An Artificial Intelligence-Based Personalized Clothing Recommendation Method for Virtual Fitting

Abstract

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

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular, to a personalized clothing recommendation method for virtual fitting based on artificial intelligence. Background Art

[0002] With the rapid development of e-commerce and mobile Internet, online clothing retail platforms have spread rapidly, and users are increasingly inclined to purchase clothing products online. The integration of virtual fitting technology and personalized recommendation technology has become one of the key paths to improve the user experience of e-commerce platforms. Traditional e-commerce platforms mainly rely on rule matching or recommendation systems based on collaborative filtering algorithms to recommend clothing related to users' historical browsing and purchase behaviors. However, such recommendation methods show significant limitations in dealing with problems such as user interest migration, body type difference adaptation, and cold start of new users, and it is difficult to effectively support the personalized, visual, and intelligent clothing recommendation requirements.

[0003] On the other hand, the early development of virtual fitting technology mainly relied on methods based on static templates or simple image synthesis, which superimposed the user's uploaded avatar or photo on the image of the commodity clothing or performed image mapping processing. This method not only lacks a sense of reality and immersion, but also cannot accurately simulate the fitting effect according to the user's actual body type parameters, resulting in strong visual deception and low practical reference value of the fitting image. At the same time, existing fitting systems are often isolated from the recommendation system and fail to achieve an end-to-end integrated recommendation process based on user preferences and body type characteristics. In other words, the virtual fitting system only exists as a display module and does not participate in the clothing screening and intelligent matching process.

[0004] In recent years, with the maturity of deep learning technologies, especially graph neural networks, contrastive learning, self-supervised learning, etc., more and more research attempts to introduce multi-modal feature fusion, semantic graph modeling, and personalized embedding representation methods to improve the accuracy of recommendation systems and user satisfaction. For example, some research uses graph neural networks to perform embedding learning on the relationship graph between users and commodities, or jointly embeds and trains images and texts to improve the semantic expression ability of recommendation results. However, most of these methods focus on the accuracy optimization of "recommendation itself" and have not fully considered how the recommendation results are linked and verified with the user's actual dressing experience, lacking a closed-loop integration mechanism from "recommendation" to "fitting".

[0005] In addition, traditional recommendation systems generally lack the ability to dynamically optimize the model structure. Once the structural parameters and training hyperparameters of the model are set, it is difficult to dynamically adjust them according to feedback during the recommendation task. The current system lacks an optimization algorithm framework that is deeply coupled with user behavior feedback. When facing complex task scenarios (such as multi-layer semantic alignment, multi-angle image generation, and multi-objective recommendation effect evaluation), it often falls into problems such as local optimality or unstable convergence, which limits the promotion ability and adaptive generalization ability of the model.

[0006] Furthermore, existing recommendation systems often ignore the modeling of users' body shape differences during the recommendation process. Most systems default that users have a stable and consistent acceptance of clothing sizes and fits, ignoring the direct relationship between body shape and the try-on effect, resulting in the generated recommendation results lacking persuasiveness and credibility during the try-on stage. Even if some systems allow users to select body shape types, their modeling methods are very rough, only staying at the level of text annotation or template selection, and unable to truly reflect the impact of body shape parameters on the quality of try-on image generation.

[0007] More notably, the current interaction feedback mechanism between virtual try-on and recommendation systems is not yet perfect. Behavioral data such as clicks, browsing, ratings, and collections generated by users after viewing try-on images are not effectively collected and modeled, and even less can they act in reverse on the evolution of the semantic graph structure and the reconstruction of training samples, resulting in the system lacking the ability of continuous learning and personalized evolution. Most traditional recommendation systems make static recommendations based on static models, and it is difficult to form a closed-loop structure of "recommendation - feedback - optimization - re-recommendation".

[0008] In summary, the existing virtual try-on and personalized clothing recommendation systems have significant technical defects and deficiencies in the following aspects: First, there is a lack of deep integration between the recommendation algorithm and user body shape modeling, resulting in the lack of suitability of the recommendation results for actual wearing; second, the semantic graph modeling and embedding method fail to efficiently align with users' multi-modal features (images, texts, behaviors); third, the recommendation model lacks a dynamic optimization mechanism at the structural level and cannot adaptively adjust the structure and parameter configuration according to feedback data; fourth, there is a lack of an end-to-end closed-loop feedback update mechanism, and user behavior feedback fails to be effectively used for the dynamic adjustment of the graph structure and sample strategy; fifth, the virtual try-on image generation process does not combine semantic recommendation results for joint modeling, lacking targeted and personalized rendering capabilities.

[0009] Therefore, how to provide an artificial intelligence-based virtual try-on personalized clothing recommendation method is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0010] An object of the present invention is to provide a personalized clothing recommendation method for virtual fitting based on artificial intelligence. The present invention integrates multi-modal semantic graph modeling, cross-modal contrast learning network, virtual fitting image generation technology and behavior feedback-driven optimization algorithm, and details how to realize a closed-loop process of personalized clothing recommendation and multi-angle fitting image generation based on user image data, text description, body shape parameters and interaction behaviors, with the advantages of high recommendation accuracy, real fitting effect, strong personality matching and strong system adaptive optimization ability.

[0011] A personalized clothing recommendation method for virtual fitting based on artificial intelligence according to an embodiment of the present invention includes the following steps:

[0012] S1. Collect user image data, text description data and user historical interaction behavior data to generate a multi-modal user feature set;

[0013] S2. Construct a multi-modal heterogeneous semantic graph, and perform modeling and embedding representation on the structure of the multi-modal heterogeneous semantic graph;

[0014] S3. Extract the structured embedding vectors of user nodes and clothing nodes from the multi-modal heterogeneous semantic graph, and input the structured embedding vectors and the multi-modal user feature set into a cross-modal contrast learning network model, and perform semantic alignment training in a shared embedding space by constructing positive and negative sample pairs;

[0015] S4. Optimize the structural parameters and training hyperparameters of the cross-modal contrast learning network model through a raccoon optimization algorithm to generate an optimized cross-modal contrast learning network model;

[0016] S5. Apply the optimized cross-modal contrast learning network model to the recommendation task, calculate the semantic matching degree score between the user and the clothing, generate a recommended candidate set, input the recommended candidate set and the user body shape parameters into the virtual fitting image generation unit, and combine the clothing image to generate a realistic image of the user wearing the recommended clothing, and output an interactive multi-angle fitting view;

[0017] S6. Collect user behavior feedback data on the fitting image to generate a user behavior feedback feature vector, which is used to update the edge weights in the multi-modal heterogeneous semantic graph and the composition of the training samples of the cross-modal contrast learning network. Periodically execute steps S2 to S5 to form an adaptive iterative optimization closed-loop personalized recommendation process.

[0018] Optionally, the multi-modal user feature set is generated by fusing visual feature vectors, semantic feature vectors, and behavioral feature vectors, where 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 the text description data into a language understanding unit, and the behavioral feature vectors are generated by encoding the user's historical behavior data.

[0019] Optionally, S2 specifically includes:

[0020] S21. Construct a node set, including user nodes, clothing nodes, and clothing attribute nodes. The user nodes are used to represent user objects with individual identifiers, the clothing nodes are used to represent target clothing objects that can be recommended, and the clothing attribute nodes are used to represent label information on the style, color, season, and brand of the clothing;

[0021] S22. Based on the user's historical interaction behavior data, establish an edge relationship between the user nodes and the clothing nodes. The edge relationship is used to represent the behavioral associations of clicks, collections, purchases, and try-ons between the user and the clothing;

[0022] S23. Based on the clothing metadata and label information, establish an edge relationship between the clothing nodes and the clothing attribute nodes. The edge relationship is used to represent the attribute associations of the style, color, and brand possessed by the clothing;

[0023] S24. Set an initial edge weight for the constructed edge relationship. The edge weight is set according to the user interaction frequency, behavior type, and the association strength between similar labels. The edge weight is used to adjust the recommendation path and the graph neural propagation weight;

[0024] S25. Use a heterogeneous graph modeling method to perform structural modeling on the multi-modal heterogeneous semantic graph, so that the heterogeneous relationship information is retained between various types of nodes, and at the same time, a complete graph structure representation is established;

[0025] S26. Based on the graph neural network structure, perform embedding representation on the multi-modal heterogeneous semantic graph, and encode various types of nodes in the multi-modal heterogeneous semantic graph into structured embedding vectors through a multi-layer information aggregation mechanism. The embedding vectors are used as the input of the cross-modal contrast learning network model.

[0026] Optionally, S3 specifically includes:

[0027] S31. Extract the structured embedding vectors of the user nodes and the clothing nodes from the multi-modal heterogeneous semantic graph, which are respectively represented as and , and extract the fusion representation vector from the multi-modal user feature set;

[0028] S32. Introduce the graph fusion gating coefficient , construct a non-linear gating fusion mechanism to nest and fuse the graph structure semantics and modal embeddings to generate a user representation vector :

[0029] ;

[0030] Among them, is the Sigmoid function, is the element-wise multiplication, is the multi-layer perceptron network, is the graph fusion gating coefficient;

[0031] S33. Combine the user representation vector with the clothing structure embedding vector to form a cross-modal sample pair , and construct positive sample pairs and negative sample pairs based on the historical interaction information between the user and the clothing;

[0032] S34. Introduce a semantic multi-layer contrast mechanism, set the number of semantic contrast levels , and construct contrast tasks at multiple semantic granularities, including overall matching contrast, style attribute contrast, and color semantics contrast, and calculate the independent loss for each layer;

[0033] S35. In each contrast layer, introduce a multi-factor-driven dynamic temperature control mechanism, and define the temperature parameter for the th training iteration as:

[0034] ;

[0035] Among them, is the base temperature hyperparameter, , , are the time adjustment factor, the similarity variance adjustment factor, and the loss sensitivity adjustment factor respectively, represents the variance of the similarity of the positive and negative sample pairs in the current batch, represents the contrast loss value of the current batch, is the logarithmic function;

[0036] S36. Use the cross-modal contrast learning network model to perform semantic alignment training on the user representation vector and the clothing representation vector in the shared embedding space by minimizing the weighted fusion multi-layer semantic contrast loss function.

[0037] Optionally, the specific content of S4 includes:

[0038] S41. Set the optimizable structure parameters and training hyperparameters in the cross-modal contrast learning network model to form an optimization target set, including the graph fusion gating coefficient , semantic comparison level number , and a time adjustment factor , similarity variance adjustment factor and loss sensitivity adjustment factor ;

[0039] S42. Divide the optimization target set into a structure fusion subspace and a training control subspace , and initialize two raccoon sub-populations , , and each raccoon individual represents a set of parameter combinations to be optimized;

[0040] S43. In each sub-population, execute the local memory-driven mechanism, environmental perturbation mechanism, and collaborative guidance update mechanism of the raccoon optimization algorithm, perform multi-round iterative optimization on the raccoon individuals in the population, and store the optimal raccoon individual in each round in the corresponding sub-group memory bank;

[0041] S44. Introduce a dynamic memory window mechanism, and set the current iteration memory window length to:

[0042] ;

[0043] Among them, 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 sub-group memory, is the memory window length during the optimization process of the current round, is the fitness score of the previous round, is a very small positive constant;

[0044] S45. According to the memory window length limit, retain the local optimal individuals in the memory banks of each sub-population after each round of optimization, and replace the outdated historical solutions;

[0045] S46. Introduce an attention transfer mechanism for graph semantic perception. In every round migration period, calculate the migration attention vector between populations based on the change in the edge density between user nodes and attribute nodes in the multi-modal heterogeneous semantic graph:

[0046] ;

[0047] Among them, , represent the influence weights of the sub-population on the user node and clothing node embeddings, represents the change value of the edge weight density from the user to the attribute in the graph, represents the sub-population 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:

[0048] ;

[0049] in, Subpopulation Middle The parameter representation vector of each raccoon individual, Subpopulation The parameter vector of the raccoon individual with the best fitness in the current round;

[0050] S47, select the optimal individual parameters of the current round from the two sub-populations , , merged into the global optimal parameter combination ;

[0051] S48. Define the composite fitness function for triple consistency evaluation as follows:

[0052] ;

[0053] in, Indicates that the positive sample pair is The average similarity in the layer semantic space, is the reconstruction error of the multi-angle virtual try-on images of the recommended clothing. It is a ranking quality indicator between user behavior feedback and recommendation ranking. The ranking quality indicator is calculated by comparing the actual click, favorite and purchase behaviors of users with the recommendation list, and is used to measure the degree of match between the recommendation ranking results and the user's actual preferences. , , is the balance coefficient;

[0054] S49. Sort all raccoon individuals according to the fitness function calculation results to determine the current global optimal parameter combination ;

[0055] 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.

[0056] Optionally, the S5 specifically includes:

[0057] 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;

[0058] 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;

[0059] 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;

[0060] 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;

[0061] 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;

[0062] 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.

[0063] Optionally, 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.

[0064] Optionally, the S6 specifically includes:

[0065] S61, after the user finishes browsing the recommended clothing trial images, collecting user behavior feedback data;

[0066] 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;

[0067] 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;

[0068] 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;

[0069] S65. Periodically re - execute the processes of multi - modal semantic graph modeling, cross - modal contrastive learning training, raccoon optimization parameter update and recommendation, and try - on image generation to form a dynamic self - updating recommendation iteration closed - loop;

[0070] S66. After the end of each round of closed - loop optimization cycle, adjust the semantic graph structure, matching strategy, and visual synthesis method in the recommendation process according to the change of the accumulated user behavior data, so as to improve the adaptability and feedback response ability of personalized recommendation.

[0071] Optionally, the user's behavior feedback data specifically includes interaction information such as clicks, browsing duration, image switching, ratings, and collections, which is used to dynamically adjust the semantic graph structure and the composition of training samples to optimize the personalized recommendation effect.

[0072] The beneficial effects of the present invention are as follows:

[0073] Based on the existing technology, the personalized clothing recommendation method for virtual fitting based on artificial intelligence proposed by the present invention constructs a complete closed - loop system from semantic understanding, interest matching to visual verification and feedback learning by introducing key technical components such as multi - modal heterogeneous semantic graphs, cross - modal contrastive learning network models, raccoon optimization algorithms, and virtual fitting image generation units. Compared with the traditional recommendation system that only relies on image retrieval or collaborative filtering, the present invention realizes a multi - modal end - to - end processing process from user data collection to the output of fitting images, and truly breaks through the barriers between the three stages of recommendation, fitting, and feedback.

[0074] The multi - modal heterogeneous semantic graph constructed by the present invention fully integrates user images, text descriptions, behavior records, and clothing attribute data, and performs embedding representation through a graph neural structure, effectively capturing the complex relationship between user interest preferences and clothing semantic labels. The structured embedding extracted based on this graph, after being fused with the user's multi - modal features, is sent into the cross - modal contrastive learning network, and the hierarchical semantic alignment mechanism is used to improve the matching accuracy of user - clothing pairs in the semantic space. By introducing the raccoon optimization algorithm to search and optimize the model structure and hyperparameters, the generalization ability and recommendation stability of the model are significantly enhanced, especially maintaining good results in scenarios with large personality differences and serious cold starts.

[0075] In addition, the present invention innovatively combines recommended candidate results with user body type parameters, introduces a virtual fitting image generation unit to generate interactive multi-angle fitting images, enabling users to not only "know what is recommended" but also "see the effect after wearing", greatly enhancing user participation and recommendation trust. The fitting images not only exist as display results but also interact with user behaviors such as clicks, ratings, and collections, forming quantifiable behavioral feedback features. The system dynamically updates the edge weights of the semantic graph and the structure of training samples based on this feedback information, and retrains the recommendation model periodically, thereby achieving adaptive capture of user interests and real-time personalized recommendation capabilities.

[0076] In summary, the present invention breaks through the technical bottlenecks of traditional recommendation systems in terms of personalized adaptability, visual interactivity, and self-optimization ability, achieving a comprehensive optimization effect of more accurate recommendation, more realistic fitting, more intelligent feedback, and more autonomous system, and has high practicality and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0078] Figure 1 is a flowchart of a method for personalized clothing recommendation based on artificial intelligence for virtual fitting proposed by the present invention;

[0079] Figure 2 is a flowchart of the raccoon optimization algorithm for jointly optimizing the model structure parameters and training hyperparameters of a method for personalized clothing recommendation based on artificial intelligence for virtual fitting proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0080] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0081] Refer to Figure 1 and Figure 2 , a method for personalized clothing recommendation based on artificial intelligence for virtual fitting, includes the following steps:

[0082] S1. Collect the image data, text description data of the user, and the historical interaction behavior data of the user to generate a multi-modal user feature set;

[0083] S2. Construct a multi-modal heterogeneous semantic graph, and perform modeling and embedding representation on the structure of the multi-modal heterogeneous semantic graph;

[0084] S3. Extract the structured embedding vectors of the user nodes and clothing nodes from the multi-modal heterogeneous semantic graph, and input the structured embedding vectors and the multi-modal user feature set into the cross-modal contrastive learning network model. Conduct semantic alignment training in the shared embedding space by constructing positive and negative sample pairs;

[0085] S4. Optimize 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;

[0086] S5. Apply the optimized cross-modal contrastive learning network model to the recommendation task, calculate the semantic matching degree score between the user and the clothing, generate a recommendation candidate set, input the recommendation candidate set and the user body type parameters into the virtual fitting image generation unit together, and combine with the clothing image to generate a realistic image of the user wearing the recommended clothing, and output an interactive multi-angle fitting view;

[0087] S6. Collect the behavioral feedback data of the user on the fitting image, generate a user behavioral feedback feature vector, which is used to update the edge weights in the multi-modal heterogeneous semantic graph and the composition of the training samples of the cross-modal contrastive learning network. Periodically execute steps S2 to S5 to form an adaptive iterative optimization closed-loop personalized recommendation process.

[0088] The virtual fitting personalized clothing recommendation method proposed by the present invention constructs a closed-loop recommendation system integrating data collection, multi-modal modeling, semantic alignment, intelligent optimization, visual generation and feedback learning through S1 to S6, and has significant beneficial effects. This method not only integrates multi-modal user features such as images, texts and behaviors, but also constructs a heterogeneous semantic graph to comprehensively express the multi-dimensional relationship between user interests and clothing semantics. Through the training of the cross-modal contrastive learning network model and the parameter tuning of the raccoon optimization algorithm, the semantic matching accuracy and the model generalization ability are effectively improved. The system links the recommendation results with the user body type parameters to generate high-fidelity multi-angle fitting images, realizing the direct implementation from "recommendation" to "visual dressing". More importantly, the behavioral feedback of the user on the fitting image is collected in real time and used to update the semantic graph structure and the composition of the training samples, forming an adaptive optimization closed-loop of "recommendation - fitting - feedback - re-recommendation". This method significantly improves the personalization level, recommendation credibility and user experience of clothing recommendation, and has good practicability and popularization value.

[0089] In this embodiment, the multi-modal user feature set is generated by fusing visual feature vectors, semantic feature vectors and behavioral feature vectors. Among them, 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 the text description data into a language understanding unit, and the behavioral feature vectors are generated by encoding the user's historical behavior data.

[0090] In this embodiment, S2 specifically includes:

[0091] S21. Construct a node set, including user nodes, clothing nodes, and clothing attribute nodes. The user nodes are used to represent user objects with individual identifiers, the clothing nodes are used to represent target clothing objects that can be recommended, and the clothing attribute nodes are used to represent label information on the styles, colors, seasons, and brands of clothing;

[0092] S22. Based on the user's historical interaction behavior data, establish an edge relationship between the user nodes and the clothing nodes. The edge relationship is used to represent the behavioral associations of clicks, collections, purchases, and try-ons between the user and the clothing;

[0093] S23. Based on the clothing metadata and label information, establish an edge relationship between the clothing nodes and the clothing attribute nodes. The edge relationship is used to represent the attribute associations of the styles, colors, and brands possessed by the clothing;

[0094] S24. Set an initial edge weight for the constructed edge relationship. The edge weight is set according to the association strength between the user interaction frequency, behavior type, and similar labels. The edge weight is used to adjust the recommendation path and the graph neural propagation weight;

[0095] S25. Use the heterogeneous graph modeling method to perform structural modeling on the multi-modal heterogeneous semantic graph, so that the heterogeneous relationship information is retained between various types of nodes, and at the same time, a complete graph structure representation is established;

[0096] S26. Based on the graph neural network structure, perform embedding representation on the multi-modal heterogeneous semantic graph, and encode various types of nodes in the multi-modal heterogeneous semantic graph into structured embedding vectors through a multi-layer information aggregation mechanism. The embedding vectors are used as the input of the cross-modal contrast learning network model.

[0097] The present invention systematically constructs a multi-modal heterogeneous semantic graph, which has the ability to accurately express the multi-dimensional relationship between users and clothing, and has significant beneficial effects. By introducing user nodes, clothing nodes, and clothing attribute nodes, and combining the user's historical click, favorite, purchase, and fitting behaviors, a richly expressive interaction graph structure is constructed, enabling the recommendation system to not only understand user behavior but also identify the semantic motivation behind user preferences. At the same time, by using clothing metadata and label information to construct a clothing attribute relationship network, the interpretability and classification ability of clothing semantic information are enhanced. By setting edge weights, the system can dynamically adjust the connection weights between nodes according to the interaction frequency and behavior intensity, providing data support for the propagation path and attention mechanism in the graph neural network. The introduction of the heterogeneous graph structure ensures the independence and scalability of different types of node relationships, while the embedded representation of the graph neural network enables the node semantic features to be uniformly encoded into trainable vector representations, facilitating subsequent cross-modal contrast learning. Overall, this graph modeling method significantly improves the expression ability of the recommendation system in user interest modeling and semantic association learning, providing a high-quality structural basis for subsequent improvement of recommendation effects and personalized semantic alignment.

[0098] In this embodiment, step S3 specifically includes:

[0099] S31. Extract the structured embedding vectors of user nodes and clothing nodes from the multi-modal heterogeneous semantic graph, denoted as and respectively, and extract the fused representation vector from the multi-modal user feature set;

[0100] S32. Introduce the graph fusion gating coefficient , construct a non-linear gating fusion mechanism, and perform nested fusion of the graph structure semantics and modal embeddings to generate the user representation vector :

[0101] ;

[0102] where is the Sigmoid function, is the element-wise multiplication, is the multi-layer perceptron network, is the graph fusion gating coefficient;

[0103] S33. Combine the user representation vector with the clothing structure embedding vector to form a cross-modal sample pair , and construct positive sample pairs and negative sample pairs according to the historical interaction information between the user and the clothing;

[0104] S34. Introduce a semantic multi-layer contrast mechanism and set the number of semantic contrast levels , construct contrast tasks with multiple semantic granularities, including overall matching contrast, style attribute contrast, and color semantics contrast, and calculate independent losses for each layer;

[0105] S35. In each contrast layer, introduce a multi-factor-driven dynamic temperature control mechanism, and define the temperature parameter for the th training iteration as:

[0106] ;

[0107] Among them, is the base temperature hyperparameter, , , are the time adjustment factor, similarity variance adjustment factor, and loss sensitivity adjustment factor respectively, represents the variance of the similarity between positive and negative sample pairs in the current batch, represents the contrast loss value of the current batch, is the logarithmic function;

[0108] S36. Use a cross-modal contrast learning network model to train the semantic alignment of the user representation vector and the clothing representation vector in the shared embedding space by minimizing the weighted fusion of the multi-layer semantic contrast loss function.

[0109] The present invention proposes a cross-modal contrast learning method that fuses graph structure semantics and multi-modal feature information, and has the significant beneficial effects of improving the semantic matching accuracy and model robustness of personalized recommendations. By extracting the graph structure embeddings of users and clothing and combining multi-modal user features, the system uses a gating mechanism to achieve flexible control of graph semantic enhancement, enabling the final user representation to possess both structural information and modal perception ability. By constructing positive and negative sample pairs and introducing a semantic multi-layer contrast mechanism, the system can learn the semantic similarity between users and clothing at multiple granularity levels, significantly improving the ability to capture implicit semantic levels in user preferences. At the same time, a multi-factor-driven dynamic temperature control mechanism is introduced, and the temperature parameter is dynamically adjusted in combination with the time progress, similarity distribution, and training loss, making the gradient of the contrast loss more stable and the model training process more adaptable and generalizable. Finally, the network model optimizes the representation consistency of users and clothing in the shared embedding space by weighted fusion of multi-layer semantic contrast losses, achieving a recommendation system modeling ability with richer semantic levels, more accurate recommendations, and faster convergence, providing a high-quality matching score basis for subsequent recommendation candidate generation.

[0110] In this embodiment, the specific content of S4 includes:

[0111] 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 , the number of semantic contrast levels , and the time adjustment factor , the similarity variance adjustment factor and the loss sensitivity adjustment factor ;

[0112] S42. Divide the optimization target set into a structural fusion subspace and a training control subspace based on the parameter functional attributes, and initialize two raccoon subpopulations , , where each raccoon individual represents a set of parameter combinations to be optimized;

[0113] S43. In each subpopulation, execute the local memory-driven mechanism, environmental perturbation mechanism, and collaborative guidance update mechanism of the raccoon optimization algorithm to perform multiple rounds of iterative optimization on the raccoon individuals in the population, and store the optimal raccoon individual in each round in the corresponding subgroup memory bank;

[0114] S44. Introduce a dynamic memory window mechanism, and set the current iteration memory window length as:

[0115] ;

[0116] Among them, 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 optimization process of the current round, is the fitness score of the previous round, is a very small positive constant;

[0117] S45. According to the memory window length limit, retain the local optimal individuals in the memory banks of each subpopulation after each round of optimization, and replace the outdated historical solutions;

[0118] S46. Introduce an attention transfer mechanism for graph semantic perception. In every round migration period, calculate the inter-population migration attention vector based on the change in the edge density between user nodes and attribute nodes in the multi-modal heterogeneous semantic graph:

[0119] ;

[0120] Among them, , Indicates the influence weight of the sub-population on the user node and clothing node embeddings, Indicates the change value of the edge weight density from the user to the attribute in the graph, Indicates the sub-population To the sub-population The attention weight coefficient of migration, Is the weight adjustment factor for the graph edge structure change, Is the normalization operation, and finally performs the migration operation between populations:

[0121] ;

[0122] Among them, Is the parameter representation vector of the th raccoon individual in the sub-population ; Is the parameter vector of the raccoon individual with the optimal fitness in the current round in the sub-population ;

[0123] S47. Select the optimal individual parameters , from the two sub-populations respectively, and merge them into the global optimal parameter combination ;

[0124] S48. Define the composite fitness function for triple consistency evaluation as follows:

[0125] ;

[0126] Among them, Indicates the average similarity of the positive sample pairs in the th layer semantic space, Is the reconstruction error of the multi-angle virtual try-on image of the recommended clothing, Is the sorting quality index between the user behavior feedback and the recommended sorting. The sorting quality index is calculated by comparing the actual click, favorite, and purchase behaviors of the user with the recommended list according to the discounted cumulative gain, and is used to measure the matching degree between the recommended sorting result and the user's true preference, , , Is the balance coefficient;

[0127] S49. Sort all raccoon individuals according to the calculation result of the fitness function to determine the current global optimal parameter combination ;

[0128] S410. Take the optimal parameter combination Applied to the cross-modal contrast learning network model, the graph fusion mechanism, semantic contrast hierarchy structure, and temperature regulation strategy are updated to output the finally optimized cross-modal contrast learning network model.

[0129] The present invention proposes a method for optimizing the parameters of a cross-modal contrast learning network based on an improved raccoon optimization algorithm, which focuses on integrating a dynamic memory window mechanism and a sub-population attention transfer mechanism for graph semantic perception, achieving joint and efficient optimization of structural parameters and training hyperparameters. First, the present invention divides the model parameters into two sub-spaces of structural fusion and training control, and initializes the raccoon sub-populations respectively, enabling each type of parameter to adaptively evolve in an independent space. By introducing the dynamic memory window mechanism, the system can dynamically adjust the depth of the memory bank according to the fitness fluctuations during the optimization process, thus avoiding the risk of local convergence to the early optimal solution. Further, the sub-population attention transfer mechanism for graph semantic perception guides the weighted transfer of parameter knowledge between different sub-groups based on the real-time changes in the edge weight density between users and attribute nodes, effectively enhancing the semantic relevance of the optimization direction and the cross-task generalization ability. Finally, the system jointly evaluates the semantic matching accuracy, image generation quality, and user feedback ranking metrics through a triple consistency fitness function to ensure that the selected optimal parameter combination has comprehensive performance in real recommendation scenarios. Overall, this method effectively improves the model optimization efficiency, the adaptation ability of the recommendation system, and the stability of the training process, and has high practicality and innovation.

[0130] In this embodiment, the specific steps of S5 are as follows:

[0131] S51. Deploy the optimized cross-modal contrast learning network model to the recommendation task module, input the fused feature vector of the user and the clothing embedding vector, and perform semantic similarity calculation;

[0132] S52. According to the semantic similarity calculation result, perform matching degree scoring and ranking on the candidate clothing, and select several pieces of clothing with the highest matching degree to form a recommendation candidate set;

[0133] S53. Collect the body parameter information of the user, including the body dimension features of height, weight, shoulder width, waist circumference, and hip circumference, and perform standardized modeling on the body parameter information of the user;

[0134] S54. Input the image features of each piece of clothing in the recommendation candidate set, together with the user's body parameters and fused features, into the virtual fitting image generation unit to perform the image synthesis operation of the clothing fitting effect;

[0135] S55. During the image synthesis process, set multiple viewing angles respectively to generate corresponding multi-angle realistic fitting images for each candidate clothing;

[0136] 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.

[0137] The present invention constructs a personalized recommendation execution process that integrates semantic recommendation and visual fitting, which has significant practicality and user experience improvement effects. By deploying the optimized cross-modal contrast learning network model in the recommendation task module, the system can accurately calculate the semantic similarity between the user's fusion features and the clothing embedding, and complete the screening and sorting of candidate clothing based on the matching degree, ensuring that the recommendation results are highly consistent with the user's potential interests. On this basis, the user's body shape parameter information is introduced and standardized modeling is performed, so that the recommendation not only stays at the semantic level, but also realizes the accurate characterization of the user's body shape characteristics. Subsequently, the system jointly inputs the recommendation results and the user's body shape information into the virtual fitting image generation unit to generate multi-angle wearing images with realistic effects, significantly improving the intuitive visualization experience of the recommended content. Through multi-view output and interactive display interface, users can make realistic judgments and personalized decisions on the recommended clothing at the visual level, thereby enhancing user trust and participation. Overall, this method not only improves the accuracy of recommendations, but also opens up the recommendation closed loop from "interest identification" to "wear verification", with significant user friendliness and commercial application value.

[0138] In this embodiment, 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.

[0139] In this implementation manner, S6 specifically includes:

[0140] S61, after the user finishes browsing the recommended clothing trial images, collecting user behavior feedback data;

[0141] 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;

[0142] 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;

[0143] 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;

[0144] S65. Periodically re - execute the multi - modal semantic graph modeling, cross - modal contrastive learning training, raccoon optimization parameter update and recommendation, and try - on image generation process to form a dynamically self - updating recommendation iteration loop;

[0145] S66. After the end of each round of closed - loop optimization cycle, adjust the semantic graph structure, matching strategy and visual synthesis method in the recommendation process according to the change of the accumulated user behavior data to improve the adaptability and feedback response ability of personalized recommendation.

[0146] The present invention constructs a dynamic adaptive optimization mechanism driven by user behavior feedback, significantly improving the learning ability and long - term performance of the personalized recommendation system. After the user views the try - on images of the recommended clothing, the method timely collects behavior data such as clicks, dwell time, ratings, and collections, and encodes them into behavior feedback feature vectors to characterize the change of user preferences. The system adjusts the edge weights in the multi - modal heterogeneous semantic graph based on this feedback feature to realize the structural reconstruction of user interests, and makes the connection strength between user nodes and clothing or attribute nodes evolve dynamically. At the same time, based on the updated graph structure, the system automatically updates the semantic alignment training samples to ensure that the training process continuously conforms to the user's latest preferences. By periodically re - executing the graph modeling, model training, parameter optimization and recommendation process, the system constructs a closed - loop architecture of "recommendation - feedback - optimization - re - recommendation" and has the ability of continuous learning. Finally, the system can also adjust the semantic propagation path and image synthesis strategy according to historical behavior changes to further improve the personalized adaptability and response speed. Generally speaking, this method realizes the feedback perception and system self - evolution ability in the true sense, and has the beneficial effects of long - term personality fitting and continuous improvement of recommendation accuracy.

[0147] In this embodiment, the user's behavior feedback data specifically includes interaction information such as clicks, browsing duration, image switching, ratings, and collections, which is used to dynamically adjust the semantic graph structure and the composition of training samples to optimize the personalized recommendation effect.

[0148] Example 1:

[0149] To verify the feasibility of the present invention in implementation, the present invention is applied to a large e - commerce platform. The platform selects 200 new users who have not made purchase behaviors in the past 30 days. Among them, 100 people use the existing traditional clothing recommendation system as the control group, and 100 people use the intelligent recommendation system deployed with the method of the present invention as the experimental group. The system completes the complete process from clothing recommendation to realistic image generation and then to feedback learning in an automated and personalized manner.

[0150] When users first visit the platform, users in the experimental group need to upload clear front half-body photos, fill in clothing preferences and brief body type descriptions. The platform automatically recognizes image features, combines the user's browsing history, behavior trajectory and text information to generate a multi-modal user feature vector. The system then models the semantic relationship between users and clothing based on the constructed multi-modal heterogeneous semantic graph and inputs it into the cross-modal contrastive learning network model for semantic alignment training. The model optimized by the raccoon optimization algorithm is used to generate a personalized clothing recommendation candidate set. After combining the body type parameters such as the user's height, weight, shoulder width, waist circumference, and hip circumference, the system generates multi-angle realistic try-on images for each recommended piece of clothing for users to view and try on interactively online.

[0151] Compared with the traditional system, the system of the present invention significantly improves the recommendation quality and user experience. From the experimental monitoring data, the accuracy rate of the top-3 recommendations in the experimental group reached 85.1%, an increase of about 16 percentage points compared with the traditional system; the click-through rate of the virtual try-on images reached 75.2%, much higher than 42.6% of the control group. The average browsing time of users on the try-on image page increased from 36 seconds to 59 seconds, indicating that users pay great attention to the content of the try-on images; the purchase conversion rate increased from 32.8% to 42.1%, and the body type satisfaction score also increased from 3.9 to 4.7 points. Users generally reported that the try-on effect was real and the matching degree was high. In addition, the average number of behavioral feedbacks per user in the experimental group was 22 times, which was 1.7 times that of the control group, providing sufficient data for the subsequent adaptive optimization of the system.

[0152] Table 1 Comparison table of key indicators between the system of the present invention and the traditional recommendation system

[0153]

[0154] According to the data in Table 1, it can be clearly seen that the personalized virtual try-on recommendation system proposed by the present invention is comprehensively superior to the traditional recommendation system in multiple core indicators, demonstrating significant performance advantages and improved user experience. First of all, in terms of the accuracy rate of the top-3 recommendations, the system of the present invention reached 85.1%, while the traditional recommendation system was only 69.1%, with an increase of up to 16 percentage points. This shows that the recommendation model trained by graph modeling, cross-modal contrastive learning and raccoon optimization algorithm can more accurately identify user interests and match the most suitable clothing products, significantly improving the accuracy of recommendations.

[0155] In terms of the ratio of users clicking on virtual fitting images, the click-through rate of users in the experimental group was 75.2%, while that of the control group was only 42.6%. This difference reflects that the virtual fitting images generated by the present invention have higher attractiveness and interactive value, can effectively guide users to explore the recommended results in depth, and is a direct manifestation of the authenticity and personality adaptability of the image generation unit. The average browsing duration of users on the recommended try-on page also increased from 36 seconds in the traditional system to 59 seconds, indicating that the system of the present invention can attract users' attention to the recommended content more durably, improving user engagement and system stickiness. This "immersive recommendation" experience is particularly crucial for improving conversion.

[0156] In terms of the purchase conversion rate, the system of the present invention also has obvious advantages, reaching 42.1%, which is nearly 10 percentage points higher than 32.8% of the traditional system. This means that more accurate recommendations and visual fitting effects help users make more confident purchase decisions, greatly enhancing the potential of commodity sales from a commercial perspective.

[0157] In terms of subjective experience, the average score of users' satisfaction with the body shape in the fitting images reached 4.7 points (out of 5), while it was 3.9 points in the traditional system, indicating that the dressing images generated by the method of the present invention after combining user body shape parameters are more in line with the users' real body shapes, improving users' recognition and satisfaction.

[0158] Finally, in terms of the number of behavioral feedbacks, the average number of interaction behaviors per user in the system of the present invention was 22 times, while it was only 13 times in the traditional system, proving that this system not only enhances user interaction activity, but also accumulates more high-quality training data for subsequent recommendation optimization of the system, strengthening the continuous learning ability of the model.

[0159] Based on the above analysis, the system of the present invention, by introducing multi-modal feature fusion, semantic graph modeling, intelligent optimization algorithms and visual fitting experience, not only improves the accuracy and personalization of the recommended results, but also enhances the user interaction experience and commercial conversion value, fully verifying the technical feasibility and market promotion potential of this method in real application scenarios.

[0160] The above is only the preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and all should be covered within the protection scope of the present invention.

Claims

1. A personalized clothing recommendation method for virtual fitting based on artificial intelligence, characterized in that, It includes the following steps: S1. Collect the image data, text description data of the user, and the historical interaction behavior data of the user to generate a multi-modal user feature set; S2. Construct a multi-modal heterogeneous semantic graph, and model and embed the structure of the multi-modal heterogeneous semantic graph; S3. Extract the structured embedding vectors of the user node and the clothing node from the multi-modal heterogeneous semantic graph, and input the structured embedding vectors and the multi-modal user feature set into the cross-modal contrast learning network model, and perform semantic alignment training in the shared embedding space by constructing positive and negative sample pairs; S4. Optimize the structure parameters and training hyperparameters of the cross-modal contrast learning network model through the raccoon optimization algorithm to generate an optimized cross-modal contrast learning network model; S5. Apply the optimized cross-modal contrast learning network model to the recommendation task, calculate the semantic matching degree score between the user and the clothing, generate a recommendation candidate set, input the recommendation candidate set and the user body type parameters into the virtual try-on image generation unit together, and combine the clothing image to generate a realistic image of the user wearing the recommended clothing, and output an interactive multi-angle try-on view; S6. Collect the behavioral feedback data of the user on the try-on image, generate a user behavioral feedback feature vector, which is used to update the edge weights in the multi-modal heterogeneous semantic graph and the composition of the training samples of the cross-modal contrast learning network, and periodically execute steps S2 to S5 to form an adaptive iterative optimization closed-loop personalized recommendation process; The specific content of S2 includes: S21. Construct a node set, including user nodes, clothing nodes, and clothing attribute nodes. 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 the label information of the style, color, season, and brand of the clothing; S22. Based on the user's historical interaction behavior data, establish an edge relationship between the user node and the clothing node. The edge relationship is used to represent the behavioral associations of click, favorite, purchase, and try-on between the user and the clothing; S23. Based on the clothing metadata and label information, establish an edge relationship between the clothing node and the clothing attribute node. The edge relationship is used to represent the attribute associations of the style, color, and brand possessed by the clothing; S24. Set an initial edge weight for the constructed edge relationship. The edge weight is set according to the user interaction frequency, behavior type, and the association strength between similar labels. The edge weight is used to adjust the recommendation path and the graph neural propagation weight; S25. Use the heterogeneous graph modeling method to model the structure of the multi-modal heterogeneous semantic graph, so that the heterogeneous relationship information is retained between various types of nodes, and at the same time, a complete graph structure representation is established; S26. Based on the graph neural network structure, perform an embedding representation of the multi-modal heterogeneous semantic graph, and encode various types of nodes in the multi-modal heterogeneous semantic graph into structured embedding vectors through a multi-layer information aggregation mechanism. The embedding vectors are used as the input of the cross-modal contrast learning network model.

2. The personalized clothing recommendation method for virtual fitting based on artificial intelligence according to claim 1, wherein The multimodal user feature set is generated by fusing a visual feature vector, a semantic feature vector and a behavioral feature vector, wherein the visual feature vector is extracted by inputting the user's image data into a convolutional neural network, the semantic feature vector is extracted by inputting text description data into a language understanding unit, and the behavioral feature vector is generated by encoding the user's historical behavior data.

3. The personalized clothing recommendation method for virtual fitting based on artificial intelligence according to claim 1, wherein The S3 specifically includes: S31. Extract the structured embedding vectors of the user node and the clothing node from the multimodal heterogeneous semantic graph, denoted as and , respectively, and extract the fused representation vector from the multimodal user feature set; S32. Introduce the graph fusion gating coefficient , construct a non-linear gating fusion mechanism, perform nested fusion on the graph structure semantics and modal embeddings, and generate a user representation vector : ； Among them, is the Sigmoid function, is the element-wise multiplication, is the multi-layer perceptron network, is the graph fusion gating coefficient; S33. Represent the user vector and the clothing structure embedding vector to form a cross-modal sample pair , and construct positive sample pairs and negative sample pairs according to the historical interaction information between the user and the clothing; S34. Introduce a semantic multi-level contrast mechanism and set the number of semantic contrast levels , construct contrast tasks with multiple semantic granularities, including overall matching contrast, style attribute contrast, and color semantic contrast, and calculate independent losses for each level; S35. In each comparison layer, introduce a multi-factor-driven dynamic temperature control mechanism, and define the temperature parameter for the th training iteration as: ； Among them, is the base temperature hyperparameter, are the time adjustment factor, similarity variance adjustment factor, and loss sensitivity adjustment factor respectively, represents the variance of the similarity of positive and negative sample pairs in the current batch, represents the contrast loss value of the current batch, is the logarithmic function; 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. A personalized clothing recommendation method for virtual fitting based on artificial intelligence according to claim 1, characterized in that, The S4 specifically includes: 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 , the number of semantic contrast levels , and the time adjustment factor , the similarity variance adjustment factor and the loss sensitivity adjustment factor ; S42. Divide the optimization target set into a structural fusion subspace based on the parameter functional attributes and the training control subspace , and initialize two raccoon subpopulations 、 , and each raccoon individual represents a set of parameter combinations to be optimized; 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, introduce a dynamic memory window mechanism, and set the current iteration memory window length to: ； Among them, is the initial window length, is the feedback sensitivity adjustment coefficient, is the fitness score of the current round. The window length is used to control the number of optimal solutions retained by the subgroup memory, is the memory window length during the optimization process of the current round, is the fitness score of the previous round, is a very small positive constant; S45. For each memory bank of each sub-population after each round of optimization, according to the memory window length limit, retain the locally optimal individuals and replace the outdated historical solutions; S46. Introduce an attention transfer mechanism that perceives graph semantics. Every round of the migration cycle, calculate the migration attention vector between populations based on the change in the edge density between user nodes and attribute nodes in the multi-modal heterogeneous semantic graph: ； Among them, , represent the influence weights of the sub-population on the user node and clothing node embeddings, represents the change value of the edge weight density from the user to the attribute in the graph spectrum, represents the sub-population to the sub-population the attention weight coefficient of migration, is the weight adjustment factor for the change of the edge structure of the graph spectrum, is the normalization operation, and finally the migration operation between populations is executed: ； Among them, is the parameter representation vector of the $i$-th raccoon individual in the sub-population, is the parameter vector of the raccoon individual with the optimal fitness in the current round in the sub-population; S47. Select the current round's optimal individual parameters from the two subpopulations respectively , and merge them into the global optimal parameter combination ; S48. Define a composite fitness function for triple consistency evaluation as follows: ； Among them, represents the average similarity of positive sample pairs in the layer semantic space, is the reconstruction error of the multi-angle virtual try-on image of the recommended clothing, is the sorting quality index between the user behavior feedback and the recommended sorting, is the balance coefficient; S49. Sort all raccoon individuals according to the calculation results of the fitness function to determine the current global optimal parameter combination ; S410. Apply the optimal parameter combination to the cross-modal contrastive learning network model, update the graph fusion mechanism, semantic contrast hierarchy structure, and temperature regulation strategy, and output the finally optimized cross-modal contrastive learning network model.

5. A personalized clothing recommendation method for virtual fitting based on artificial intelligence according to claim 1, characterized in that The S5 specifically includes: 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, 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, 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, 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, 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. 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. The personalized clothing recommendation method based on artificial intelligence for virtual fitting according to claim 5, characterized in that, The user's body parameter information specifically includes body dimension features 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 features.

7. A personalized clothing recommendation method for virtual fitting based on artificial intelligence according to claim 1, characterized in that The S6 specifically includes: S61, after the user finishes browsing the recommended clothing trial images, collecting user behavior feedback data; 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, 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. Re-extract the structural embedding representations of the user and the clothing nodes according to the updated multi-modal heterogeneous semantic graph structure, and use them to generate a new semantic alignment training sample composition, including the relationship update of positive sample pairs and negative sample pairs; S65. Periodically re-execute the multi-modal semantic graph modeling, cross-modal contrast learning training, raccoon optimization parameter update and recommendation, and try-on image generation processes to form a dynamically self-updating recommendation iteration closed loop; S66. After each round of closed-loop optimization cycle, adjust the semantic graph structure, matching strategy, and visual synthesis method in the recommendation process according to the cumulative changes in user behavior data to improve the adaptability and feedback response ability of personalized recommendations.

8. A personalized clothing recommendation method for virtual fitting based on artificial intelligence according to claim 7, characterized in that The user's behavior feedback data specifically includes interaction information such as clicks, browsing duration, image switching, ratings, and collections, which is used to dynamically adjust the semantic graph structure and training sample composition to optimize the personalized recommendation effect.

Images