CN120146972B - Virtual fitting personalized clothing recommendation method based on artificial intelligence - Google Patents

Abstract

Translated fromChinese

本发明公开了一种基于人工智能的虚拟试衣个性化服装推荐方法，包括如下步骤：S1、采集用户图像、文本描述与行为数据，生成多模态用户特征集；S2、构建多模态异构语义图谱，并进行结构建模与嵌入表示；S3、图谱嵌入与特征融合，构建正负样本对，进行语义对齐训练；S4、利用浣熊优化算法，优化网络结构参数与训练超参数；S5、计算语义匹配得分，生成推荐候选，输入体型参数，生成多角度虚拟试衣图像；S6、采集用户行为反馈，更新图谱边权与训练样本，执行闭环优化。本发明的作用是实现基于用户多模态特征与体型参数的个性化服饰推荐与虚拟试衣图像生成，构建推荐—试穿—反馈闭环，提升推荐准确性与用户体验。

The invention discloses a virtual fitting personalized clothing recommendation method based on artificial intelligence, comprising the following steps: S1, collecting user images, text descriptions and behavior data, generating a multimodal user feature set; S2, constructing a multimodal heterogeneous semantic map, and performing structural modeling and embedding representation; S3, map embedding and feature fusion, constructing positive and negative sample pairs, and performing semantic alignment training; S4, using the raccoon optimization algorithm to optimize network structure parameters and training hyperparameters; S5, calculating semantic matching scores, generating recommendation candidates, inputting body shape parameters, and generating multi-angle virtual fitting images; S6, collecting user behavior feedback, updating the map edge weight and training samples, and performing closed-loop optimization. The function of the present invention is to realize personalized clothing recommendation and virtual fitting image generation based on user multimodal features and body shape parameters, construct a recommendation-try-on-feedback closed loop, and improve recommendation accuracy and user experience.

Description

Virtual fitting personalized clothing recommendation method based on artificial intelligence

Technical Field

The invention relates to the technical field of artificial intelligence, in particular to a virtual fitting personalized clothing recommendation method based on artificial intelligence.

Background

With the rapid development of electronic commerce and mobile internet, online apparel retail platforms are rapidly spreading, and users are increasingly inclined to purchase apparel products in an online manner. The fusion of the virtual fitting technology and the personalized recommendation technology has become one of the key paths for improving the user experience of the e-commerce platform. Traditional e-commerce platforms mainly rely on rule matching or recommendation systems based on collaborative filtering algorithms to recommend clothes related to historical browsing and purchasing behaviors to users. However, such recommendation methods present significant limitations in dealing with problems of user interest migration, body conformation difference adaptation, new user cold start, etc., and are difficult to effectively support personalized, visual, and intelligent clothing recommendation needs.

On the other hand, early development of virtual fitting technologies mainly relied on static template-based or simple image synthesis methods to position and superimpose or map head portraits or photos uploaded by users with merchandise apparel images. The method not only lacks sense of reality and immersion, but also can not accurately simulate the try-on effect according to the real body type parameters of the user, so that the try-on image has strong visual deception and low actual reference value. Meanwhile, the existing fitting system is often isolated from a recommendation system, and cannot realize end-to-end fusion recommendation flow based on user preference and body type characteristics. In other words, the virtual fitting system exists only as a display module and cannot participate in the clothes screening and intelligent matching process.

In recent years, with the maturity of deep learning technology, especially graphic neural network, contrast learning, self-supervision learning and other technologies, more and more research attempts are being made to introduce multi-modal feature fusion, semantic graph modeling and personalized embedded representation methods so as to improve the accuracy and user satisfaction of a recommendation system. For example, there are studies on embedding learning of a relationship graph between a user and a commodity by using a graph neural network, or on joint embedding training of images and texts to promote semantic expression ability of a recommendation result. However, most of such methods focus on the accuracy optimization of the recommendation itself, and how the recommendation result is linked and verified with the actual wearing experience of the user is not fully considered, and a closed loop fusion mechanism from the recommendation to the try-on is lacked.

In addition, conventional recommendation systems generally lack the ability to dynamically optimize model structures. Once the structural parameters and the training super parameters of the model are set, the structural parameters and the training super parameters are difficult to dynamically adjust according to feedback in the process of recommending tasks. The current system lacks an optimization algorithm framework coupled with the feedback depth of the user behavior, and when facing complex task scenes (such as multi-layer semantic alignment, multi-angle image generation and multi-objective recommendation effect evaluation), the current system often falls into problems of local optimum or unstable convergence and the like, so that the popularization capability and the self-adaptive generalization capability of the model are limited.

Further, existing recommendation systems often ignore the modeling of the user's body type differences during the recommendation process. Most of the system defaults that users have stable and consistent acceptability to clothes sizes and formats, and direct relations between body types and try-on effects are ignored, so that generated recommended results lack persuasion and credibility in the try-on stage. Even if some systems allow users to select body types, the modeling mode is quite rough, and only stays at the text labeling or template selection level, so that the influence of body type parameters on the generation quality of the try-on image cannot be truly reflected.

More notably, the interactive feedback mechanism between the virtual fitting and the recommendation system is not perfect at present. The behavior data such as clicking, browsing, scoring, collecting and the like generated after the user views the try-on image are not effectively collected and modeled, and further cannot be reversely acted on the evolution of the semantic map structure and the reconstruction of the training sample, so that the system lacks the capability of continuous learning and individual evolution. Most of traditional recommendation systems make static recommendation based on static models, and a closed loop structure of recommendation-feedback-optimization-re-recommendation is difficult to form.

In summary, the conventional virtual fitting and personalized clothing recommendation system has obvious technical defects and shortcomings that firstly, a recommendation algorithm and user body type modeling lack of deep fusion, so that recommendation results lack of adaptability to real wearing, secondly, semantic map modeling and embedding modes cannot be aligned with multi-modal characteristics (images, texts and behaviors) of a user efficiently, thirdly, the recommendation model lacks a dynamic optimization mechanism of a structural layer, cannot adaptively adjust a structure and parameter configuration according to feedback data, thirdly, an end-to-end closed loop feedback updating mechanism is lacking, user behavior feedback cannot be effectively used for dynamic adjustment of a map structure and a sample strategy, and thirdly, the virtual fitting image generation process does not combine semantic recommendation results to perform joint modeling, so that pertinence and personalized rendering capability are lacked.

Therefore, how to provide a virtual fit personalized clothing recommendation method based on artificial intelligence is a problem to be solved by those skilled in the art.

Disclosure of Invention

The invention aims to provide an artificial intelligence-based virtual fitting personalized clothing recommendation method, which fuses multi-mode semantic map modeling, a cross-mode contrast learning network, a virtual fitting image generation technology and a behavior feedback-driven optimization algorithm, and details how to realize a closed loop flow of clothing personalized recommendation and multi-angle fitting image generation based on image data, text description, body type parameters and interactive behaviors of a user.

According to the embodiment of the invention, the virtual fitting personalized clothing recommendation method based on artificial intelligence comprises the following steps of:

S1, acquiring image data, text description data and historical interaction behavior data of a user to generate a multi-mode user feature set;

S2, constructing a multi-mode heterogeneous semantic graph, and modeling and embedding a multi-mode heterogeneous semantic graph structure to represent;

s3, extracting structured embedded vectors of user nodes and clothing nodes from the multi-modal heterogeneous semantic graphs, inputting the structured embedded vectors and the multi-modal user feature sets into a cross-modal contrast learning network model, and carrying out semantic alignment training in a shared embedded space by constructing positive and negative sample pairs;

s4, optimizing the structural parameters and the training super parameters of the cross-modal comparison learning network model through a raccoon optimization algorithm to generate an optimized cross-modal comparison learning network model;

S5, applying the optimized cross-modal comparison learning network model to a recommendation task, calculating a semantic matching degree score between a user and clothes, generating a recommendation candidate set, inputting the recommendation candidate set and user body type parameters into a virtual fitting image generation unit together, generating a fitting image of the user wearing the recommendation clothes by combining the clothes image, and outputting an interactable multi-angle fitting view;

And S6, collecting behavior feedback data of the user on the try-on image, generating a user behavior feedback feature vector, and updating the edge weight in the multi-modal heterogeneous semantic map and the training sample composition of the cross-modal contrast learning network, and periodically executing the steps S2 to S5 to form a closed-loop personalized recommendation flow of the self-adaptive iterative optimization.

Optionally, the multi-mode user feature set is generated by fusing a visual feature vector, a semantic feature vector and a behavior feature vector, wherein the visual feature vector is extracted by inputting image data of a user to a convolutional neural network, the semantic feature vector is extracted by inputting text description data to a language understanding unit, and the behavior feature vector is generated by encoding historical behavior data of the user.

Optionally, the S2 specifically includes:

S21, constructing a node set, wherein the node set comprises user nodes, clothing nodes and clothing attribute nodes, the user nodes are used for representing user objects with individual identifications, the clothing nodes are used for representing recommended target clothing objects, and the clothing attribute nodes are used for representing label information of styles, colors, seasons and brands of clothing;

s22, establishing an edge relation between a user node and a clothing node based on user history interaction behavior data, wherein the edge relation is used for representing that clicking, collecting, purchasing and fitting behavior association exists between the user and the clothing;

S23, establishing an edge relation between the clothing nodes and the clothing attribute nodes based on the clothing metadata and the tag information, wherein the edge relation is used for representing attribute association of styles, colors and brands of clothing;

s24, setting an initial side weight for the constructed side relation, wherein the side weight is set according to the interaction frequency of the user, the behavior type and the association strength between similar labels, and is used for adjusting the recommended path and the graph nerve propagation weight;

s25, carrying out structural modeling on the multi-mode heterogeneous semantic graph by adopting a heterogeneous graph modeling method, so that heterogeneous relation information among various nodes is reserved, and meanwhile, a complete graph structural representation is established;

S26, embedding and representing the multi-modal heterogeneous semantic graph based on the graph neural network structure, and encoding various nodes in the multi-modal heterogeneous semantic graph into a structured embedded vector through a multi-layer information aggregation mechanism, wherein the embedded vector is used for inputting a cross-modal contrast learning network model.

Optionally, the step S3 specifically includes:

S31, extracting structured embedded vectors of user nodes and clothing nodes from the multi-mode heterogeneous semantic graph, wherein the structured embedded vectors are respectively expressed asAndAnd extracting a fusion representation vector from the multimodal user feature set;

S32, introducing a map fusion gating coefficientConstructing a nonlinear gating fusion mechanism, embedding and fusing the semantic of the map structure and the modality, and generating a user representation vector:

;

Wherein, theAs a function of the Sigmoid,For the element-level multiplication to be performed,In the case of a multi-layer perceptron network,Fusing gating coefficients for the map;

S33, representing the vector of the userGarment structure embedding vectorConstructing cross-modal sample pairsConstructing a positive sample pair and a negative sample pair according to historical interaction information of a user and clothes;

s34, introducing a semantic multi-layer comparison mechanism, and setting the number of semantic comparison layer numbersConstructing a plurality of semantic granularity comparison tasks, including overall matching comparison, style attribute comparison and color semantic comparison, and performing independent loss calculation on each layer;

s35, introducing a multi-factor driving dynamic temperature control mechanism into each contrast layer to define the firstThe temperature parameters of the training iterations are:

;

Wherein, theIs used as a basic temperature super-parameter,,,A time adjustment factor, a similarity variance adjustment factor, and a loss sensitivity adjustment factor,Representing the variance of the positive and negative samples of the current batch versus the similarity,A comparative loss value representing the current lot,As a logarithmic function;

S36, training the semantic alignment of the user representation vector and the clothing representation vector in the shared embedded space by using the cross-modal contrast learning network model and minimizing the multi-layer semantic contrast loss function of the weighted fusion.

Optionally, the step S4 specifically includes:

S41, setting optimized structural parameters and training superparameters in a cross-modal contrast learning network model to form an optimized target set comprising map fusion gating coefficientsSemantic contrast layer progressionTime adjustment factorSimilarity variance adjustment factorAnd loss sensitivity adjustment factor;

S42, dividing the optimization target set into structure fusion subspaces based on parameter function attributesWith training control subspacesAnd initializing two raccoon sub-populations、Each raccoon individual represents a set of parameter combinations to be optimized;

S43, executing a local memory driving mechanism, an environment disturbance mechanism and a collaborative guiding updating mechanism of a raccoon optimization algorithm in each sub-population, performing multi-round iterative optimization on the individual raccoon population, and storing each round of optimal raccoon individual into a corresponding sub-population memory bank;

S44, introducing a dynamic memory window mechanism, and setting the length of the current iterative memory window as:

;

Wherein, theFor the initial window length to be the same,For the feedback sensitivity adjustment factor to be used,For the current round fitness score, the window length is used to control the number of optimal solutions retained by the subgroup memory,For the length of the memory window in the current round of optimization,For the fitness score of the previous round,Is a very small normal number constant;

s45, pressing the memory banks of each sub-population after each round of optimizationMemorizing window length limitation, reserving a local optimal individual, and replacing an outdated historical solution;

S46, introducing an attention migration mechanism of semantic perception of the map every other timeIn the round migration period, calculating an inter-population migration attention vector based on the edge density change between the user node and the attribute node in the multi-mode heterogeneous semantic graph:

;

Wherein, the、Indicating that the sub-population embeds impact weights on the user nodes and apparel nodes,Representing the change value of the weight intensity from the user to the attribute in the map,Representing sub-populationsSub-population ofThe attention weighting coefficient of the migration,For the weight adjustment factor of the structure change of the map edge,For normalization operations, inter-population migration operations are ultimately performed:

;

Wherein, theIs a sub-populationMiddle (f)The parameters of the individual raccoons represent vectors,Is a sub-populationThe current wheel adaptability of the raccoon individual parameter vector is optimal;

S47, respectively selecting the optimal individual parameters of the current round from the two sub-populations、Merging into globally optimal parameter combinations;

S48, defining a compound fitness function of triple consistency evaluationThe following are provided:

;

Wherein, theRepresenting positive sample pairs at the firstAverage similarity in the layer semantic space,In order to recommend the reconstruction error of the multi-angle virtual try-on image of the clothing,The ranking quality index between the user behavior feedback and the recommendation ranking is calculated according to the damage accumulation gain by comparing the actual clicking, collecting and purchasing behaviors of the user with the recommendation list, and is used for measuring the matching degree of the recommendation ranking result and the actual preference of the user,,,Is a balance coefficient;

s49, sorting all raccoon individuals according to the fitness function calculation result, and determining the current global optimal parameter combination;

S410, combining the optimal parametersThe method is applied to the cross-modal comparison learning network model, and the map fusion mechanism, the semantic comparison hierarchical structure and the temperature regulation strategy are updated to output the finally optimized cross-modal comparison learning network model.

Optionally, the step S5 specifically includes:

s51, deploying the optimized cross-modal comparison learning network model into a recommendation task module, inputting a fusion feature vector and a clothing embedding vector of a user, and executing semantic similarity calculation;

S52, according to the semantic similarity calculation result, matching degree scoring and sorting are carried out on candidate clothes, and a plurality of clothes with highest matching degree are selected from the candidate clothes to form a recommendation candidate set;

S53, acquiring body type parameter information of a user, including body type dimension characteristics of height, weight, shoulder width, waistline and hip circumference, and carrying out standardized modeling on the body type parameter information of the user;

S54, inputting the image characteristics of each piece of clothing in the recommended candidate set, the body type parameters and the fusion characteristics of the user into a virtual fitting image generation unit together, and executing image synthesis operation of clothing fitting effects;

S55, in the image synthesis process, respectively setting a plurality of observation visual angles, and generating corresponding multi-angle fitting images aiming at each piece of candidate clothes;

s56, organizing the generated multi-angle image set into an interactive display interface, wherein a user can conduct visual preview on each piece of recommended clothing, and the visual preview comprises interactive operations of image switching, rotation, scaling and body type fitting effect comparison.

Optionally, the body type parameter information of the user specifically includes body type dimension characteristics of height, weight, shoulder width, waistline and hip circumference, and the body type dimension characteristics are used for driving the virtual fitting image generating unit to generate a fitting image conforming to the body type characteristics of the user.

Optionally, the step S6 specifically includes:

S61, after the user finishes the try-on image browsing of the recommended clothing, collecting behavior feedback data of the user;

S62, preprocessing and encoding the collected behavior feedback data to generate a user behavior feedback feature vector for describing the current preference change of the user;

S63, updating edge weights in the multi-mode heterogeneous semantic graphs based on the user behavior feedback feature vectors, wherein the updating comprises adjustment of connection strength between user nodes and clothing nodes and between user nodes and attribute nodes, and reflects user interest reconstruction trend;

s64, re-extracting the structure embedded representation of the user and the clothing node according to the updated multi-mode heterogeneous semantic map structure, wherein the structure embedded representation is used for generating a new semantic alignment training sample composition and comprises the relation update of a positive sample pair and a negative sample pair;

S65, periodically re-executing multi-mode semantic map modeling, cross-mode contrast learning training and raccoon optimization parameter updating, recommending and fitting image generation processes to form a dynamic self-updated recommended iteration closed loop;

And S66, after the closed loop optimization period of each round is finished, adjusting a semantic graph structure, a matching strategy and a visual synthesis mode in a recommendation process according to the accumulated user behavior data change condition, and improving adaptability and feedback response capability of personalized recommendation.

Optionally, the behavior feedback data of the user specifically includes click, browsing duration, image switching, grading and interaction information of collection, and is used for dynamically adjusting semantic map structures and training sample composition, so as to optimize personalized recommendation effects.

The beneficial effects of the invention are as follows:

According to the artificial intelligence-based virtual fitting personalized clothing recommendation method, a complete closed loop system from semantic understanding, interest matching to visual verification and feedback learning is constructed by introducing key technical components such as a multi-mode heterogeneous semantic map, a cross-mode contrast learning network model, a raccoon optimization algorithm, a virtual fitting image generation unit and the like on the basis of the prior art. Compared with the traditional recommendation system which only relies on image retrieval or collaborative filtering, the invention realizes the multi-mode end-to-end processing flow from user data acquisition to try-on image output, and truly opens up barriers between three stages of recommendation, fitting and feedback.

The multi-mode heterogeneous semantic map constructed by the invention fully integrates user images, text descriptions, behavior records and clothing attribute data, and is embedded and represented through a graph neural structure, so that the complex relationship between user interest preferences and clothing semantic tags is effectively captured. Based on the structured embedding of the atlas extraction, the atlas extraction and the multi-modal characteristics of the user are fused and then sent into a cross-modal contrast learning network, and the matching accuracy of the user-clothing pair in the semantic space is improved by using a layered semantic alignment mechanism. By introducing raccoon optimization algorithm to search and optimize the model structure and super parameters, the generalization capability and the recommendation stability of the model are obviously enhanced, and particularly, the method still has good effect in the scenes with large personality difference and serious cold start.

In addition, the recommendation candidate result is innovatively combined with the user body type parameter, and the virtual fitting image generating unit is introduced to generate the interactive multi-angle fitting image, so that a user can not only know what to recommend, but also see the wearing effect, and the user participation feeling and recommendation trust degree are greatly improved. The try-on image not only exists as a display result, but also is in linkage with clicking, scoring, collection and other behaviors of the user, so that quantifiable behavior feedback characteristics are formed. The system dynamically updates the semantic map side weights and the training sample structure through the feedback information, and periodically retrains the recommendation model, so that the self-adaptive capturing and real-time personalized recommendation capability of the user interests is realized.

In conclusion, the invention breaks through the technical bottleneck of the traditional recommendation system in the aspects of personal suitability, visual interactivity and self-optimizing capability, achieves the comprehensive optimizing effects of more accurate recommendation, more realistic fitting, more intelligent feedback and more autonomous system, and has higher practicability and popularization value.

Drawings

The accompanying drawings are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate the invention and together with the embodiments of the invention, serve to explain the invention. In the drawings:

FIG. 1 is a flow chart of a virtual fitting personalized clothing recommendation method based on artificial intelligence;

fig. 2 is a flowchart of a raccoon optimization algorithm of an artificial intelligence-based virtual fitting personalized clothing recommendation method for jointly optimizing model structural parameters and training super-parameters.

Detailed Description

The invention will now be described in further detail with reference to the accompanying drawings. The drawings are simplified schematic representations which merely illustrate the basic structure of the invention and therefore show only the structures which are relevant to the invention.

Referring to fig. 1 and 2, an artificial intelligence based virtual fitting personalized clothing recommendation method includes the steps of:

S1, acquiring image data, text description data and historical interaction behavior data of a user to generate a multi-mode user feature set;

S2, constructing a multi-mode heterogeneous semantic graph, and modeling and embedding a multi-mode heterogeneous semantic graph structure to represent;

s3, extracting structured embedded vectors of user nodes and clothing nodes from the multi-modal heterogeneous semantic graphs, inputting the structured embedded vectors and the multi-modal user feature sets into a cross-modal contrast learning network model, and carrying out semantic alignment training in a shared embedded space by constructing positive and negative sample pairs;

s4, optimizing the structural parameters and the training super parameters of the cross-modal comparison learning network model through a raccoon optimization algorithm to generate an optimized cross-modal comparison learning network model;

S5, applying the optimized cross-modal comparison learning network model to a recommendation task, calculating a semantic matching degree score between a user and clothes, generating a recommendation candidate set, inputting the recommendation candidate set and user body type parameters into a virtual fitting image generation unit together, generating a fitting image of the user wearing the recommendation clothes by combining the clothes image, and outputting an interactable multi-angle fitting view;

And S6, collecting behavior feedback data of the user on the try-on image, generating a user behavior feedback feature vector, and updating the edge weight in the multi-modal heterogeneous semantic map and the training sample composition of the cross-modal contrast learning network, and periodically executing the steps S2 to S5 to form a closed-loop personalized recommendation flow of the self-adaptive iterative optimization.

According to the virtual fitting personalized clothing recommendation method provided by the invention, a closed-loop recommendation system integrating data acquisition, multi-mode modeling, semantic alignment, intelligent optimization, visual generation and feedback learning is constructed through the steps S1 to S6, so that the virtual fitting personalized clothing recommendation method has remarkable beneficial effects. The method not only integrates the multi-mode user characteristics such as images, texts, behaviors and the like, but also constructs a heterogeneous semantic map to comprehensively express the multi-dimensional relationship between the user interests and the clothing semantics. Through the training of the cross-modal contrast learning network model and the parameter tuning of the raccoon optimization algorithm, the semantic matching accuracy and the model generalization capability are effectively improved. The system links the recommended result and the user body type parameter to generate a multi-angle try-on image with high sense of reality, and realizes direct landing from 'recommendation' to 'visual wearing'. More importantly, the behavior feedback of the user on the fitting image is acquired in real time and used for updating the semantic graph structure and training sample composition, so that a self-adaptive optimization closed loop of recommendation, fitting, feedback and recommendation is formed. The method remarkably improves the individuation level, recommendation credibility and user experience of clothing recommendation, and has good practicability and popularization value.

In this embodiment, the multi-modal user feature set is generated by fusing a visual feature vector, a semantic feature vector and a behavior feature vector, wherein the visual feature vector is generated by inputting image data of a user to a convolutional neural network for extraction, the semantic feature vector is generated by inputting text description data to a language understanding unit for extraction, and the behavior feature vector is generated by encoding historical behavior data of the user.

In this embodiment, the S2 specifically includes:

S21, constructing a node set, wherein the node set comprises user nodes, clothing nodes and clothing attribute nodes, the user nodes are used for representing user objects with individual identifications, the clothing nodes are used for representing recommended target clothing objects, and the clothing attribute nodes are used for representing label information of styles, colors, seasons and brands of clothing;

s22, establishing an edge relation between a user node and a clothing node based on user history interaction behavior data, wherein the edge relation is used for representing that clicking, collecting, purchasing and fitting behavior association exists between the user and the clothing;

S23, establishing an edge relation between the clothing nodes and the clothing attribute nodes based on the clothing metadata and the tag information, wherein the edge relation is used for representing attribute association of styles, colors and brands of clothing;

s24, setting an initial side weight for the constructed side relation, wherein the side weight is set according to the interaction frequency of the user, the behavior type and the association strength between similar labels, and is used for adjusting the recommended path and the graph nerve propagation weight;

s25, carrying out structural modeling on the multi-mode heterogeneous semantic graph by adopting a heterogeneous graph modeling method, so that heterogeneous relation information among various nodes is reserved, and meanwhile, a complete graph structural representation is established;

S26, embedding and representing the multi-modal heterogeneous semantic graph based on the graph neural network structure, and encoding various nodes in the multi-modal heterogeneous semantic graph into a structured embedded vector through a multi-layer information aggregation mechanism, wherein the embedded vector is used for inputting a cross-modal contrast learning network model.

The multi-mode heterogeneous semantic map is systematically constructed, has the capability of accurately expressing the multi-dimensional relationship between the user and the clothes, and has obvious beneficial effects. By introducing user nodes, clothing nodes and clothing attribute nodes and combining the historical click, collection, purchase and trial-pass actions of the user, an interactive graph structure with rich expression is constructed, so that a recommendation system not only can understand the user behavior, but also can identify semantic causes behind the user preference. Meanwhile, a clothing attribute relation network is constructed by using clothing metadata and tag information, so that the interpretability and classification capability of clothing semantic information are enhanced. By setting the edge weight, the system can dynamically adjust the connection weight between nodes according to the interaction frequency and the behavior intensity, and provide data support for the propagation path and the attention mechanism in the graph neural network. The introduction of the heterogeneous graph structure ensures the independence and the expandability of the node relations of different types, and the embedded representation of the graph neural network enables the node semantic features to be uniformly encoded into the trainable vector representation, so that the subsequent cross-mode contrast learning is facilitated. In the whole, the map modeling method greatly improves the expression capability of the recommendation system in the aspects of user interest modeling and semantic association learning, and provides a high-quality structural basis for subsequent recommendation effect improvement and personalized semantic alignment.

In this embodiment, the step S3 specifically includes:

S31, extracting structured embedded vectors of user nodes and clothing nodes from the multi-mode heterogeneous semantic graph, wherein the structured embedded vectors are respectively expressed asAndAnd extracting a fusion representation vector from the multimodal user feature set;

S32, introducing a map fusion gating coefficientConstructing a nonlinear gating fusion mechanism, embedding and fusing the semantic of the map structure and the modality, and generating a user representation vector:

;

Wherein, theAs a function of the Sigmoid,For the element-level multiplication to be performed,In the case of a multi-layer perceptron network,Fusing gating coefficients for the map;

S33, representing the vector of the userGarment structure embedding vectorConstructing cross-modal sample pairsConstructing a positive sample pair and a negative sample pair according to historical interaction information of a user and clothes;

s34, introducing a semantic multi-layer comparison mechanism, and setting the number of semantic comparison layer numbersConstructing a plurality of semantic granularity comparison tasks, including overall matching comparison, style attribute comparison and color semantic comparison, and performing independent loss calculation on each layer;

s35, introducing a multi-factor driving dynamic temperature control mechanism into each contrast layer to define the firstThe temperature parameters of the training iterations are:

;

Wherein, theIs used as a basic temperature super-parameter,,,A time adjustment factor, a similarity variance adjustment factor, and a loss sensitivity adjustment factor,Representing the variance of the positive and negative samples of the current batch versus the similarity,A comparative loss value representing the current lot,As a logarithmic function;

S36, training the semantic alignment of the user representation vector and the clothing representation vector in the shared embedded space by using the cross-modal contrast learning network model and minimizing the multi-layer semantic contrast loss function of the weighted fusion.

The invention provides a cross-modal contrast learning method for fusing the semantic of a map structure and multi-modal feature information, which has the remarkable beneficial effects of improving the matching precision of personalized recommended semantic and the robustness of a model. By extracting the pattern structure of the user and the clothing, embedding and combining the multi-mode user characteristics, the system realizes flexible control on pattern semantic enhancement by using a gating mechanism, so that the end user representation has both structure information and mode perception capability. By constructing positive and negative sample pairs and introducing a semantic multi-layer comparison mechanism, the system can learn semantic similarity between a user and clothes on multiple granularity levels, and the capturing capability of the implicit semantic hierarchy in user preference is remarkably improved. Meanwhile, a multi-factor driven dynamic temperature control mechanism is introduced, and temperature parameters are dynamically adjusted by combining time progress, similarity distribution and training loss, so that the gradient of contrast loss is more stable, and the model training process has more adaptability and generalization capability. Finally, the network model optimizes the representation consistency of the user and the clothes in the shared embedded space by weighting and fusing multi-layer semantic comparison loss, realizes the modeling capability of a recommendation system with richer semantic hierarchy, more accurate recommendation and quicker convergence, and provides a high-quality matching scoring basis for the generation of subsequent recommendation candidates.

In this embodiment, the S4 specifically includes:

S41, setting optimized structural parameters and training superparameters in a cross-modal contrast learning network model to form an optimized target set comprising map fusion gating coefficientsSemantic contrast layer progressionTime adjustment factorSimilarity variance adjustment factorAnd loss sensitivity adjustment factor;

S42, dividing the optimization target set into structure fusion subspaces based on parameter function attributesWith training control subspacesAnd initializing two raccoon sub-populations、Each raccoon individual represents a set of parameter combinations to be optimized;

S43, executing a local memory driving mechanism, an environment disturbance mechanism and a collaborative guiding updating mechanism of a raccoon optimization algorithm in each sub-population, performing multi-round iterative optimization on the individual raccoon population, and storing each round of optimal raccoon individual into a corresponding sub-population memory bank;

S44, introducing a dynamic memory window mechanism, and setting the length of the current iterative memory window as:

;

Wherein, theFor the initial window length to be the same,For the feedback sensitivity adjustment factor to be used,For the current round fitness score, the window length is used to control the number of optimal solutions retained by the subgroup memory,For the length of the memory window in the current round of optimization,For the fitness score of the previous round,Is a very small normal number constant;

s45, pressing the memory banks of each sub-population after each round of optimizationMemorizing window length limitation, reserving a local optimal individual, and replacing an outdated historical solution;

S46, introducing an attention migration mechanism of semantic perception of the map every other timeIn the round migration period, calculating an inter-population migration attention vector based on the edge density change between the user node and the attribute node in the multi-mode heterogeneous semantic graph:

;

Wherein, the、Indicating that the sub-population embeds impact weights on the user nodes and apparel nodes,Representing the change value of the weight intensity from the user to the attribute in the map,Representing sub-populationsSub-population ofThe attention weighting coefficient of the migration,For the weight adjustment factor of the structure change of the map edge,For normalization operations, inter-population migration operations are ultimately performed:

;

Wherein, theIs a sub-populationMiddle (f)The parameters of the individual raccoons represent vectors,Is a sub-populationThe current wheel adaptability of the raccoon individual parameter vector is optimal;

S47, respectively selecting the optimal individual parameters of the current round from the two sub-populations、Merging into globally optimal parameter combinations;

S48, defining a compound fitness function of triple consistency evaluationThe following are provided:

;

Wherein, theRepresenting positive sample pairs at the firstAverage similarity in the layer semantic space,In order to recommend the reconstruction error of the multi-angle virtual try-on image of the clothing,The ranking quality index between the user behavior feedback and the recommendation ranking is calculated according to the damage accumulation gain by comparing the actual clicking, collecting and purchasing behaviors of the user with the recommendation list, and is used for measuring the matching degree of the recommendation ranking result and the actual preference of the user,,,Is a balance coefficient;

s49, sorting all raccoon individuals according to the fitness function calculation result, and determining the current global optimal parameter combination;

S410, combining the optimal parametersThe method is applied to the cross-modal comparison learning network model, and the map fusion mechanism, the semantic comparison hierarchical structure and the temperature regulation strategy are updated to output the finally optimized cross-modal comparison learning network model.

The invention provides a cross-modal contrast learning network parameter optimization method based on an improved raccoon optimization algorithm, which mainly integrates a dynamic memory window mechanism and a sub-population attention migration mechanism of atlas semantic perception, and realizes the combined efficient optimization of structural parameters and training super-parameters. Firstly, the invention divides the model parameters into two subspaces of structural fusion and training control, and initializes raccoon sub-population respectively, so that each type of parameters can adaptively evolve in independent space. By introducing a dynamic memory window mechanism, the system can dynamically adjust the depth of the memory bank according to adaptability fluctuation in the optimization process, so that the local convergence risk of an early optimal solution is avoided. Further, the sub-population attention migration mechanism for semantic perception of the map guides weighted migration of parameter knowledge among different subgroups based on real-time change of edge weight density between a user and attribute nodes, and semantic relevance and task-crossing generalization capability of an optimization direction are effectively improved. Finally, the system jointly evaluates semantic matching precision, image generation quality and user feedback ordering indexes through a triple consistency fitness function, and ensures that the selected optimal parameter combination has omnibearing performance in a real recommended scene. Overall, the method effectively improves model optimization efficiency, adaptation capacity of a recommendation system and stability of a training process, and has high practicability and innovation.

In this embodiment, the step S5 specifically includes:

s51, deploying the optimized cross-modal comparison learning network model into a recommendation task module, inputting a fusion feature vector and a clothing embedding vector of a user, and executing semantic similarity calculation;

S52, according to the semantic similarity calculation result, matching degree scoring and sorting are carried out on candidate clothes, and a plurality of clothes with highest matching degree are selected from the candidate clothes to form a recommendation candidate set;

S53, acquiring body type parameter information of a user, including body type dimension characteristics of height, weight, shoulder width, waistline and hip circumference, and carrying out standardized modeling on the body type parameter information of the user;

S54, inputting the image characteristics of each piece of clothing in the recommended candidate set, the body type parameters and the fusion characteristics of the user into a virtual fitting image generation unit together, and executing image synthesis operation of clothing fitting effects;

S55, in the image synthesis process, respectively setting a plurality of observation visual angles, and generating corresponding multi-angle fitting images aiming at each piece of candidate clothes;

s56, organizing the generated multi-angle image set into an interactive display interface, wherein a user can conduct visual preview on each piece of recommended clothing, and the visual preview comprises interactive operations of image switching, rotation, scaling and body type fitting effect comparison.

The invention constructs a personalized recommendation execution flow integrating semantic recommendation and visual try-on, and has remarkable practicability and user experience improving effect. By disposing the optimized cross-modal comparison learning network model in the recommendation task module, the system can accurately calculate semantic similarity between the fusion characteristics of the user and the clothes embedding, and complete screening and sequencing of candidate clothes based on the matching degree, so that the recommendation result is ensured to be highly fit with the potential interests of the user. On the basis, user body shape parameter information is introduced and standardized modeling is carried out, so that recommendation is not only stopped at a semantic level, but also accurate depiction of user body shape characteristics is realized. And then, the system inputs the recommendation result and the user body type information to the virtual fitting image generation unit in a combined way, so as to generate a multi-angle wearing image with a simulation effect, and the visual experience of the recommendation content is remarkably improved. Through the multi-view output and interactive display interface, the user can carry out realism judgment and personalized decision on the recommended clothes at the visual level, so that the user trust and participation sense are enhanced. In the whole, the method not only improves the recommendation accuracy, but also opens up a recommendation closed loop from 'interest identification' to 'cross-lap verification', and has remarkable user friendliness and commercial application value.

In this embodiment, the body shape parameter information of the user specifically includes body shape dimension characteristics of height, weight, shoulder width, waistline and hip circumference, and is used to drive the virtual fitting image generating unit to generate a fitting image according with the body shape characteristics of the user.

In this embodiment, the step S6 specifically includes:

S61, after the user finishes the try-on image browsing of the recommended clothing, collecting behavior feedback data of the user;

S62, preprocessing and encoding the collected behavior feedback data to generate a user behavior feedback feature vector for describing the current preference change of the user;

S63, updating edge weights in the multi-mode heterogeneous semantic graphs based on the user behavior feedback feature vectors, wherein the updating comprises adjustment of connection strength between user nodes and clothing nodes and between user nodes and attribute nodes, and reflects user interest reconstruction trend;

s64, re-extracting the structure embedded representation of the user and the clothing node according to the updated multi-mode heterogeneous semantic map structure, wherein the structure embedded representation is used for generating a new semantic alignment training sample composition and comprises the relation update of a positive sample pair and a negative sample pair;

S65, periodically re-executing multi-mode semantic map modeling, cross-mode contrast learning training and raccoon optimization parameter updating, recommending and fitting image generation processes to form a dynamic self-updated recommended iteration closed loop;

And S66, after the closed loop optimization period of each round is finished, adjusting a semantic graph structure, a matching strategy and a visual synthesis mode in a recommendation process according to the accumulated user behavior data change condition, and improving adaptability and feedback response capability of personalized recommendation.

The invention constructs a dynamic self-adaptive optimization mechanism based on user behavior feedback drive, and remarkably improves the learning ability and long-term performance of the personalized recommendation system. According to the method, after the user finishes the try-on image browsing of the recommended clothes, behavior data such as clicking, stay time, scoring, collection and the like are timely collected and encoded into behavior feedback feature vectors for describing user preference changes. The system adjusts the side weights in the multi-mode heterogeneous semantic graph based on the feedback characteristics, so that the structural reconstruction of the user interests is realized, and the connection strength between the user nodes and the clothes or attribute nodes dynamically evolves. Meanwhile, based on the updated map structure, the system automatically updates the semantic alignment training sample, and ensures that the training process continuously fits the latest preference of the user. By periodically re-executing the map modeling, model training, parameter optimization and recommendation processes, the system builds a closed-loop architecture of recommendation, feedback, optimization and re-recommendation, and has continuous learning capability. Finally, the system can also adjust the semantic propagation path and the image synthesis strategy according to the historical behavior change, so that the individuation adaptability and the response speed are further improved. Overall, the method realizes the feedback sensing and the self-evolution capability of the system in the real sense, and has the beneficial effects of long-term individual fitting and accurate and continuous recommendation.

In this embodiment, the behavior feedback data of the user specifically includes click, browsing duration, image switching, scoring and interaction information of collection, which are used for dynamically adjusting semantic map structures and training sample composition, and optimizing personalized recommendation effects.

Example 1:

in order to verify the feasibility of the invention in implementation, the invention is applied to a large electronic commerce platform, the platform selects 200 new users which do not generate purchasing behavior within 30 days, wherein 100 people use the existing traditional clothing recommendation system as a comparison group, and 100 people use an intelligent recommendation system which is provided with the method of the invention as an experiment group, and the system completes the complete flow from clothing recommendation to simulated image generation to feedback learning in an automatic and personalized mode.

When a user accesses the platform for the first time, the experiment group user needs to upload a clear front half body photograph, fill out dressing preference and brief body type description, automatically identify image features by the platform, and generate multi-mode user feature vectors by combining user browsing history, behavior tracks and text information. The system then models semantic relationships between users and clothes based on the constructed multi-modal heterogeneous semantic graphs, and inputs a cross-modal contrast learning network model to perform semantic alignment training. The model regulated by raccoon optimization algorithm is used for generating personalized clothing recommendation candidate sets, and the system generates multi-angle fitting images for the recommended clothing after combining body type parameters such as the height, weight, shoulder width, waistline, hip circumference and the like of the user, so that the user can interactively browse and fit on line.

Compared with the traditional system, the system provided by the invention has the advantages that the recommendation quality and the user experience are obviously improved. From experimental monitoring data, the recommended Top-3 accuracy rate in the experimental group reaches 85.1%, the accuracy rate is improved by about 16 percent compared with the traditional system, the click rate of the virtual fitting image reaches 75.2%, and the click rate is far higher than that of the control group by 42.6%. The average browsing time of the user on the try-on image page is increased from 36 seconds to 59 seconds, the high attention of the user on the try-on image content is displayed, the purchase conversion rate is increased from 32.8% to 42.1%, the body type satisfaction degree score is also increased from 3.9 to 4.7 minutes, and the user commonly feeds back the try-on effect is real and the matching degree is high. In addition, the average number of behavior feedback generated by each user in the experimental group is 22 times and is 1.7 times that of the control group, so that sufficient data is provided for subsequent self-adaptive optimization of the system.

Table 1 key index comparison table of the inventive system and conventional recommendation system

According to the data in table 1, it can be clearly seen that the personalized virtual fitting recommendation system provided by the invention is comprehensively superior to the traditional recommendation system in terms of a plurality of core indexes, and shows remarkable performance advantages and user experience improvement. Firstly, in the aspect of recommending the Top-3 accuracy, the system reaches 85.1%, and the traditional recommendation system is only 69.1%, and the lifting amplitude is up to 16 percentage points. The recommendation model trained by the atlas modeling, cross-modal contrast learning and raccoon optimization algorithm can more accurately identify the user interests and match the user interests to the most suitable clothing commodity, and the recommendation accuracy is remarkably improved.

At the rate at which the user clicks on the virtual fitting image, the experimental group user click rate was 75.2%, while the control group was only 42.6%. The difference reflects that the virtual fitting image generated by the invention has higher attraction and interaction value, can effectively guide the user to explore the recommendation result deeply, and is a direct embodiment of the simulation and individuality suitability of the image generation unit. The average browsing time length of the user in recommending the try-on page is also increased from 36 seconds to 59 seconds of the traditional system, so that the system disclosed by the invention can attract the user to pay attention to the recommended content more permanently, and the participation degree of the user and the system viscosity are improved. This "immersive recommendation" experience is particularly critical to promoting conversion.

The system of the invention also has obvious advantages in purchasing conversion, reaching 42.1%, and improving the conversion rate by nearly 10 percent compared with 32.8% of the traditional system. This means that more accurate recommendation and visual try-on effect is helpful for users to make more confident purchasing decisions, and commodity sales potential is greatly improved from a commercial perspective.

In the aspect of subjective experience, the body type satisfaction degree of a user on a try-on image is scored to be 4.7 minutes (full score of 5 minutes), and the traditional system is 3.9 minutes, which shows that the wearing image generated by combining the body type parameters of the user in the method is more fit with the real body shape of the user, and the acceptance degree and satisfaction degree of the user are improved.

Finally, in the aspect of behavior feedback number, the system provided by the invention has 22 times of interaction behaviors generated by users on average per person, and the traditional system is only 13 times, so that the system is proved to not only enhance the interaction activity of the users, but also accumulate more high-quality training data for the subsequent recommendation optimization of the system, and strengthen the continuous learning capability of the model.

In the system, through introducing multi-mode feature fusion, semantic map modeling, intelligent optimization algorithm and visual fitting experience, the accuracy and individuation degree of a recommendation result are improved, the interactive experience and commercial conversion value of a user are enhanced, and the technical feasibility and market popularization potential of the method in a real application scene are fully verified.

The foregoing is only a preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art, who is within the scope of the present invention, should make equivalent substitutions or modifications according to the technical scheme of the present invention and the inventive concept thereof, and should be covered by the scope of the present invention.

Claims (8)

Translated fromChinese

1.一种基于人工智能的虚拟试衣个性化服装推荐方法，其特征在于，包括如下步骤：1. A virtual fitting personalized clothing recommendation method based on artificial intelligence, characterized in that it includes the following steps:S1、采集用户的图像数据、文本描述数据以及用户的历史交互行为数据，生成多模态用户特征集；S1, collect user image data, text description data and user historical interaction behavior data to generate a multimodal user feature set;S2、构建多模态异构语义图谱，并对多模态异构语义图谱结构进行建模与嵌入表示；S2. Construct a multimodal heterogeneous semantic graph, and model and embed the structure of the multimodal heterogeneous semantic graph;S3、从多模态异构语义图谱中提取用户节点与服饰节点的结构化嵌入向量，并将结构化嵌入向量和多模态用户特征集输入跨模态对比学习网络模型，通过构建正负样本对在共享嵌入空间中进行语义对齐训练；S3, extracting structured embedding vectors of user nodes and clothing nodes from the multimodal heterogeneous semantic graph, and inputting the structured embedding vectors and multimodal user feature sets into the cross-modal contrastive learning network model, and performing semantic alignment training in the shared embedding space by constructing positive and negative sample pairs;S4、通过浣熊优化算法对跨模态对比学习网络模型的结构参数与训练超参数进行优化，生成优化后的跨模态对比学习网络模型；S4. Optimizing the structural parameters and training hyperparameters of the cross-modal contrastive learning network model through the raccoon optimization algorithm to generate an optimized cross-modal contrastive learning network model;S5、将优化后的跨模态对比学习网络模型应用于推荐任务中，计算用户与服饰之间的语义匹配度得分，生成推荐候选集合，将推荐候选集合与用户体型参数共同输入至虚拟试衣图像生成单元，结合服饰图像生成用户穿着推荐服饰的拟真图像，输出可交互的多角度试衣视图；S5. Apply the optimized cross-modal contrastive learning network model to the recommendation task, calculate the semantic matching score between the user and the clothing, generate a recommendation candidate set, input the recommendation candidate set and the user's body shape parameters into the virtual fitting image generation unit, combine the clothing image to generate a realistic image of the user wearing the recommended clothing, and output an interactive multi-angle fitting view;S6、采集用户对试穿图像的行为反馈数据，生成用户行为反馈特征向量，用于更新多模态异构语义图谱中的边权值与跨模态对比学习网络的训练样本构成，周期性执行步骤S2至步骤S5，形成自适应迭代优化的闭环个性化推荐流程；S6, collecting user behavior feedback data on the try-on images, generating user behavior feedback feature vectors, which are used to update the edge weights in the multimodal heterogeneous semantic graph and the training sample composition of the cross-modal contrastive learning network, and periodically executing steps S2 to S5 to form a closed-loop personalized recommendation process with adaptive iterative optimization;所述S2具体包括：The S2 specifically includes:S21、构建节点集合，包括用户节点、服饰节点以及服饰属性节点，所述用户节点用于表示具有个体标识的用户对象，服饰节点用于表示可推荐的目标服饰对象，服饰属性节点用于表示服饰的风格、颜色、季节和品牌的标签信息；S21, constructing a node set, including a user node, a clothing node, and a clothing attribute node, wherein the user node is used to represent a user object with an individual identifier, the clothing node is used to represent a target clothing object that can be recommended, and the clothing attribute node is used to represent label information of the style, color, season, and brand of the clothing;S22、基于用户历史交互行为数据，建立用户节点与服饰节点之间的边关系，所述边关系用于表示用户与服饰之间存在点击、收藏、购买和试穿的行为关联；S22, based on the user's historical interaction behavior data, establishing an edge relationship between the user node and the clothing node, wherein the edge relationship is used to indicate that there is a behavior association between the user and the clothing, such as click, favorite, purchase, and try-on;S23、基于服饰元数据与标签信息，建立服饰节点与服饰属性节点之间的边关系，所述边关系用于表示服饰所具备的风格、颜色和品牌的属性关联；S23, based on the clothing metadata and label information, establishing an edge relationship between the clothing node and the clothing attribute node, wherein the edge relationship is used to represent the attribute association of the style, color and brand of the clothing;S24、为构建的边关系设置初始边权值，边权值依据用户交互频次、行为类型和相似标签之间的关联强度进行设定，所述边权值用于推荐路径与图神经传播权重的调整；S24, setting an initial edge weight for the constructed edge relationship, the edge weight is set according to the user interaction frequency, behavior type and the strength of association between similar tags, and the edge weight is used to adjust the recommended path and graph neural propagation weight;S25、采用异构图建模方法对所述多模态异构语义图谱进行结构建模，使各类节点间保留异构关系信息，同时建立完整的图结构表示；S25, using a heterogeneous graph modeling method to perform structural modeling on the multimodal heterogeneous semantic graph, so that heterogeneous relationship information is retained between various types of nodes, and a complete graph structure representation is established at the same time;S26、基于图神经网络结构对多模态异构语义图谱进行嵌入表示，将多模态异构语义图谱中各类节点通过多层信息聚合机制编码为结构化嵌入向量，所述嵌入向量用于跨模态对比学习网络模型的输入。S26. Based on the graph neural network structure, the multimodal heterogeneous semantic graph is embedded and represented, and various nodes in the multimodal heterogeneous semantic graph are encoded into structured embedding vectors through a multi-layer information aggregation mechanism. The embedding vectors are used as input for the cross-modal contrastive learning network model.2.根据权利要求1所述的一种基于人工智能的虚拟试衣个性化服装推荐方法，其特征在于，所述多模态用户特征集是融合视觉特征向量、语义特征向量和行为特征向量生成的，其中视觉特征向量是通过将用户的图像数据输入至卷积神经网络进行提取，语义特征向量是通过将文本描述数据输入至语言理解单元进行提取，行为特征向量是通过将用户的历史行为数据进行编码进行生成的。2. According to the artificial intelligence-based virtual fitting personalized clothing recommendation method of claim 1, it is characterized in that the multimodal user feature set is generated by fusing visual feature vectors, semantic feature vectors and behavioral feature vectors, wherein the visual feature vectors are extracted by inputting the user's image data into a convolutional neural network, the semantic feature vectors are extracted by inputting text description data into a language understanding unit, and the behavioral feature vectors are generated by encoding the user's historical behavior data.3.根据权利要求1所述的一种基于人工智能的虚拟试衣个性化服装推荐方法，其特征在于，所述S3具体包括：3. The method for recommending personalized clothing based on virtual fitting by artificial intelligence according to claim 1, wherein S3 specifically comprises:S31、从多模态异构语义图谱中提取用户节点与服饰节点的结构化嵌入向量，分别表示为和，并从多模态用户特征集中提取融合表示向量；S31. Extract the structured embedding vectors of user nodes and clothing nodes from the multimodal heterogeneous semantic graph, which are represented as and , and extract the fusion representation vector from the multimodal user feature set ;S32、引入图谱融合门控系数，构建非线性门控融合机制，将图谱结构语义与模态嵌入进行嵌套融合，生成用户表示向量：S32, introduce the gating coefficient of the atlas fusion , construct a nonlinear gated fusion mechanism, nest and fuse the graph structure semantics with the modal embedding, and generate a user representation vector :； ;其中，为Sigmoid函数，为元素级乘法，为多层感知器网络，为图谱融合门控系数；in, is the Sigmoid function, is element-wise multiplication, is a multi-layer perceptron network, is the atlas fusion gating coefficient;S33、将用户表示向量与服饰结构嵌入向量构成跨模态样本对，依据用户与服饰的历史交互信息，构建正样本对与负样本对；S33, user representation vector Embedding vector with clothing structure Constructing cross-modal sample pairs , construct positive sample pairs and negative sample pairs based on the historical interaction information between users and clothing;S34、引入语义多层对比机制，设置语义对比层级数，构建多个语义粒度的对比任务，包括整体匹配对比、风格属性对比和颜色语义对比，对每一层进行独立损失计算；S34. Introduce a semantic multi-layer comparison mechanism and set the number of semantic comparison levels , construct comparison tasks of multiple semantic granularities, including overall matching comparison, style attribute comparison, and color semantic comparison, and perform independent loss calculation for each layer;S35、在每一对比层中，引入多因子驱动的动态温度控制机制，定义第次训练迭代的温度参数为：S35. In each contrast layer, a multi-factor driven dynamic temperature control mechanism is introduced to define the The temperature parameter for the training iteration is:； ;其中，为基础温度超参数，分别为时间调节因子、相似度方差调节因子和损失敏感度调节因子，表示当前批次正负样本对相似度的方差，表示当前批次的对比损失值，为对数函数；in, is the base temperature hyperparameter, They are time adjustment factor, similarity variance adjustment factor and loss sensitivity adjustment factor respectively. Represents the variance of the similarity between the positive and negative sample pairs in the current batch, Represents the contrast loss value of the current batch, is a logarithmic function;S36、使用跨模态对比学习网络模型，通过最小化加权融合的多层语义对比损失函数，对用户表示向量与服饰表示向量在共享嵌入空间中的语义对齐训练。S36. Use a cross-modal contrastive learning network model to minimize the weighted fusion multi-layer semantic contrast loss function to train the semantic alignment of user representation vectors and clothing representation vectors in a shared embedding space.4.根据权利要求1所述的一种基于人工智能的虚拟试衣个性化服装推荐方法，其特征在于，所述S4具体包括：4. The method for recommending personalized clothing based on virtual fitting by artificial intelligence according to claim 1, wherein S4 specifically comprises:S41、设定跨模态对比学习网络模型中可优化的结构参数与训练超参数，组成优化目标集，包括图谱融合门控系数、语义对比层级数，以及时间调节因子、相似度方差调节因子和损失敏感度调节因子；S41. Set the optimizable structural parameters and training hyperparameters in the cross-modal contrastive learning network model to form an optimization target set, including the graph fusion gating coefficient , semantic contrast level , and the time adjustment factor , Similarity variance adjustment factor and loss sensitivity adjustment factor ;S42、基于参数功能属性将优化目标集划分为结构融合子空间 与训练控制子空间，并初始化两个浣熊子种群、，每个浣熊个体表示一组待优化参数组合；S42, based on parameter function attributes, divide the optimization target set into structural fusion subspaces and training control subspace , and initialize two raccoon sub-populations , , each raccoon individual represents a set of parameter combinations to be optimized;S43、在各子种群中，执行浣熊优化算法的局部记忆驱动机制、环境扰动机制与协同引导更新机制，对种群浣熊个体进行多轮迭代优化，并将每轮最优浣熊个体存入对应子群记忆库中；S43, in each sub-population, executing the local memory driving mechanism, environmental disturbance mechanism and collaborative guidance updating mechanism of the raccoon optimization algorithm, performing multiple rounds of iterative optimization on the raccoon individuals in the population, and storing the best raccoon individuals in each round into the corresponding sub-population memory bank;S44、引入动态记忆窗口机制，设当前迭代记忆窗口长度为：S44, introduce a dynamic memory window mechanism, and set the current iteration memory window length to:； ;其中，为初始窗口长度，为反馈敏感调节系数，为当前轮适应度得分，窗口长度用于控制子群记忆保留的最优解数量，为当前轮次优化过程中的记忆窗口长度，为上一轮的适应度得分，为极小正数常量；in, is the initial window length, is the feedback sensitivity adjustment coefficient, is the fitness score of the current round, and the window length is used to control the number of optimal solutions retained in the subgroup memory. is the memory window length in the current round of optimization, is the fitness score of the previous round, is a very small positive constant;S45、对每轮优化后各子种群的记忆库按记忆窗口长度限制，保留局部最优个体，替换过时的历史解；S45, the memory library of each sub-population after each round of optimization is Limit the length of the memory window to retain the local optimal individual and replace the outdated historical solution;S46、引入图谱语义感知的注意力迁移机制，在每隔轮迁移周期内，基于多模态异构语义图谱中用户节点与属性节点之间边密度变化，计算种群间迁移注意向量：S46, introduce the attention transfer mechanism of graph semantic perception, During the round migration cycle, based on the edge density changes between user nodes and attribute nodes in the multimodal heterogeneous semantic graph, the inter-population migration attention vector is calculated:； ;其中，、表示子种群对用户节点和服饰节点嵌入影响权重，表示图谱中用户到属性边权密度变化值，表示子种群向子种群迁移的注意力权重系数，为图谱边结构变化权重调节因子，为归一化操作，最终执行种群间迁移操作：in, , represents the influence weight of the subpopulation on the embedding of user nodes and clothing nodes, Indicates the change value of the edge weight density from user to attribute in the graph, Represents subpopulation Subpopulation The attention weight coefficient of the transfer, is the weight adjustment factor for the graph edge structure change, For normalization operation, finally perform inter-population migration operation:； ;其中，为子种群中第i个浣熊个体的参数表示向量，为子种群中当前轮适应度最优浣熊个体参数向量；in, Subpopulation The parameter representation vector of the i-th raccoon individual in , Subpopulation The parameter vector of the raccoon individual with the best fitness in the current round;S47、从两个子种群中分别选出当前轮最优个体参数、，合并为全局最优参数组合；S47, select the optimal individual parameters of the current round from the two sub-populations , , merged into the global optimal parameter combination ;S48、定义三重一致性评估的复合适应度函数如下：S48. Define the composite fitness function for triple consistency evaluation as follows:； ;其中，表示正样本对在第层语义空间中的平均相似度，为推荐服饰的多角度虚拟试穿图像的重建误差，为用户行为反馈与推荐排序间的排序质量指标，为平衡系数；in, Indicates that the positive sample pair is The average similarity in the layer semantic space, is the reconstruction error of multi-angle virtual try-on images of recommended clothing. It is a ranking quality indicator between user behavior feedback and recommendation ranking. is the balance coefficient;S49、依据适应度函数计算结果对所有浣熊个体进行排序，确定当前全局最优参数组合；S49. Sort all raccoon individuals according to the fitness function calculation results to determine the current global optimal parameter combination ;S410、将所述最优参数组合应用于跨模态对比学习网络模型中，更新图谱融合机制、语义对比层级结构与温度调控策略，输出最终优化后的跨模态对比学习网络模型。S410, combining the optimal parameters Applied to the cross-modal contrastive learning network model, it updates the graph fusion mechanism, semantic contrast hierarchy structure and temperature control strategy, and outputs the final optimized cross-modal contrastive learning network model.5.根据权利要求1所述的一种基于人工智能的虚拟试衣个性化服装推荐方法，其特征在于，所述S5具体包括：5. The method for recommending personalized clothing based on virtual fitting by artificial intelligence according to claim 1, wherein S5 specifically comprises:S51、将优化后的跨模态对比学习网络模型部署至推荐任务模块中，输入用户的融合特征向量与服饰嵌入向量，执行语义相似度计算；S51, deploying the optimized cross-modal contrastive learning network model to the recommendation task module, inputting the user's fused feature vector and clothing embedding vector, and performing semantic similarity calculation;S52、根据所述语义相似度计算结果，对候选服饰进行匹配度评分与排序，并从中选取匹配度最高的若干服饰构成推荐候选集合；S52, scoring and sorting the candidate clothing according to the semantic similarity calculation result, and selecting a number of clothing with the highest matching degree to form a recommended candidate set;S53、采集用户的体型参数信息，包括身高、体重、肩宽、腰围和臀围的体型维度特征，并对用户的体型参数信息进行标准化建模；S53, collecting the user's body parameter information, including body dimension characteristics of height, weight, shoulder width, waist circumference and hip circumference, and performing standardized modeling on the user's body parameter information;S54、将推荐候选集合中的每件服饰的图像特征，与用户体型参数及融合特征共同输入至虚拟试衣图像生成单元，执行服饰试穿效果的图像合成操作；S54, inputting the image features of each piece of clothing in the recommended candidate set, the user's body parameters and the fusion features into a virtual fitting image generation unit, and performing an image synthesis operation of the clothing fitting effect;S55、在图像合成过程中，分别设定多个观察视角，针对每件候选服饰生成对应的多角度拟真试穿图像；S55, during the image synthesis process, multiple observation perspectives are set respectively, and corresponding multi-angle simulated try-on images are generated for each candidate garment;S56、将生成的多角度图像集合组织为可交互式展示界面，用户能够对每件推荐服饰进行可视化预览，包括图像切换、旋转、缩放与体型拟合效果对比的交互操作。S56. Organizing the generated multi-angle image collection into an interactive display interface, the user can perform a visual preview of each recommended clothing, including interactive operations of image switching, rotation, scaling and body fitting effect comparison.6.根据权利要求5所述的一种基于人工智能的虚拟试衣个性化服装推荐方法，其特征在于，所述用户的体型参数信息具体包括身高、体重、肩宽、腰围和臀围的体型维度特征，用于驱动虚拟试衣图像生成单元生成符合用户身形特征的拟真试穿图像。6. According to the artificial intelligence-based virtual fitting personalized clothing recommendation method of claim 5, it is characterized in that the user's body parameter information specifically includes body dimension characteristics of height, weight, shoulder width, waist circumference and hip circumference, which are used to drive the virtual fitting image generation unit to generate a simulated fitting image that conforms to the user's body shape characteristics.7.根据权利要求1所述的一种基于人工智能的虚拟试衣个性化服装推荐方法，其特征在于，所述S6具体包括：7. The method for recommending personalized clothing based on virtual fitting by artificial intelligence according to claim 1, wherein S6 specifically comprises:S61、在用户完成推荐服饰的试穿图像浏览后，采集用户的行为反馈数据；S61, after the user finishes browsing the recommended clothing trial images, collecting user behavior feedback data;S62、对所采集的行为反馈数据进行预处理与编码，生成用于刻画用户当前偏好变化的用户行为反馈特征向量；S62, preprocessing and encoding the collected behavior feedback data to generate a user behavior feedback feature vector for describing the current preference change of the user;S63、基于所述用户行为反馈特征向量，对多模态异构语义图谱中的边权值进行更新，包括用户节点与服饰节点之间、用户节点与属性节点之间的连接强度调整，反映用户兴趣重构趋势；S63, based on the user behavior feedback feature vector, updating the edge weights in the multimodal heterogeneous semantic graph, including adjusting the connection strength between the user node and the clothing node, and between the user node and the attribute node, to reflect the user interest reconstruction trend;S64、根据更新后的多模态异构语义图谱结构，重新提取用户与服饰节点的结构嵌入表示，用于生成新的语义对齐训练样本构成，包括正样本对与负样本对的关系更新；S64, re-extracting the structural embedding representation of the user and clothing nodes according to the updated multimodal heterogeneous semantic graph structure, for generating a new semantic alignment training sample composition, including updating the relationship between the positive sample pairs and the negative sample pairs;S65、周期性地重新执行多模态语义图谱建模、跨模态对比学习训练、浣熊优化参数更新与推荐及试穿图像生成流程，形成动态自更新的推荐迭代闭环；S65, periodically re-execute the multimodal semantic graph modeling, cross-modal contrastive learning training, raccoon optimization parameter update and recommendation, and trial image generation process to form a dynamic self-updating recommendation iterative closed loop;S66、在每一轮闭环优化周期结束后，根据累计的用户行为数据变化情况，对推荐流程中的语义图结构、匹配策略与视觉合成方式进行调整，提升个性化推荐的适应性与反馈响应能力。S66. After each closed-loop optimization cycle, the semantic graph structure, matching strategy and visual synthesis method in the recommendation process are adjusted according to the changes in the accumulated user behavior data to improve the adaptability and feedback response capabilities of personalized recommendations.8.根据权利要求7所述的一种基于人工智能的虚拟试衣个性化服装推荐方法，其特征在于，所述用户的行为反馈数据具体包括点击、浏览时长、图像切换、评分与收藏的交互信息，用于动态调整语义图谱结构与训练样本构成，优化个性化推荐效果。8. According to the artificial intelligence-based virtual fitting personalized clothing recommendation method of claim 7, it is characterized in that the user's behavioral feedback data specifically includes clicks, browsing time, image switching, ratings and collection interaction information, which is used to dynamically adjust the semantic graph structure and training sample composition to optimize the personalized recommendation effect.