate corresponding material cluster features; construct a material feature template library based on the material cluster features; acquire materials to be classified, extract the structural features, color distribution and texture features of the materials to be classified, and generate material features to be classified; classify the features of the materials to be classified using a support vector machine algorithm to obtain the classification features of the materials to be classified; Domain category; during the classification, the material clustering features in the material feature template library are used as training samples of the support vector machine algorithm; if the confidence of the domain category of the material to be classified is lower than a preset threshold, the similarity between the material feature to be classified and each material clustering feature in the material feature template library is calculated, and the domain category corresponding to the material clustering feature with the highest similarity is used as the domain category of the material to be classified; according to the classification result, the material feature template library is updated; during the update, the new material clustering features are added to the material feature template library, and the redundant features and noise features in the material feature template library are removed.

[0017] Specifically, firstly, a large amount of material data is obtained from the Internet, image libraries and other channels to build a material collection containing more than 100,000 images. Then, the material collection is preprocessed to unify the image size to 256*256 pixels, and the image data is normalized to the range of 0-1. Then, the K-means clustering algorithm is used to cluster the material collection, and the number of clusters K is set to 100, the number of iterations is 50, and 100 material clusters are obtained. For each material cluster, its structural features, color distribution and texture features are extracted. The structural features are obtained by extracting the SIFT features of the image and quantizing and encoding them using the bag of words model to obtain a 128-dimensional feature vector; the color distribution is obtained by calculating the histogram of the image in the HSV color space to obtain a 256-dimensional feature vector; the texture features are obtained by calculating the local binary pattern histogram of the image to obtain a 59-dimensional feature vector. These three features are cascaded to generate a 443-dimensional material cluster feature vector to build a material feature template library. When new materials to be classified arrive, their structural features, color distribution and texture features are also extracted to generate features of the materials to be classified. Then the support vector machine algorithm is used to classify the features of the materials to be classified, and the radial basis kernel function is adopted. The regularization parameter C is set to 10, and the 100 cluster features in the material feature template library are used as training samples to train the classification model. The features of the materials to be classified are predicted to obtain their domain categories and confidence levels. If the confidence level is lower than the threshold of 6, the classification result is considered unreliable, and the Euclidean distance between the features of the materials to be classified and the cluster features in the material feature template library is calculated. The domain category corresponding to the cluster feature with the closest distance is taken as the domain category of the materials to be classified. Finally, the material feature template library is updated according to the classification results, the features of the materials to be classified are added to the corresponding clusters, and each cluster is refined using the density clustering algorithm to remove the redundant features and noise features with low density, so as to ensure the refinement and efficiency of the material feature template library. After multiple rounds of iteration and optimization, the material feature template library will be able to accurately and efficiently classify massive materials automatically.

[0018] Step S103, designing a general rule expression method based on a Turing machine, which includes a state node for describing the static characteristics of the material and a state transfer function for describing the dynamic change characteristics of the material.

[0019] The video is decoded through the OpenCV library, the video frame is extracted, and the color histogram and SIFT feature points of the material are extracted from it, and the color histogram and SIFT feature points are processed to obtain feature parameters; according to the feature parameters, it is matched with the predefined state nodes to determine the current state of the material; if the state of the material changes, the state transfer is triggered according to the preset transfer function; the transfer function determines the next state according to the current state and input conditions, and this process is implemented through the state machine engine SMC; at the same time, the Redis memory database is used to cache the state information of the material; through this method, accurate description of material features and real-time status monitoring are achieved.

[0020] Specifically, first decode the video through the OpenCV library, extract the video frame, and scale the frame image to a uniform size, such as 640*480 pixels. Then preprocess each frame of the image, including grayscale, histogram equalization and Gaussian filtering, to eliminate noise and lighting effects. Then extract the color histogram and SIFT feature points of the image. The color histogram uses the HSV color space, calculates the histogram of 32 bins for the three channels of H, S, and V, and normalizes them into feature vectors. SIFT feature point extraction uses the DoG scale space and 128-dimensional descriptor, obtains 100 visual words through clustering, and uses the bag-of-words model to generate feature vectors for each frame of the image. The color histogram and SIFT feature vector are spliced ​​to obtain a 164-dimensional feature parameter. According to the feature parameter, the Euclidean distance is matched with the predefined state node, and the distance threshold is set to 8 to determine the current state of the material. If the current state is different from the state of the previous frame, the state transition is triggered. The transfer function searches the state transition table according to the current state and input conditions, determines the next state, and records the state transition time and duration. The state machine engine is implemented using the SMC library, and the state diagram is defined in JSON format. The state transition table is stored in the Redis database for fast query and update. At the same time, the state information of the material is cached in Redis in the form of key-value pairs, with the key being the material ID and the value being the state ID and timestamp. In this way, automatic state recognition and real-time monitoring of video materials are achieved, which can accurately describe the dynamic characteristics of the material and provide support for subsequent material classification, retrieval, and editing.

[0021] Step S104, design a dedicated rule module and parameter set for each application field, and generate calculation rules that meet the characteristics of the materials in this field by configuring state nodes and transfer functions.

[0022] According to the characteristics and requirements of the application field, a suitable set of parameters is designed to characterize the key features and constraints of the field. Through the analysis of business scenarios, the core business objects and their attributes are identified and abstracted into state nodes. According to the dependencies and conversion rules between business objects, the transfer functions between state nodes are constructed. For each state node, starting from the business requirements, the corresponding calculation rules are designed to process the data and logic in the state. The design of the calculation rules should fully consider the business constraints and special circumstances to ensure that the generated results meet the actual needs. Through data collection and analysis, the key data in the business process is obtained as the input of the calculation rules. Using the constructed state nodes, transfer functions and calculation rules, the business data is processed and calculated to generate result data that conforms to the business logic. The calculation results are compared with the expected output to evaluate the accuracy and effect of the rule module. According to the evaluation results, the state nodes, transfer functions and calculation rules are optimized and adjusted to improve the adaptability and accuracy of the system. Through log monitoring and data analysis, the changes of business data are continuously tracked to timely discover the factors that may affect the rule module. According to the results of monitoring and analysis, the parameter set and calculation rules are dynamically adjusted to keep them consistent with business needs. After multiple rounds of iterative optimization, a stable, efficient and accurate business rule engine was finally formed to provide intelligent support for the operation of the business system.

[0023] Specifically, in the video material management system, a set of 164-dimensional feature parameters is first defined, of which the color histogram accounts for 96 dimensions and the SIFT feature accounts for 100 dimensions, which respectively represent the visual features of the material such as color, texture and key points. According to the content and attributes of the video material, five state nodes are designed: static, motion, close-up, panoramic and transition, which correspond to the typical states of the video material in different shooting scenes. The static state indicates that the picture is basically motionless, the motion state indicates that there is an obvious moving target in the picture, the close-up state indicates that the picture is mainly a close-up, the panoramic state indicates that the picture is a panoramic or distant view, and the transition state indicates that the picture is undergoing a transition effect. According to the general rules of video shooting, the transfer function between state nodes is defined. For example, the static state can be transferred to the motion state or close-up state, the motion state can be transferred to the static state or panoramic state, the close-up state can be transferred to the static state or transition state, the panoramic state can be transferred to the motion state or transition state, and the transition state can be transferred to any other state. The triggering condition of the transfer function is determined according to the current state and the input feature parameters. For example, when the current state is static and the area of ​​the moving target accounts for more than 20%, the transfer to the motion state is triggered. For each state node, a corresponding calculation rule is designed. For example, in the static state, the average brightness and contrast of the picture are counted. If the brightness is less than 30 or the contrast is less than 20, it is marked as insufficient lighting. In the moving state, the trajectory and speed of the moving target are counted. If the speed is greater than 50 pixels / second, it is marked as fast motion. In the close-up state, the face and limbs are detected. If the face is detected, the facial expression and sight direction are further identified. In the panoramic state, the composition and subject of the picture are analyzed. If the subject is located in the center of the picture, it is marked as a centered composition. In the transition state, the type and duration of the transition are identified. If the duration is greater than 1 second, it is marked as a slow transition. Through the analysis and statistics of a large number of video materials, a batch of labeled data is obtained as the training set of the calculation rules. The SVM algorithm is used to classify the feature parameters in each state, and a set of discriminants are obtained. The discriminant is embedded in the calculation rule to automatically label and classify the new video material. By comparing with the manual labeling results, it is found that the accuracy of the rule has reached more than 85%. For cases of misclassification, the feature parameters and discriminants were further optimized, and the processing logic for some special scenes, such as motion blur, occlusion, and light and shadow changes, was added. At the same time, the state distribution and transition frequency of video materials were analyzed through log data, and it was found that different types of videos had significant differences in state distribution. For example, news videos have more static and close-up states, while entertainment videos have more motion and transition states. Taking advantage of this feature, the parameters of the state transition function were adaptively adjusted, and different transition thresholds and durations were set according to different video types, which improved the accuracy of state recognition.Finally, a Turing machine-based automatic video material annotation and classification system was established, which can automatically generate status labels and keywords based on the video content, providing a powerful support tool for video retrieval and editing.

[0024] Step S105, constructing a material calculation rule knowledge base, using a graph database structure to store the calculation rules, rule modules and parameter sets summarized and refined in various fields.

[0025] Obtain material calculation rules, rule modules and parameter set knowledge in various fields, preliminarily screen and organize the parameter set knowledge, eliminate redundant and irrelevant information, and retain key information; structure the screened parameter set knowledge, extract key information, and build a rule knowledge base; the key information includes rule conditions, rule results and parameter value ranges; use Neo4j graph database to store the rule knowledge base; use rule conditions and parameter value ranges as nodes, and rule results as the relationship between nodes to form a rule knowledge graph; obtain the rule subset applicable to the target material by matching the attributes of the target material with the rule conditions; use the Drools rule engine to perform intelligent calculation processing on the target material based on the retrieved rule subset and parameters to generate multiple variant elements material; perform quality assessment on the generated variant material, the assessment includes using the material similarity algorithm SSIM to assess the diversity of the material, and using the novelty algorithm ImageRank to score the overall quality, diversity, and innovation of the material, respectively, to obtain a scoring result; according to the scoring result, dynamically optimize and adjust the rules and parameters in the rule knowledge graph, the dynamic optimization and adjustment includes adjusting the parameter value range of the generation rule for the material with an overall quality score lower than a preset threshold, adding new generation rules for the material with a diversity score lower than a preset threshold, and optimizing and combining existing rules to form new generation rules for the material with an innovation score lower than a preset threshold; continuously iterate and optimize the knowledge graph to improve the effect and quality of intelligent material generation; output an optimized high-quality and diverse material collection.

[0026] Specifically, in the intelligent material generation system, a large number of material calculation rules and parameter sets are first collected from multiple fields. For example, for video materials, rules and parameters in terms of lens movement, color style, and picture composition are collected; for audio materials, rules and parameters in terms of rhythm, melody, and timbre are collected. Then, the collected rules and parameters are screened and sorted, some overly simple or repeated rules are eliminated, the most critical and representative rules are retained, and key information such as rule conditions, rule results, and parameter value ranges are extracted to build a structured rule knowledge base. The knowledge base uses the Neo4j graph database for storage, with rule conditions and parameter value ranges as nodes and rule results as the relationship between nodes, forming a rule knowledge graph. When a target material needs to be intelligently calculated, the system first analyzes the attribute characteristics of the target material, and then retrieves the rules and parameters related to it in the knowledge graph. For example, for a video material, the system detects that its lens is mainly slow panning and the color saturation is low, then it will retrieve rules related to slow lens movement and low-saturation color style. By matching material attributes and rule conditions, the system finds a set of rule subsets that apply to the material. Next, the system uses the Drools rule engine to perform intelligent calculations on the material based on the rule subsets and corresponding parameters to generate multiple variant materials. For example, for previous video materials, the system may generate a variant that accelerates the movement of the lens and increases the color saturation, and may also generate a variant that reduces the speed of the lens movement and increases the sense of particle size in the picture. For the generated variant materials, the system uses the SSIM similarity algorithm to evaluate their diversity, the ImageRank novelty algorithm to evaluate their innovation, and manually scores the overall quality of the materials. The scoring results show that the overall quality of the generated variant materials reached 2 points (out of 5 points), the diversity index reached 78 (the closer to 100, the better), and the innovation index reached 69 (the closer to 100, the better). Based on the scoring results, the system dynamically optimizes the rules and parameters in the knowledge graph. For example, for materials with low quality scores, the system adjusts the parameter value range of its generation rules and lowers the upper limit of the lens movement speed from 30 pixels / second to 20 pixels / second; for materials with low diversity index, the system adds some new generation rules, such as screen segmentation and special effects superposition; for materials with low innovation index, the system tries to combine existing rules to form new generation rules, such as combining slow lens movement with high-contrast color style to form a new "slow motion + high contrast" style. Through continuous iteration and optimization of the knowledge graph, the system significantly improves the effect and quality of intelligent material generation, and ultimately outputs a high-quality and diverse material collection, providing rich choices for material application.

[0027] Step S106, adopting the transfer learning algorithm to transfer cross-domain rules, taking the rule modules and parameters learned in one domain as the prior knowledge of the target domain, and adapting to the characteristics of the target domain material by fine-tuning the dedicated rule modules and parameter sets.

[0028] Acquire a source domain data set, clean and standardize the source domain data set to obtain a preprocessed source domain data set; use principal component analysis to extract a discriminative source domain feature vector for the preprocessed source domain data set; input the extracted feature vector into a neural network model, and obtain a source domain prediction model through training; acquire a target domain data set, clean and standardize the target domain data set to obtain a preprocessed target domain data set, and then use principal component analysis to extract a discriminative target domain feature vector for the preprocessed target domain data set; input the extracted target domain feature vector into a neural network model, and obtain a target domain prediction model through training; the source domain prediction model The parameters are used as the initialization parameters of the target domain prediction model to initialize the target domain prediction model; the target domain feature vector is input into the target domain prediction model, and the target domain prediction model is trained using the back propagation algorithm to obtain the trained target domain prediction model; during the training process, if the performance index of the target domain prediction model on the validation set reaches a preset threshold, the training is terminated to obtain the final target domain prediction model; if the performance index of the target domain prediction model on the validation set does not reach the preset threshold, the training is continued until the preset threshold is reached; the target domain test set is input into the final target domain prediction model, the prediction result of the target domain is output, and the knowledge transfer from the source domain to the target domain is completed.

[0029] Specifically, when cleaning and standardizing the source domain data set, first extract the original data from the database through SQL query statements, and then use the Pandas library in Python to remove duplicate values ​​and process missing values, such as filling missing numerical data with the average value of the column. Next, the data is Z-score standardized, that is, the mean of the data is adjusted to 0 and the standard deviation is adjusted to 1 to eliminate the dimensional influence between different features. In the feature extraction stage, the principal component analysis (PCA) algorithm is used. Through the PCA module in the Scikit-learn library, the principal components with a cumulative contribution rate of more than 85% of the explained variance are selected, thereby effectively reducing the data dimension while retaining the most important information. The obtained feature vector is used to train the neural network model. Here, the TensorFlow framework is used to build a multi-layer perceptron (MLP) and initialize the weights and biases. In the target domain model training stage, the parameters of the source domain model are used as the initial parameters, and the Adam optimizer is used to update the parameters. The learning rate is set to 0.01, the batch size is 32, and the model is optimized by calculating the cross entropy loss function. If the accuracy of the model does not improve within five consecutive validation cycles, the early stopping method is used to terminate the training early to prevent overfitting. Finally, the test set data of the target domain is input into the final prediction model, and the data category is judged by the probability value output by the model, completing the knowledge transfer process from the source domain to the target domain.

[0030] Step S107, design a rule optimization mechanism based on reinforcement learning, adjust the dedicated rule module and parameter set, take the comprehensive quality of the generated material as the optimization goal, and maximize the domain adaptability of the generated material.

[0031] Acquire a pre-established initial rule module and parameter set based on specific domain requirements, wherein the initial rule module is based on pre-established domain knowledge, and the parameter set is optimized and configured according to the characteristics of the target domain;

[0032] B1. Generate at least one batch of material samples according to the initial rule module and parameter set; B2. Obtain a preset evaluation standard, use the preset evaluation standard to evaluate the quality of the at least one batch of material samples, and obtain a quality score corresponding to each of the material samples, wherein the preset evaluation standard includes the integrity, innovation and applicability of the sample; B3. Use the quality score corresponding to each material sample as a reward signal, use the policy gradient algorithm to adjust the initial rule module and parameter set according to the reward signal, and obtain an adjusted rule module and parameter set; B4. Determine whether the adjusted rule module and parameter set improves the quality of the material. If so, update the adjusted rule module and parameter set to a new initial rule module and parameter set. If not, keep the original initial rule module and parameter set unchanged; B5. Repeat steps B1-B4 until the material quality meets the preset optimization target or reaches the maximum number of iterations; B6. Output the optimized rule module and parameter set.

[0033] Specifically, in an intelligent system based on the needs of a specific domain, an initial set of rule modules and parameters is first established, which are optimized and configured based on historical data. For example, if the target domain is image recognition, the initial parameters may include the weights of the convolutional neural network (CNN), which are obtained by pre-training on a large number of annotated images. Then, a batch of image samples are generated using these initial rules and parameters, which enhance the generalization ability of the model by simulating changes in real-world scenes (such as lighting, angle, etc.). After that, the quality of the generated samples is evaluated using preset evaluation criteria, such as sample clarity, edge recognition accuracy, and object classification accuracy, and each sample is given a comprehensive quality score. These scores are then used as reward signals to adjust the initial rule modules and parameter sets through a policy gradient algorithm. Specifically, the policy gradient method in reinforcement learning, such as the REINFORCE algorithm, is used to update the network parameters based on the reward signal to improve the overall quality of the generated samples. The system will continuously monitor whether the adjusted rule modules and parameter sets actually improve the sample quality. If there is no improvement, the original settings will be maintained; if there is an improvement, it will be updated to a new initial rule module and parameter set. This process will be repeated until the sample quality reaches the preset optimization target or the maximum number of iterations is reached, and finally the optimized rule module and parameter set are output for use in practical applications. This method ensures that the system gradually improves its performance in continuous self-optimization and adapts to complex and changing application scenarios.

[0034] The above only lists some preferred embodiments of the present invention, but the present invention is not limited thereto, and many improvements and changes can be made. As long as the improvements and changes are made on the basis of the basic principles of the present invention, they should be regarded as falling within the protection scope of the present invention.

Claims

1. A method for recommending creative art design materials based on artificial intelligence, characterized in that: The method comprises: Materials from different fields are obtained, and deep learning algorithms are used to extract features from the materials to obtain feature vectors that reflect the essential attributes of the materials; cluster analysis is performed on the feature vectors, and the common law representation of materials from different fields is obtained based on the clustering results, which includes the structural characteristics, color distribution, and texture characteristics of the materials; a general rule expression method based on a Turing machine is designed, which includes state nodes for describing the static characteristics of the materials and state transfer functions for describing the dynamic change characteristics of the materials; for each application field, a dedicated rule module and parameter set are designed, and calculation rules that meet the characteristics of the materials in the field are generated by configuring state nodes and transfer functions; a material calculation rule knowledge base is constructed, and a graph database structure is used to store the calculation rules, rule modules, and parameter sets summarized and refined in various fields; a transfer learning algorithm is used to transfer cross-domain rules, and the rule modules and parameters learned in one field are used as the prior knowledge of the target field, and the dedicated rule modules and parameter sets are fine-tuned to adapt to the characteristics of the materials in the target field; a rule optimization mechanism based on reinforcement learning is designed, and the dedicated rule modules and parameter sets are adjusted to optimize the comprehensive quality of the generated materials and maximize the field adaptability of the generated materials.

2. The method according to claim 1, characterized in that Obtain materials from different fields and use deep learning algorithms to extract features from the materials to obtain feature vectors that reflect the essential attributes of the materials, including: A1. Obtain material data from different fields, and distinguish and label them according to the source of the material; A2. Preprocessing the material data, including data cleaning and normalization operations; A3. For the material data, a pre-trained deep learning model is used to perform feature extraction to obtain a high-dimensional feature vector that reflects the essential attributes of the material; A4. Clustering the extracted high-dimensional feature vectors by using a K-means clustering algorithm to automatically discover the commonalities and differences between different materials; A5. According to the clustering results, a domain feature library is constructed, where the domain feature library is used to store typical feature vectors of each domain. A6. When new material data is obtained, repeat steps A1-A3 to perform preprocessing operations on it and extract its feature vector; A7, calculating the cosine similarity between the feature vector of the new material and the typical vector in the domain feature library to determine the domain to which the new material belongs; A8. After determining the field to which the new material belongs, perform fine-grained attribute comparison with the materials in that field, and obtain detailed attribute labels of the new material through attribute label matching algorithm; A9. Associate the domain attributes and detailed attributes of the new material to form structured material attribute information, and store it in the material attribute database.

3. The method according to claim 1, characterized in that Perform cluster analysis on the feature vectors, and obtain the common law representation of materials in different fields according to the clustering results. The representation includes the structural characteristics, color distribution, and texture characteristics of the materials, including: Acquire a material set, wherein the material set includes a plurality of materials to be classified; For the material set, a K-means clustering algorithm is used to perform clustering processing to obtain multiple material clusters; For each of the material clusters, extract the structural features, color distribution and texture features of the material cluster to generate corresponding material cluster features; Constructing a material feature template library according to the material clustering features; Acquire a material to be classified, extract structural features, color distribution and texture features of the material to be classified, and generate features of the material to be classified; Using a support vector machine algorithm to classify the features of the material to be classified, and obtaining the field category of the material to be classified; During classification, the material clustering features in the material feature template library are used as training samples for the support vector machine algorithm; If the confidence of the domain category of the material to be classified is lower than a preset threshold, the similarity between the feature of the material to be classified and the features of each material cluster in the material feature template library is calculated, and the domain category corresponding to the material cluster feature with the highest similarity is used as the domain category of the material to be classified; According to the classification result, updating the material feature template library; During updating, new material clustering features are added to the material feature template library, and redundant features and noise features in the material feature template library are removed.

4. The method according to claim 1, characterized in that: Design a general rule expression based on Turing machine, which includes state nodes for describing the static characteristics of materials and state transfer functions for describing the dynamic characteristics of materials, including: Decode the video through the OpenCV library, extract the video frame, and extract the color histogram and SIFT feature points of the material from it, and process the color histogram and SIFT feature points to obtain feature parameters; According to the characteristic parameters, match with the predefined state nodes to determine the current state of the material; If the state of the material changes, the state transfer is triggered according to the preset transfer function; The transfer function determines the next state according to the current state and input conditions, and this process is implemented by the state machine engine SMC; At the same time, the Redis memory database is used to cache the status information of the material.

5. The method according to claim 1, characterized in that For each application field, we design dedicated rule modules and parameter sets, and generate calculation rules that meet the characteristics of the materials in that field by configuring state nodes and transfer functions, including: According to the characteristics and requirements of the application field, a set of parameters is designed to characterize the key features and constraints of the field; By analyzing the business scenarios, we identify the core business objects and their attributes and abstract them into state nodes. Construct the transfer function between state nodes according to the dependency relationship and conversion rules between business objects; For each state node, design corresponding calculation rules based on business needs to process the data and logic in that state; The design of calculation rules should fully consider business constraints and special circumstances to ensure that the generated results meet actual needs; Through data collection and analysis, key data in the business process is obtained as input for calculation rules; Use the constructed state nodes, transfer functions and calculation rules to process and calculate business data to generate result data that conforms to business logic; Compare the calculation results with the expected output to evaluate the accuracy and effectiveness of the rule module; According to the evaluation results, the state nodes, transfer functions and calculation rules are optimized and adjusted to improve the adaptability and accuracy of the system.

6. The method according to claim 1, characterized in that Construct a material calculation rule knowledge base, using a graph database structure to store the calculation rules, rule modules and parameter sets summarized and refined in various fields, including: Obtain material calculation rules, rule modules and parameter set knowledge in various fields, preliminarily screen and organize the parameter set knowledge, eliminate redundant and irrelevant information, and retain key information; Structuring the filtered parameter set knowledge, extracting key information, and building a rule knowledge base; the key information includes rule conditions, rule results, and parameter value ranges; The Neo4j graph database is used to store the rule knowledge base; rule conditions and parameter value ranges are used as nodes, and rule results are used as relationships between nodes to form a rule knowledge graph; For the target material, retrieve the calculation rules and parameters related to it from the rule knowledge graph; Acquire a rule subset applicable to the target material by matching the attributes of the target material with the rule conditions; According to the retrieved rule subset and parameters, the Drools rule engine is used to perform intelligent calculation processing on the target material to generate multiple variant materials; Performing a quality assessment on the generated variant materials, the assessment includes using a material similarity algorithm SSIM to assess the diversity of the materials, and using a novelty algorithm ImageRank to score the overall quality, diversity, and innovation of the materials, respectively, to obtain a scoring result; According to the scoring results, the rules and parameters in the rule knowledge graph are dynamically optimized and adjusted, wherein the dynamic optimization and adjustment includes adjusting the parameter value range of the generation rules for materials whose overall quality score is lower than a preset threshold, adding new generation rules for materials whose diversity score is lower than a preset threshold, and optimizing and combining existing rules to form new generation rules for materials whose innovation score is lower than a preset threshold; Continuously iterate and optimize the knowledge graph to improve the effect and quality of intelligent material generation; Output optimized high-quality and diverse collections of assets.

7. The method according to claim 1, characterized in that The transfer learning algorithm is used to transfer rules across domains. The rule modules and parameters learned in one domain are used as the prior knowledge of the target domain. By fine-tuning the dedicated rule modules and parameter sets, the characteristics of the target domain materials are adapted, including: Acquire a source domain data set, and perform cleaning and standardization processing on the source domain data set to obtain a preprocessed source domain data set; For the preprocessed source domain data set, principal component analysis is used to extract a distinguishing source domain feature vector; The extracted feature vectors are input into the neural network model, and the source domain prediction model is obtained through training; Acquire a target domain data set, clean and standardize the target domain data set to obtain a preprocessed target domain data set, and then use principal component analysis to extract a discriminative target domain feature vector for the preprocessed target domain data set; Inputting the extracted target domain feature vector into a neural network model, and obtaining a target domain prediction model through training; Using the parameters of the source domain prediction model as initialization parameters of the target domain prediction model to initialize the target domain prediction model; Inputting the target domain feature vector into the target domain prediction model, and training the target domain prediction model using a back propagation algorithm to obtain a trained target domain prediction model; During the training process, if the performance index of the target domain prediction model on the validation set reaches a preset threshold, the training is terminated to obtain the final target domain prediction model; If the performance index of the target domain prediction model on the validation set does not reach the preset threshold, continue training until it reaches the preset threshold; The target domain test set is input into the final target domain prediction model, and the prediction result of the target domain is output to complete the knowledge transfer from the source domain to the target domain.

8. The method according to claim 1, characterized in that Design a rule optimization mechanism based on reinforcement learning, adjust the dedicated rule modules and parameter sets, take the comprehensive quality of generated materials as the optimization goal, and maximize the domain adaptability of generated materials, including: Acquire a pre-established initial rule module and parameter set based on specific domain requirements, wherein the initial rule module is based on pre-established domain knowledge, and the parameter set is optimized and configured according to the characteristics of the target domain; B1. Generate at least one batch of material samples according to the initial rule module and parameter set; B2. Obtaining a preset evaluation standard, and using the preset evaluation standard to perform quality evaluation on the at least one batch of material samples to obtain a quality score corresponding to each of the material samples, wherein the preset evaluation standard includes the integrity, innovation and applicability of the sample; B3, taking the quality score corresponding to each material sample as a reward signal, and using a policy gradient algorithm to adjust the initial rule module and parameter set according to the reward signal to obtain an adjusted rule module and parameter set; B4. Determine whether the adjusted rule module and parameter set improve the quality of the material. If so, update the adjusted rule module and parameter set to a new initial rule module and parameter set. If not, keep the original initial rule module and parameter set unchanged. B5. Repeat steps B1-B4 until the material quality meets the preset optimization target or the maximum number of iterations is reached; B6. Output the optimized rule module and parameter set.

Images