Multi-modal personalized costume design method and system

Abstract

The invention provides a multi-modal personalized costume design method and system, and relates to the field of costume design and information technology crossing, and the method comprises the steps: collecting multi-modal data of a user; key features of multi-modal data are obtained by extracting key features of each type of modal data, a multi-layer perceptron is constructed, the multi-modal data are analyzed, and user personalized feature vectors are generated; constructing a decision tree according to the personalized feature vector of the user, generating a costume design rule base according to the decision tree, and calculating the similarity between the to-be-designed costume and the costume in the training set by applying a KNN algorithm; performing personalized design through the similarity to generate a multi-modal personalized costume design scheme; by constructing a technical closed loop of'multi-modal perception-deep modeling-knowledge reasoning-similar evolution ', not only is the pain point that traditional costume design depends on subjective experience solved, but also a new-generation explainable and evolvable personalized design normal form is created through fusion of data intelligence and domain knowledge.

Description

Technical Field

[0001] The present invention relates to the cross - technical field of clothing design and information technology, and particularly to a multi - modal personalized clothing design method and system. Background Art

[0002] With the increasing growth of consumers' personalized needs for clothing, traditional clothing design methods have gradually exposed pain points such as relying on subjective experience, low design efficiency, and difficulty in meeting diverse needs. Traditional clothing design often relies on designers' personal experience and aesthetics, making it difficult to comprehensively and accurately capture market dynamics and consumer preferences. At the same time, with the rapid development of digital and intelligent technologies, the clothing design field is also facing an urgent need to transform towards intelligence and personalization. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a multi - modal personalized clothing design method and system. By constructing a technical closed - loop of "multi - modal perception - deep modeling - knowledge reasoning - similarity evolution", the present invention not only solves the pain point of traditional clothing design relying on subjective experience, but also creates a new generation of interpretable and evolvable personalized design paradigm through the integration of data intelligence and domain knowledge.

[0004] To solve the above - mentioned technical problem, the technical solution of the present invention is as follows: In a first aspect, a multi - modal personalized clothing design method, the method includes: Collect multi - modal data of users; By extracting the key features of each type of modal data, obtain the key features of the multi - modal data; According to the key features of the multi - modal data, construct a multi - layer perceptron; According to the multi - layer perceptron, analyze the multi - modal data to generate a user - personalized feature vector; According to the user - personalized feature vector, construct a decision tree, and according to the decision tree, generate a clothing design rule base; According to the clothing design rule base, apply the KNN algorithm to calculate the similarity between the clothing to be designed and the clothing in the training set; Through the similarity, perform personalized design to generate a multi - modal personalized clothing design scheme.

[0005] Further, by extracting the key features of each type of modal data to obtain the key features of the multi - modal data, it includes: Obtain multi - modal data, where the multi - modal data includes visual modal data, text modal data, physiological modal data, and behavioral modal data; Pre - process the multi - modal data to obtain pre - processed multi - modal data; Extract features from each type of preprocessed modal data to obtain the corresponding single-modal feature vectors; Assign a weight to the feature vector of each modality, through Fuse the single-modal feature vectors into a comprehensive multi-modal feature vector to obtain the key features of the multi-modal data, where 、 、 、 respectively represent the visually modal data, text modal data, physiological modal data, and behavioral modal data obtained originally; 、 、 、 respectively represent the feature vectors of the visual modality, text modality, physiological modality, and behavioral modality; 、 、 、 respectively represent the weights of the feature vectors of the visual modality, text modality, physiological modality, and behavioral modality; represents the comprehensive multi-modal feature vector.

[0006] Furthermore, based on the key features of the multi-modal data, construct a multi-layer perceptron, including: Based on the key features of the multi-modal data, construct the architecture of the multi-layer perceptron model, which includes an input layer, a hidden layer, the dimensions of the output layer, and an activation function; Train and optimize the architecture of the multi-layer perceptron model to obtain the optimized architecture of the multi-layer perceptron model.

[0007] Furthermore, based on the multi-layer perceptron, analyze the multi-modal data to generate a user personalized feature vector, including: Process the multi-modal data according to the multi-layer perceptron model to obtain the predicted value of the personalized feature vector; Evaluate the prediction result of the personalized feature vector to obtain the evaluation result; According to the evaluation result, correct the predicted value of the personalized feature vector to generate the user personalized feature vector.

[0008] Furthermore, based on the user personalized feature vector, construct a decision tree, and according to the decision tree, generate a clothing design rule base, including: According to the user personalized feature vector, divide the data, construct the nodes and branches of the decision tree, and set the parameters of the decision tree; Prune the decision tree to obtain the processed decision tree; Starting from the root node of the decision tree, traverse each node and branch. Each node represents a judgment condition for a feature dimension, and the branch represents the decision path corresponding to different feature values; According to the judgment conditions and branch directions of the nodes, generate clothing design rules, organize and classify all the generated rules to form a clothing design rule library.

[0009] Furthermore, according to the clothing design rule library, apply the KNN algorithm to calculate the similarity between the clothing to be designed and the clothing in the training set, including: Extract the clothing features of the training set according to the clothing design rule library; Determine the value of K according to the extracted clothing features of the training set; According to the value of K, through Calculate the distances between the clothing features vectors of the training set clothing and the clothing to be designed to obtain all the distance values, where, represents the th clothing feature vector of the training set clothing and the clothing feature vector of the clothing to be designed the Euclidean distance between them, is the th clothing feature vector of the training set clothing dimensional eigenvalue, is the dimensional eigenvalue of the clothing feature vector of the clothing to be designed; Sort all the distance values, and determine the similarity between the clothing to be designed and the clothing in the training set according to the sorting result.

[0010] Furthermore, through the similarity, perform personalized design to generate a multi-modal personalized clothing design plan, including: Screen out clothing samples that match the similarity of the clothing to be designed through the similarity; Extract the common design elements by analyzing the screened clothing samples with similarity; According to the common design elements, through adjustment and optimization, combine them together to generate a multi-modal personalized clothing design plan.

[0011] On the second aspect, a multi-modal personalized clothing design system includes: An acquisition module, used to collect multi-modal data of users; by extracting the key features of each modal data to obtain the key features of the multi-modal data; A construction module, used to construct a multi-layer perceptron according to the key features of the multi-modal data; according to the multi-layer perceptron, analyze the multi-modal data through the Nadam optimizer to generate a user personalized feature vector; according to the user personalized feature vector, construct a decision tree, and generate a clothing design rule library according to the decision tree; A processing module, configured to calculate the similarity between the clothing to be designed and the clothing in the training set according to the clothing design rule base and by applying the KNN algorithm; and perform personalized design based on the similarity to generate a multi-modal personalized clothing design solution.

[0012] In a third aspect, a computing device includes: One or more processors; A storage device for storing one or more programs, which when executed by the one or more processors cause the one or more processors to implement the method.

[0013] In a fourth aspect, a computer-readable storage medium stores a program that, when executed by a processor, implements the method.

[0014] The above solution of the present invention has at least the following beneficial effects: By collecting the multi-modal data of users, comprehensively understanding the user needs, generating more accurate personalized feature vectors, and making the designed clothing more in line with user expectations; using a multi-layer perceptron (MLP) to automatically learn the complex relationships between multi-modal features, improving the accuracy of personalized feature extraction, and designing more innovatively and uniquely; constructing a technical closed-loop to realize the full-process automation from data collection to design, reducing manual intervention and improving design efficiency; applying the KNN algorithm to quickly match similar design cases, learning from successful experiences, and further improving design efficiency and quality; using a decision tree to construct a clothing design rule base, transforming complex features into interpretable rules, facilitating designers to understand and optimize designs; pruning the decision tree to improve the generalization ability of the model and enhancing the flexibility and adaptability of designs; the generated multi-modal personalized clothing design solutions better meet the personalized needs of users, improving user experience and satisfaction; the method and system of the present invention provide strong support for the intelligentization of clothing design, promoting the industry to transform towards intelligentization and personalization. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a schematic flowchart of a multi-modal personalized clothing design method provided by an embodiment of the present invention.

[0016] Figure 2 is a schematic diagram of a multi-modal personalized clothing design monitoring system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0017] Exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in a form and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present invention can be fully conveyed to those skilled in the art.

[0018] As Figure 1 shown, an embodiment of the present invention provides a multimodal personalized clothing design method, and the method includes the following steps: Step 11, collecting multimodal data of a user; Step 12, obtaining the key features of the multimodal data by extracting the key features of each modality data; Step 13, constructing a multi-layer perceptron according to the key features of the multimodal data; Step 14, analyzing the multimodal data according to the multi-layer perceptron to generate a user personalized feature vector; Step 15, constructing a decision tree according to the user personalized feature vector, and generating a clothing design rule base according to the decision tree; Step 16, applying the KNN algorithm according to the clothing design rule base to calculate the similarity between the clothing to be designed and the clothing in the training set; Step 17, performing personalized design through the similarity to generate a multimodal personalized clothing design scheme.

[0019] In the embodiment of the present invention, collecting multimodal data, including visual, text, physiological and behavioral data, helps to comprehensively understand the user, avoid the limitations of single data, and improve the comprehensiveness of the design. By extracting the key features of each modality and generating a comprehensive feature vector through a weighted fusion mechanism, the user portrait is more comprehensive, and the accuracy and robustness of feature representation are improved. According to the key features of the multimodal data, a multi-layer perceptron is constructed, and the nonlinear mapping ability of the multi-layer perceptron (MLP) is used to automatically learn the complex interaction relationships between multimodal features, improve the effect of personalized feature extraction, and generate a user personalized feature vector. By analyzing the data through the MLP, accurate user features are obtained, providing a basis for subsequent decision-making. A decision tree and a rule base are constructed to convert complex features into interpretable rules, enhancing the transparency and adjustability of the design, and facilitating designers to understand and optimize. Applying the KNN algorithm to calculate the similarity can quickly find similar design cases, improve the design efficiency, and reduce repetitive labor. Performing personalized design through the similarity, combining multimodal data to generate innovative solutions, balancing classic and personalized, and improving the design quality.

[0020] In a preferred embodiment of the present invention, the above step 12 may include: Step 121, obtain multimodal data, where the multimodal data includes visual modal data, text modal data, physiological modal data, and behavioral modal data; Step 122, preprocess the multimodal data to obtain preprocessed multimodal data; Step 123, extract features from each type of preprocessed modal data to obtain corresponding unimodal feature vectors; Step 124, assign a weight to the feature vector of each modality, and fuse the unimodal feature vectors into a comprehensive multimodal feature vector to obtain the key features of the multimodal data, where 、 、 、 respectively represent the originally obtained visual modal data, text modal data, physiological modal data, and behavioral modal data, 、 、 、 respectively represent the feature vectors of the visual modality, text modality, physiological modality, and behavioral modality, 、 、 、 respectively represent the weights of the feature vectors of the visual modality, text modality, physiological modality, and behavioral modality, represents the comprehensive multimodal feature vector.

[0021] In the embodiment of the present invention, by integrating four types of heterogeneous data, namely visual, text, physiological, and behavioral data, a "full-dimensional portrait" of user needs is constructed. The multi-source heterogeneous data is preprocessed, including normalization, noise reduction, and missing value filling, to eliminate dimension differences, correct sensor errors, and improve the signal-to-noise ratio of the data. For each type of modal data, key feature vectors are extracted, effectively reducing the data dimension while retaining key information, facilitating subsequent fusion and analysis. Special methods are used to extract features for different modalities to ensure the effectiveness and specificity of the features. Through a dedicated algorithm, the dimension of the unimodal feature vector is compressed to 5%-10% of the original data, while retaining more than 92% of the discriminant information. The attention mechanism is introduced to automatically learn the modal weights, and the softmax function is used to ensure that the weights sum to 1, solving the "scene-function" matching contradiction. The feature vectors of each modality are fused to form a comprehensive feature vector, integrating the information of each modality, making up for the deficiencies of a single modality, and improving the robustness and accuracy of the model. This multimodal feature engineering system through a three-stage process of "decomposition - extraction - fusion" not only retains the modality-specific information but also realizes cross-modal knowledge transfer, providing a high-quality feature basis for subsequent personalized modeling and promoting the development of clothing design towards intelligence and precision.

[0022] In the embodiments of the present invention, the specific steps include: Step 121, collect raw data containing different modalities from multiple data sources, specifically including visual modality data (V), text modality data (T), physiological modality data (P), and behavioral modality data (B); Step 122, for the obtained raw multi-modal data, perform standardization processing such as denoising and cleaning on the visual modality data, text modality data, physiological modality data, and behavioral modality data, map the data of different modalities to a unified scale (such as the 0-1 range), and eliminate the dimension difference; if there are timestamp differences in the data (such as different sampling rates of video frames and physiological signals), align the time axis through interpolation or resampling; Step 123, according to the preprocessed data, extract the core features of each modality respectively to generate corresponding single-modal feature vectors; extract features such as facial key points, expression classification, and action amplitude for the visual modality; extract features such as keywords, semantic themes, and sentiment polarity through a word vector model or sentiment analysis algorithm for the text modality; calculate physiological indicators such as heart rate variability, peak skin conductivity, and EEG band energy (such as alpha waves, beta waves) for the physiological modality; count behavioral features such as gesture frequency, operation path complexity, and interaction duration for the behavioral modality; Step 124, according to the importance of each modality to the analysis target, assign weights to the single-modal feature vectors. The weights can be determined by a data-driven method. By performing a linear combination of the single-modal feature vectors to generate a comprehensive multi-modal feature vector and obtain the key features of the multi-modal data.

[0023] In the determination of each weight, assume that in the multi-modal personalized clothing design method, a large amount of historical data is collected, including visual modality data, text modality data, physiological modality data, and behavioral modality data. Perform preprocessing on the historical data, including operations such as data cleaning and normalization, to ensure the comparability and consistency of the data of each modality. Then, divide the processed data into a training set and a test set, and evaluate the feature importance of the training set data through a random forest machine learning algorithm. The random forest machine learning algorithm can automatically calculate the prediction ability of each feature for the target variable (such as user satisfaction) and give the corresponding feature importance score. According to the feature importance score, perform normalization processing on the feature importance score to obtain the relative importance of each feature, and then use it as the weight of the data of this modality. For example, assume that the feature importance evaluation results show that: the feature importance score of the visual modality data is 0.5, the feature importance score of the text modality data is 0.3, the feature importance score of the physiological modality data is 0.15, and the feature importance score of the behavioral modality data is 0.05; these scores can be normalized to obtain the weights: the weight of the visual modality data is and the weight of the text modality data is , the weight of the physiological modality data is , the weight of the behavioral modality data is , finally, linearly combine the feature vectors of each modality data according to the calculated weights to generate a comprehensive multi-modal feature vector.

[0024] In a preferred embodiment of the present invention, the above step 13 may include: Step 131, construct a multi-layer perceptron model architecture according to the key features of the multi-modal data. The multi-layer perceptron model architecture includes an input layer, a hidden layer, the dimensions of the output layer, and an activation function; Step 132, train and optimize the multi-layer perceptron model architecture to obtain an optimized multi-layer perceptron model architecture.

[0025] In the embodiment of the present invention, constructing a multi-layer perceptron model architecture according to the key features of the multi-modal data can ensure that the model effectively captures and integrates the unique information in different modality data; by precisely matching the multi-modal data features to design the model architecture, the model can better adapt to different sources and types of data, enhance the performance of the model on unseen data, improve the generalization ability, and avoid the occurrence of overfitting. By training and optimizing the multi-layer perceptron model architecture and continuously adjusting the parameters of the model, the model can more accurately fit the training data, thereby improving the prediction accuracy of the model on the test data and enhancing the robustness of the model. Using a suitable optimization algorithm (such as stochastic gradient descent and its variants Adam, Adagrad, etc.) to train the model can accelerate the convergence speed of the model, reduce the training time, and obtain better model effects within a limited time.

[0026] In the embodiment of the present invention, the specific steps include: Step 131, preprocess the extracted feature vectors of each modality according to the key features of the multi-modal data, and splice the processed feature vectors in a certain order to form a comprehensive feature vector; the number of neurons in the input layer of the multi-layer perceptron model is set to the dimension of this comprehensive feature vector. When determining the hidden layer architecture, it is necessary to comprehensively consider the complexity of the task and the scale of the data; it is crucial to select a suitable activation function for the hidden layer, the purpose of which is to introduce non-linearity so that the model can learn the complex non-linear relationships in the data. However, in certain specific cases, such as when the output value needs to be within a specific range, other activation functions can also be considered. Determine the dimensions and activation functions of the output layer according to the type of task to construct a complete multi-layer perceptron model architecture; Step 132: Divide the multimodal data into a training set, a validation set, and a test set according to the multi-layer perceptron model architecture; use the training set to iteratively train the model. In each iteration, input the data into the model to obtain a prediction result, calculate the loss function value, and use an optimization algorithm to update the model's parameters. For example, set the number of training epochs to 100, and in each epoch, batch the training set data (batch size = 32) and input it into the model for training. During the training process, regularly evaluate the model using the validation set and calculate the performance metrics of the model on the validation set (such as accuracy, recall, F1 value, etc.). According to the performance on the validation set, adjust the hyperparameters of the model, such as the learning rate, the number of neurons in the hidden layer, the regularization parameter, etc. For example, if it is found that the model is overfitting on the validation set, the strength of the regularization term can be increased or the number of neurons in the hidden layer can be reduced. After multiple trainings and optimizations, use the test set to finally evaluate the model to obtain the optimized multi-layer perceptron model architecture.

[0027] In a preferred embodiment of the present invention, the above step 14 may include: Step 141: Process the multimodal data according to the multi-layer perceptron model to obtain the predicted value of the personalized feature vector; Step 142: Evaluate the predicted result of the personalized feature vector to obtain an evaluation result; Step 143: Correct the predicted value of the personalized feature vector according to the evaluation result to generate the user's personalized feature vector.

[0028] In the embodiments of the present invention, the multi-layer perceptron model can analyze and fuse these multimodal data, thereby mining more comprehensive and accurate predicted values of personalized feature vectors, providing strong support for subsequent user portrait construction and precise services; using the multi-layer perceptron model can quickly process these multimodal data and generate predicted values of personalized feature vectors in a timely manner, providing support for the platform's real-time recommendation and content filtering. By evaluating the predicted result of the personalized feature vector, problems and deviations in the prediction can be discovered in a timely manner, providing a clear direction for subsequent correction work to ensure the prediction quality of the personalized feature vector. The corrected user's personalized feature vector more accurately and comprehensively reflects the user's true characteristics and needs, ensuring better protection of the user's privacy information.

[0029] In the embodiments of the present invention, the specific steps include: Step 141: Collect and process the multimodal data according to the multi-layer perceptron model. Input the prepared multimodal data into the multi-layer perceptron model. The model will process the input data layer by layer according to its internal structure and parameters, extract features and perform non-linear transformations to obtain the predicted value of the personalized feature vector; Step 142: According to the specific requirements of the task, select appropriate evaluation metrics. For example, if the task is a classification task, metrics such as accuracy, recall, and F1-score can be selected; if the task is a regression task, metrics such as mean squared error (MSE) and mean absolute error (MAE) can be selected. For each sample, obtain its true personalized feature vector label, and use the selected evaluation metric to compare the predicted value of the personalized feature vector with the true label to calculate the evaluation value. For example, for a classification task, calculate the accuracy between the predicted result and the true label, and for a regression task, calculate the mean squared error between the predicted value and the true value; Step 143: According to the evaluation results, determine the correction strategy for the predicted value of the personalized feature vector. If the evaluation results show that the prediction error of the model on certain features is large, consider adding relevant feature inputs, adjusting the parameters or structure of the model. If there are problems with the data quality, further clean and preprocess the data. According to the determined correction strategy, correct the predicted value of the personalized feature vector. After correction, obtain the final user personalized feature vector.

[0030] In a preferred embodiment of the present invention, the above step 15 may include: Step 151: According to the user personalized feature vector, divide the data, construct the nodes and branches of the decision tree, and set the parameters of the decision tree; Step 152: Prune the decision tree to obtain the processed decision tree; Step 153: Starting from the root node of the decision tree, traverse each node and branch. Each node represents a judgment condition for a feature dimension, and the branch represents the decision path corresponding to different feature values; Step 154: According to the judgment conditions of the nodes and the branch directions, generate clothing design rules, and organize and classify all the generated rules to form a clothing design rule library.

[0031] In the embodiment of the present invention, constructing a decision tree based on the user personalized feature vector can fully consider the unique preferences and features of each user, construct nodes and branches that better meet the user's needs, and thus lay a foundation for generating accurate clothing design rules in the follow-up. Appropriate parameter settings can enable the decision tree to better capture the patterns and rules in the data without overfitting the training data. Through pruning, the decision tree can better adapt to unseen data, avoid overfitting, reduce the number of nodes and branches, thereby reducing the computational complexity of the model and improving the efficiency of the entire process. Traversing the nodes and branches of the decision tree can systematically analyze the relationships between different feature dimensions and their impacts on the final decision. It helps fashion designers deeply understand user needs and market trends, grasp the design key points as a whole, and improve the quality and innovation of designs. The generated fashion design rule library provides designers with a standardized design basis, making the design process more rule-based, and helps fashion enterprises achieve an organic combination of personalized customization and mass production, thereby enhancing the competitiveness and market share of the enterprises.

[0032] In the embodiment of the present invention, the specific steps include: Step 151: Analyze the user's personalized feature vector to determine the possible value range and importance of each feature; divide the data set into different subsets according to the user's personalized feature vector, and the data in each subset has similar feature values, providing a basis for constructing the nodes and branches of the decision tree in the subsequent steps; select a feature as the root node of the decision tree, for example, select the style preference as the root node, and construct corresponding branches according to different values of this feature, and determine the maximum depth of the decision tree to prevent the decision tree from being too complex. Step 152: Select a suitable pruning strategy. Common pruning strategies include pre-pruning and post-pruning. Pre-pruning is to perform pruning during the construction of the decision tree, while post-pruning is to perform pruning after the decision tree is constructed. Prune the decision tree according to the selected pruning strategy. For example, remove some branches with a small number of leaf nodes and little impact on the classification result; after pruning, use the validation data set to evaluate the performance of the decision tree again to ensure that the pruning operation improves the generalization ability of the decision tree. Step 153: Start from the root node of the decision tree, record the feature dimension and judgment condition of the current node. For each branch of the root node, enter the corresponding sub-node according to the feature value corresponding to the branch, and repeat the above process to traverse each sub-node and branch in turn until the leaf node is traversed. During the traversal process, record the information of each node and branch to form a complete decision path. For example, record the path from the root node to the leaf node as: style preference (simple) - color (white) - material (cotton). Step 154: Generate corresponding fashion design rules according to the traversed decision path, sort out all the generated design rules, and remove duplicate and contradictory rules. For example, if there are two rules "style preference (simple) - color (white)" and "style preference (simple) - color (non-white)", adjust or merge them according to the actual situation; classify the rules according to different feature dimensions or design requirements, and store the sorted and classified rules in the database to form a fashion design rule library.

[0033] In a preferred embodiment of the present invention, the above step 16 may include: Step 161: Extract the clothing features of the training set according to the fashion design rule library. Step 162: Determine the value of K according to the extracted clothing features of the training set. Step 163: According to the value of K, calculate the distances between the clothing features of the training set and the clothing features of the clothing to be designed, and obtain all the distance values, where represents the clothing feature vector of the th piece of training set clothing and the clothing feature vector of the clothing to be designed, is the th dimension eigenvalue of the clothing feature vector of the th piece of training set clothing, is the th dimension eigenvalue of the clothing feature vector of the clothing to be designed; Step 164: Sort all the distance values, and determine the similarity between the clothing to be designed and the clothing in the training set according to the sorting result.

[0034] In the embodiment of the present invention, the clothing design rule base is systematically sorted and classified. The clothing features of the training set are extracted from the rule base, and the key information related to clothing design can be accurately obtained. The features extracted based on the rule base are screened and verified, and have high quality and representativeness, which improves the quality of the training set data and helps to improve the accuracy and reliability of the subsequent K-Nearest Neighbor algorithm (KNN). By calculating the distance values, the similarity between the clothing to be designed and each piece of clothing in the training set can be clearly understood, providing a basis for the subsequent similarity sorting. Determining the similarity according to the sorted distance values can provide a clear design reference for the clothing to be designed, making the clothing to be designed more distinctive and attractive.

[0035] In the embodiment of the present invention, the specific steps include: Step 161: According to the content of the rule base, define the clothing features of the training set to be extracted, screen the defined features, remove those features that have little impact on clothing design or are difficult to quantify, ensure that the extracted features are representative and practical, collect data containing information related to the training set clothing, and use means such as image recognition technology and text mining technology to extract the defined features from the collected data; Step 162: Analyze the extracted clothing features of the training set to understand the distribution and characteristics of the features, evaluate the similarity and difference between the features, and judge the complexity and noise level of the dataset. According to the feature analysis results, formulate a selection strategy for the value of K. Generally speaking, if the dataset is relatively simple and has less noise, a smaller value of K can be selected; if the dataset is relatively complex and has more noise, a larger value of K should be selected. The cross-validation method can be used to try different values of K, and based on the results of cross-validation, determine the final value of K. For example, after multiple cross-validations, it is found that when K = 5, the accuracy of the algorithm is the highest, so the value of K is determined to be 5. Step 163: Represent the extracted clothing features of the training set and the clothing features to be designed as feature vectors respectively, ensuring that the dimensions of the clothing feature vectors of the training set and the clothing features to be designed are the same for distance calculation. According to the calculation formula of the Euclidean distance, calculate the Euclidean distance between the th clothing feature vector of the training set and the clothing feature vector to be designed. Calculate the distances between all the clothing feature vectors of the training set and the clothing feature vector to be designed in turn to obtain all the distance values. Step 164: Use a sorting algorithm (such as quicksort, mergesort, etc.) to sort all the distance values in ascending order. The sorted distance values can intuitively reflect the similarity degree between the clothing in the training set and the clothing to be designed. The smaller the distance value, the higher the similarity. According to the sorting results, determine the similarity between the clothing to be designed and the clothing in the training set. The clothing in the training set corresponding to the first K distance values can be selected as the clothing most similar to the clothing to be designed. For example, if K = 5, select the first 5 pieces of clothing in the training set with the smallest distance values and consider them to have the highest similarity to the clothing to be designed. Present the similarity sorting results to the designer in an intuitive way, such as a list, a chart, etc. The designer can refer to the similar clothing in the training set according to the similarity sorting results for design, improving the design efficiency and quality.

[0036] In a preferred embodiment of the present invention, the above step 17 may include: Step 171: Screen out the clothing samples that match the similarity of the clothing to be designed through similarity. Step 172: Extract the common design elements by analyzing the screened clothing samples with similarity. Step 173: Combine them together through adjustment and optimization according to the common design elements to generate a multi-modal personalized clothing design scheme.

[0037] In an embodiment of the present invention, by screening clothing samples by similarity, designers do not need to search one by one in a massive amount of clothing styles, and can quickly and accurately locate samples that are similar to the clothing to be designed in style, design, audience, etc., saving time for searching and screening; extracting common design elements from the screened similar clothing samples avoids the designer's complicated analysis process of a large number of different styles of clothing, and can focus on core design elements more efficiently, providing a clear direction for the generation of subsequent design solutions, and accelerating the conversion of design inspiration; extracting common design elements of samples and applying them to new designs can draw on successful design experience, so that the generated clothing design solution is more in line with market demand and aesthetic trends, and improves the rationality and feasibility of the design.

[0038] In the embodiment of the present invention, the specific steps include: Step 171, determine the similarity evaluation dimension, collect relevant data such as pictures and text descriptions (including style, fabric, color, etc.) of various types of clothing, build a comprehensive clothing database for subsequent similarity screening, select a suitable similarity algorithm based on the evaluation dimension and data characteristics, input the relevant information of the clothing to be designed into the algorithm, compare it with the clothing data in the database, sort it from high to low according to the similarity, and screen out clothing samples that match the similarity of the clothing to be designed; Step 172, through the selected clothing sample images, from the overall version, outline to the local decoration, pattern and other aspects, find out the visual features that appear repeatedly in multiple samples, such as specific neckline shapes (square neck, round neck, etc.), cuff designs (flared sleeves, puff sleeves, etc.), pattern types (stripes, polka dots, etc.), combined with the text description of the clothing, analyze the design points mentioned therein, further determine the common design elements, summarize the visual analysis and text information, remove some accidental or non-representative features, and determine the truly common design elements; Step 173, make detailed adjustments to the extracted common design elements, such as changing the size and density of the pattern, adjusting the saturation and brightness of the color, or modifying certain dimensions of the pattern to adapt to different design needs and target audiences, taking into account the functionality, comfort and aesthetics of the clothing, optimizing the design elements, and making diversified combinations of the adjusted and optimized design elements, trying different matching methods, and creating multiple different clothing styles. It can be combined with different occasions and style requirements to generate clothing design solutions suitable for daily wear, formal occasions, sports scenes and other situations.

[0039] like Figure 2 As shown, an embodiment of the present invention further provides a multi-modal personalized clothing design monitoring system, comprising: An acquisition module, configured to collect multimodal data of a user; by extracting key features of each type of modal data, to obtain key features of the multimodal data; A construction module, configured to construct a multi-layer perceptron according to the key features of the multimodal data; according to the multi-layer perceptron, analyze the multimodal data through an Adam optimizer to generate a user personalized feature vector; according to the user personalized feature vector, construct a decision tree, and according to the decision tree, generate a clothing design rule base; A processing module, configured to apply a KNN algorithm according to the clothing design rule base to calculate the similarity between the clothing to be designed and the clothing in the training set; through the similarity, perform personalized design to generate a multimodal personalized clothing design scheme.

[0040] It should be noted that this system corresponds to the above method, and all implementation manners in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0041] An embodiment of the present invention further provides a computing device, including: a processor, and a memory storing a computer program, when the computer program is run by the processor, it executes the method as described above. All implementation manners in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

Claims

1. A multimodal personalized clothing design method, characterized in that: The method comprises: Collect multimodal data from users; By extracting the key features of each modal data, the key features of multimodal data are obtained; Construct a multi-layer perceptron based on the key features of multimodal data; According to the multi-layer perceptron, the multimodal data is analyzed to generate a user personalized feature vector; According to the user's personalized feature vector, a decision tree is constructed, and based on the decision tree, a clothing design rule base is generated; According to the clothing design rule library, the KNN algorithm is applied to calculate the similarity between the clothing to be designed and the clothing in the training set; Through similarity, personalized design is performed to generate multimodal personalized clothing design solutions.

2. The multimodal personalized clothing design method according to claim 1, characterized in that: By extracting the key features of each modal data, the key features of multimodal data are obtained, including: Acquire multimodal data, which includes visual modality data, textual modality data, physiological modality data, and behavioral modality data; Preprocessing the multimodal data to obtain preprocessed multimodal data; Perform feature extraction on each modal data after preprocessing to obtain the corresponding single-modal feature vector; Assign a weight to each modality's eigenvector, by The single-modal feature vectors are fused into comprehensive multimodal feature vectors to obtain the key features of multimodal data, where: , , , Respectively represent the original acquired visual modality data, text modality data, physiological modality data and behavioral modality data, , , , Represent the feature vectors of visual modality, textual modality, physiological modality and behavioral modality respectively, , , , Represent the weights of the feature vectors of visual modality, textual modality, physiological modality and behavioral modality respectively, Represents the comprehensive multimodal feature vector.

3. The multimodal personalized clothing design method according to claim 2, characterized in that: According to the key features of multimodal data, a multi-layer perceptron is constructed, including: According to the key features of multimodal data, a multi-layer perceptron model architecture is constructed, which includes input layer, hidden layer, output layer dimensions and activation function; The multi-layer perceptron model architecture is trained and optimized to obtain an optimized multi-layer perceptron model architecture.

4. The multimodal personalized clothing design method according to claim 3, characterized in that: According to the multi-layer perceptron, the multimodal data is analyzed to generate a user-personalized feature vector, including: According to the multi-layer perceptron model, multimodal data is processed to obtain personalized feature vector prediction values; Evaluate the personalized feature vector prediction result to obtain an evaluation result; According to the evaluation result, the personalized feature vector prediction value is modified to generate the user personalized feature vector.

5. The multimodal personalized clothing design method according to claim 4, characterized in that: According to the user's personalized feature vector, a decision tree is constructed, and based on the decision tree, a clothing design rule base is generated, including: According to the user's personalized feature vector, the data is divided, the nodes and branches of the decision tree are constructed, and the parameters of the decision tree are set; Prune the decision tree to obtain a processed decision tree; Starting from the root node of the decision tree, traverse each node and branch. Each node represents a judgment condition of a feature dimension, and the branch represents the decision path corresponding to different feature values. According to the judgment conditions and branch directions of the nodes, clothing design rules are generated, and all generated rules are sorted and classified to form a clothing design rule library.

6. The multimodal personalized clothing design method according to claim 5, characterized in that: According to the clothing design rule base, the KNN algorithm is applied to calculate the similarity between the clothing to be designed and the clothing in the training set, including: Extract clothing features of the training set based on the clothing design rule library; Determine the K value based on the extracted clothing features of the training set; According to the K value, Calculate the distance between the feature vectors of the training set clothing and the clothing to be designed, and get all the distance values, where: Indicates The feature vector of clothing in the training set and the feature vector of the clothing to be designed The Euclidean distance between It is The first clothing feature vector of the training set dimensional eigenvalue, is the first feature vector of the clothing to be designed dimensional eigenvalue; All distance values ​​are sorted, and the similarity between the clothing to be designed and the clothing in the training set is determined based on the sorting results.

7. The multimodal personalized clothing design method according to claim 6, characterized in that: Through similarity, personalized design is performed to generate multimodal personalized clothing design solutions, including: Through similarity, clothing samples that match the clothing to be designed are screened out; By analyzing the selected similar clothing samples, common design elements are extracted; Based on the common design elements, they are adjusted and optimized and combined together to generate multimodal personalized clothing design solutions.

8. A multimodal personalized clothing design system, the system being used to execute the method as claimed in any one of claims 1 to 7, characterized in that: include: The acquisition module is used to collect multimodal data of users; By extracting the key features of each modal data, the key features of multimodal data are obtained; A building module is used to build a multi-layer perceptron based on the key features of the multi-modal data; based on the multi-layer perceptron, the multi-modal data is analyzed through the Nadam optimizer to generate a user personalized feature vector; based on The user personalizes the feature vector, builds a decision tree, and generates a clothing design rule base based on the decision tree; A processing module is used to calculate the similarity between the garment to be designed and the garment in the training set by applying the KNN algorithm according to the garment design rule library; Through similarity, personalized design is performed to generate multimodal personalized clothing design solutions.

9. A computing device, characterized in that include: one or more processors; A storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a program, which, when executed by a processor, implements the method according to any one of claims 1 to 7.

Images