An ai personalized dressing scheme generation method and system based on dynamic inventory matching

Abstract

The embodiment of the present disclosure discloses an AI personalized dressing scheme generation method and system based on dynamic inventory matching, which relates to the technical field of electronic commerce and artificial intelligence. The method comprises: obtaining multi-platform commodity inventory data and picture materials through a distributed crawler, and storing them in an analytical database after cleaning and feature extraction; receiving user dressing needs, combining real-time inventory status to calculate commodity inventory confidence; using an AI image enhancement model to standardize candidate commodity pictures; generating a dressing scheme based on user preferences and inventory confidence weighted matching, and supporting multimedia content synthesis; updating the inventory occupancy state according to the scheme generation result. The present disclosure establishes a real-time strong correlation between inventory data and AI generated content, introduces an inventory confidence algorithm to solve the problem of recommending out-of-stock commodities, ensures the quality of materials through pre-image enhancement, and optimizes user experience through multi-terminal adaptive layout, significantly improving the landing ability and business conversion efficiency of the dressing scheme.

Description

Technical Field

[0001] This disclosure relates to the fields of e-commerce and artificial intelligence technologies, specifically to a method and system for generating AI-powered personalized outfit suggestions based on dynamic inventory matching. This technology is particularly suitable for inventory management, personalized recommendations, and content marketing scenarios in the apparel retail industry. Background Technology

[0002] The information disclosed in this background section is intended only to enhance understanding of the overall background of this disclosure and is not necessarily an admission or implication in any way that such information constitutes related technology that is already known to those skilled in the art.

[0003] With the development of the e-commerce industry, apparel sales are gradually shifting towards personalization and content-driven approaches. Existing outfit recommendation technologies are primarily based on collaborative filtering or content recommendation algorithms, focusing on analyzing users' historical behavior or preference tags. However, existing technologies have the following drawbacks: Inventory disconnect: Recommendation systems are often independent of inventory management systems, meaning that recommended products may be out of stock or discontinued, leading to lost conversion rates. Traditional solutions often employ a "post-recommendation verification" model, which is inefficient. Low-quality materials: The quality of product images obtained by web crawlers varies, with issues such as watermarks, cluttered backgrounds, and low resolution, which directly affect the visual effect of outfit schemes and brand image; Monotonous terminal experience: The existing system has similar functions on different terminals (mobile phones, PCs, and tablets) and has not been adapted to be different for the ease of operation on mobile devices or the data processing capabilities on PCs, resulting in slow loading on mobile devices and redundant functions on PCs. Lack of closed loop: After generating outfit ideas, there is no feedback on the inventory status, making it impossible to achieve closed-loop management of "ideas as sales", and there is a lack of a model self-optimization mechanism based on user feedback.

[0004] Therefore, there is an urgent need for a technology that can deeply integrate real-time inventory, image quality enhancement, and multi-terminal adaptation to generate outfit solutions in order to solve the above problems. Summary of the Invention

[0005] Therefore, the purpose of this disclosure is to provide an AI-based personalized outfit matching solution generation method and system based on dynamic inventory matching, in order to solve the problems of disconnect between recommendations and inventory, uncontrollable material quality, and poor terminal compatibility in the existing technology.

[0006] To achieve the above objectives, this disclosure provides the following technical solution: A method for generating AI-based personalized outfit suggestions based on dynamic inventory matching, the core of which lies in establishing a real-time strong correlation between inventory data and AI-generated content. By calculating inventory confidence, the marketability of recommended products is ensured; by enhancing front-view images, the usability of the materials is ensured; and by adapting to multiple terminals, the convenience of interaction is ensured.

[0007] To achieve the above objectives, the embodiments of this disclosure provide the following technical solutions: In a first aspect of the embodiments of this disclosure, an AI-based personalized outfit matching method is provided, comprising the following steps: S1: Through the distributed data acquisition module, basic product information and related image materials are crawled from multiple preset clothing platforms. After data cleaning and deduplication, the data is stored in the analytical database. S2: Receive the outfit generation request sent by the user terminal, and parse the style constraints and user identity identifier in the request; S3: Retrieve a set of candidate products that meet the style constraints from the analytical database and obtain the real-time inventory status of each candidate product; S4: Based on the real-time inventory status, calculate the inventory confidence score of each candidate product, and combine it with image feature similarity to select the target product combination; S5: Utilizes the AI ​​image processing engine to enhance the image materials corresponding to the target product combination, generating standardized outfit materials; S6: Based on standardized outfit materials, synthesize personalized outfit schemes, write the scheme storage address and associated style number set into the scheme database, and trigger an inventory status update command at the same time.

[0008] In one embodiment of this disclosure, the data cleaning and deduplication operation in step S1 specifically includes: Use regular expressions to extract product item numbers as unique identifiers from crawled data; The image materials are subjected to resolution detection, and images with a resolution lower than the preset threshold of 720P are discarded, while high-definition main images and detail images are retained; Using the product style number as an index, compare it with existing records in the database. If the style number exists, update the inventory quantity and shelf status. If the style number does not exist, add a new record and record the source of the crawling.

[0009] In one embodiment of this disclosure, the formula for calculating the inventory confidence score in step S4 is: C inv =α*log(1+N stock )+β*e -λ*ΔT Among them, C inv N represents the inventory confidence score. stock ΔT represents the current inventory quantity, ΔT represents the time interval since the last inventory synchronization, α and β are weighting coefficients and α+β=1, and λ is the time decay factor; The specific process for selecting the target product combination involves calculating the cosine similarity S between the user preference vector (Vuser) and the product feature vector (Vitem). pref Calculate the final matching score S total =ω1*S pref +ω2*C inv Where ω1 is the user preference weight; ω2 is the inventory confidence weight; C inv Assign a confidence score to the inventory. Select S total The top-ranked products form a combination.

[0010] In one embodiment of this disclosure, the enhancement process in step S5 includes at least one of the following: Background removal processing based on semantic segmentation technology uses a deep learning model to identify the foreground clothing area and replace the background pixels with a preset solid color background. Facial redrawing based on generative adversarial networks can repair facial occlusion or optimize expressions in images containing models, while keeping the clothing texture unchanged. Watermark removal based on texture analysis identifies high-frequency noise watermark regions in images and performs pixel repair.

[0011] In one embodiment of this disclosure, a multi-terminal adaptation step is further included after step S6: Monitor the device type identifier of the user terminal and parse the user agent string; If the device type is mobile, a compressed preview file will be generated, and the interface will use a single-column flow layout, limiting complex editing functions; If the device type is web-based or tablet-based, the original resolution file will be generated. The interface adopts a multi-column layout and provides batch processing and data statistics functions. For tablets, the screen orientation sensor signal is detected, supporting automatic switching between portrait and landscape layouts.

[0012] In one embodiment of this disclosure, parsing the user identity identifier in step S2 specifically includes: Obtain user authorization identifier through third-party instant messaging platform API; The authorization identifier is compared with the user information table to determine the user's permission level; If the access level is ordinary user, then the search scope in step S3 is limited to publicly available products, and the export permission for the scheme in step S6 is also limited. If the permission level is administrator, then full data retrieval and permission management functions are enabled.

[0013] In one embodiment of this disclosure, step S1 further includes an inventory synchronization mechanism: It supports manual synchronization and scheduled automatic synchronization. Scheduled synchronization can be set to perform a full synchronization at a fixed time every day at midnight. During synchronization, compare existing data and only update change fields such as inventory quantity and shelf status, generating operation logs; If synchronization fails, an error log is recorded and an alarm notification is triggered, supporting breakpoint resumption.

[0014] In a second aspect of the embodiments of this disclosure, an AI-based personalized outfit matching system is provided for generating outfit solutions based on dynamic inventory matching, for implementing the method described in any of the above claims, comprising: The data acquisition layer is configured with distributed crawler nodes to acquire product data and images from multiple platforms and execute cleaning logic. The data storage layer includes an analytical database used to store product information tables, image material tables, user information tables, outfit scheme tables, and operation log tables. The AI ​​engine layer integrates image feature extraction models, image enhancement models, and multimedia generation models; The business logic layer is used to perform inventory confidence calculation, product matching and filtering, and inventory status synchronization. The interactive display layer provides a multi-terminal compatible interface for receiving user requests and displaying outfit ideas.

[0015] In one embodiment of this disclosure, the specific structure of the data storage layer includes: The product information table includes fields for style number primary key, brand, style tag, size range, and inventory quantity. The image resource table contains an image ID primary key, a related product number foreign key, an image type, and a storage path field. It establishes a one-to-many relationship with the product information table through the related product number foreign key. The outfit scheme table includes a scheme ID primary key, a creator ID foreign key, and a set of associated style numbers field. It is linked to the user information table through the creator ID foreign key. The analytical database is pre-partitioned to store basic product information, image materials, and user operation logs, and indexes are created to accelerate queries.

[0016] In one embodiment of this disclosure, the system further includes a feedback optimization module: Collect user behavior data related to generated outfit suggestions, including clicks, favorites, shares, and conversions to purchases; The weighting coefficients ω1 and ω2 in step S4 are adjusted based on the operational behavior data; If the user conversion rate is lower than the preset threshold, the inventory confidence weight will be automatically reduced, the image feature similarity weight will be increased, and the image feature extraction model will be retrained.

[0017] According to the embodiments of this disclosure, it has the following advantages: High feasibility: Introducing an inventory confidence algorithm to avoid recommending out-of-stock items and improve solution conversion rates.

[0018] High-quality content: Automated image cleaning and enhancement to unify visual style and improve brand image.

[0019] Flexible adaptation: Dynamically adjust functions and layout according to terminal devices to optimize user experience.

[0020] Data closed loop: The generation of solutions is linked with inventory updates to achieve refined operations and support model optimization based on feedback. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments or related technologies of this disclosure, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0022] Figure 1 is a schematic diagram of the main flow of the method provided in the embodiment of this disclosure; Figure 2 is a schematic diagram of the system layered architecture provided in an embodiment of this disclosure; Figure 3 is a schematic diagram of the inventory confidence calculation and matching logic in an embodiment of this disclosure; Figure 4 is a schematic diagram of the multi-terminal interface adaptation layout in an embodiment of this disclosure; Figure 5 is a schematic diagram comparing the effects of image enhancement processing before and after in an embodiment of this disclosure; Figure 6 is a schematic diagram of the workflow of the feedback optimization module in an embodiment of this disclosure; Figure 7 is a logical framework diagram of the solution generation application in the system provided in the embodiments of this disclosure. Detailed Implementation

[0023] The following specific embodiments illustrate the implementation of this disclosure. Those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of this disclosure. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure. Example 1: System Architecture and Core Theoretical Foundations

[0024] As shown in Figure 1, an AI-based personalized outfit matching method based on dynamic inventory matching is provided, including the following steps: Step S101 (Data Preparation): The web crawler acquires data, cleans it, and then stores it in the database. Through the distributed data acquisition module, basic product information and related image materials are crawled from multiple preset clothing platforms. After performing data cleaning and deduplication operations, the data is stored in the analytical database.

[0025] Step S102 (Request Reception): The user logs in and initiates an outfit request. The system receives the outfit generation request sent by the user's terminal and parses the style constraints and user identity identifier in the request.

[0026] Step S103 (Candidate Search): Search for products based on style tags, retrieve a set of candidate products that meet the style constraints from the analytical database, and obtain the real-time inventory status of each candidate product.

[0027] Step S104 (Confidence Calculation): Calculate inventory score, filter products, calculate inventory confidence score for each candidate product based on real-time inventory status, and filter target product combination by combining image feature similarity.

[0028] Step S105 (Image Enhancement): Process the image materials by calling the AI ​​image processing engine to enhance the image materials corresponding to the target product combination and generate standardized outfit materials.

[0029] Step S106 (Scheme Synthesis): Generate the final scheme and update the inventory. Based on standardized outfit materials, synthesize personalized outfit schemes and write the scheme storage address and associated style number set into the scheme database. At the same time, trigger the inventory status update instruction.

[0030] The core of this method lies in establishing a strong, real-time correlation between inventory data and AI-generated content. It ensures the marketability of recommended products by calculating inventory confidence levels; ensures the usability of content through pre-image enhancement; and ensures ease of interaction through multi-terminal adaptation.

[0031] As shown in Figure 2, this system adopts a layered architecture design. Based on the microservice theory, the system is decoupled into five core layers to reduce coupling and improve maintainability. The connection relationship between the data acquisition layer, data storage layer, AI engine layer, business logic layer and interactive display layer is shown.

[0032] Data Acquisition Layer: Theoretical basis: Based on distributed crawler theory, multi-node concurrent data collection is adopted to improve efficiency.

[0033] Implementation details: Deploy 4 distributed crawler nodes, each corresponding to a major apparel platform. Configure a crawling frequency of once every 2 hours. This layer is responsible for acquiring and initially cleaning the raw data. Utilize the HTTP protocol to simulate browser requests, bypassing basic anti-crawling mechanisms.

[0034] Data storage layer: Theoretical basis: Based on OLAP (Online Analytical Processing) theory, it adopts an analytical database (ADS) to support rapid querying and analysis of massive amounts of data.

[0035] Implementation details: Five core tables are pre-defined (product, image, user, plan, log). Data is partitioned for storage, such as by time or brand, to ensure query efficiency. Indexing is implemented to accelerate style number retrieval.

[0036] AI Engine Layer: Theoretical basis: Based on deep learning theory, it uses convolutional neural networks (CNN) to extract features and generative adversarial networks (GAN) to generate content.

[0037] Implementation details: Integration with the PyTorch deep learning framework. Includes an image feature extraction model (ResNet architecture), an image enhancement model (U-Net for segmentation, StyleGAN for redrawing), and a multimedia generation model. Configuring GPU computing power (≥8GB VRAM) to support matrix operation acceleration.

[0038] Business logic layer: Theoretical basis: Based on rule engine and algorithm weighting theory.

[0039] Implementation details: The core computing unit is responsible for executing the inventory confidence algorithm, handling permission verification (interfacing with DingTalk API), and inventory synchronization logic.

[0040] Interactive presentation layer: Theoretical basis: Based on Responsive Web Design (RWD) theory.

[0041] Implementation details: Based on the Vue3 + Element Plus framework, the User-Agent is identified to determine the terminal type and render different interfaces accordingly. Example 2: Inventory Confidence Algorithm and Matching Logic

[0042] This embodiment details the algorithm logic of inventory confidence, illustrated in Figure 3, which shows the input-output relationship of each variable in the formula and the weighted matching process.

[0043] In existing technical solutions, typically only the inventory N is determined. stock >0. This disclosure introduces a time decay factor to prevent "overselling" recommendations caused by data synchronization delays. The formula is as follows: C inv =α*log(1+N stock )+β*e -λ*ΔT Theoretical explanation and variable definition: C inv (Inventory Confidence): Indicates the degree of confidence that the product can be recommended at the current moment, with values ​​normalized to [0,1].

[0044] N stock (Inventory Quantity): The number of inventory items currently recorded in the database.

[0045] Logarithmic function log(1+N) stock The purpose of this is to smooth out the impact of inventory quantity by utilizing the growth characteristics of the logarithmic function. For example, the difference in sellability between 100 and 1000 units in inventory should not be too large, avoiding the dominance of maximum inventory values ​​in the scoring, and making the algorithm focus more on "availability" rather than "quantity".

[0046] ΔT (time interval): The time since the last inventory synchronization (in hours).

[0047] Exponential decay e -λ*ΔT Its purpose is to reflect the freshness of data. Data reliability decreases exponentially over time. If ΔT exceeds 24 hours, this score approaches 0, indicating to the system that the data may be outdated.

[0048] α, β (weighting coefficients): satisfy α + β = 1. In actual deployment, they can be set to α = 0.6 and β = 0.4, focusing on inventory quantity while also considering timeliness.

[0049] λ (attenuation factor): Set to 0.1 to control the attenuation rate.

[0050] Matching logic: Final matching score S total Combined with user preferences (Spref ) and inventory confidence level (C inv ).

[0051] S total =ω1*S pref +ω2*C inv Spref (Preference Similarity): Calculated by measuring the cosine similarity between a user's historical preference vector and a product feature vector. Cosine similarity assesses similarity by measuring the cosine of the angle between the two vectors in multidimensional space; a value closer to 1 indicates greater similarity. ω1, ω2 matching weights): initially set to 0.5:0.5, which can be dynamically adjusted based on the feedback in Example 10.

[0052] This logic ensures that the generated solution not only meets the user's aesthetic preferences but also has a high degree of purchase feasibility, thus solving the inventory disconnect problem mentioned in the background technology. Example 3: Details of Image Enhancement Processing

[0053] In this embodiment, an AI image processing engine is invoked to enhance the image materials corresponding to the target product combination, generating standardized outfit materials. The enhancement process includes at least one of the following: Background removal processing based on semantic segmentation technology uses a deep learning model to identify the foreground clothing area and replace the background pixels with a preset solid color background. Facial redrawing based on generative adversarial networks can repair facial occlusion or optimize expressions in images containing models, while keeping the clothing texture unchanged. The watermark removal process based on texture analysis identifies high-frequency noise watermark regions in the image and performs pixel-level repair. This is illustrated in Figure 5. Before generating outfit schemes, the system automatically performs pipeline processing on the selected product images, based on image segmentation and generation techniques from computer vision. The results of the original image, background removal, watermark removal, and facial redrawing are shown.

[0054] White background generation (semantic segmentation): Technical principle: Using semantic segmentation models (such as DeepLabV3+ or U-Net), each pixel in the image is classified as either "foreground (clothing)" or "background".

[0055] Implementation: The model outputs a binary mask, replacing background pixels with pure white (RGB: 255,255,255) to eliminate environmental interference and unify the visual style.

[0056] Watermark Removal (Image Inpainting): Technical principle: Based on texture analysis algorithms, high-frequency noise regions (usually watermarks) in images are detected. The watermark regions are then filled using partial differential equations or deep learning models, utilizing the texture information of surrounding pixels.

[0057] Implementation: Identify the watermark location, generate repaired pixels, and ensure the image is clean.

[0058] Facial redrawing (Generative Adversarial Network GAN): Technical principle: The generator of GAN is used to generate realistic human faces, and the discriminator distinguishes between real and fake faces.

[0059] Implementation: If the image contains a model, call the pre-trained GAN model to optimize the lighting and shadows on the model's face or occlude privacy information while protecting clothing details (by locking the clothing area through an attention mechanism). This step ensures that all materials entering the "Outfit Scheme Table" are standardized high-definition images, solving the problem of inconsistent material quality mentioned in Example 1. Example 4: Multi-terminal adaptive interaction

[0060] This embodiment implements the terminal adaptation and interaction process based on the above system, including the following multi-terminal adaptation steps: Monitor the device type identifier of the user terminal and parse the user agent string; If the device type is mobile, a compressed preview file will be generated, and the interface will use a single-column flow layout, limiting complex editing functions; If the device type is web-based or tablet-based, the original resolution file will be generated. The interface adopts a multi-column layout and provides batch processing and data statistics functions. For tablets, the system detects the screen orientation sensor signal and supports automatic layout switching between portrait and landscape modes. This is illustrated in Figure 4. The system identifies the device by parsing the User-Agent in the HTTP request header and implements adaptation based on responsive design principles. The interface differences between mobile devices, web browsers, and tablets are compared.

[0061] Mobile version: Layout: It adopts a single column flow layout, which conforms to the finger swiping habits.

[0062] Features: Hides complex statistical charts and enlarges the "Upload Image" and "Solution Preview" buttons. Supports gesture-based switching between schemes.

[0063] Performance: Image loading employs lazy loading and compression strategies (WebP format) to reduce bandwidth consumption.

[0064] Web version: Layout: A multi-column layout is used. The left side is the filter bar, the middle is the solution flow, and the right side is the data statistics.

[0065] Features: Supports batch export of Excel inventory data and displays inventory trend charts (based on the ECharts library).

[0066] Pad end: Layout: Detects screen orientation sensor.

[0067] Vertical screen: Similar to the layout of mobile devices.

[0068] Landscape mode: Automatically switches to split-screen mode (product library on the left, scheme editing area on the right), and optimizes the size of the touch hotspot (minimum click area 44x44 pixels).

[0069] In addition, if a weak network environment is detected (judged by API response time), the resolution of the AI-generated video will be automatically reduced to prioritize smooth interaction. Example 5: Database Structure and Synchronization Mechanism

[0070] The method in this embodiment uses a distributed data acquisition module to crawl basic product information and related image materials from multiple preset clothing platforms. After performing data cleaning and deduplication operations, the data is stored in an analytical database, and a data synchronization mechanism is included to supplement the database design details.

[0071] Table structure design: Product Information Table: Style Number (PK), Brand, Style Tag, Size Range, Inventory Quantity, Listing Status, Source of Data Crawling. The style number is used as the primary key to ensure uniqueness.

[0072] Image resource table: Image ID (PK), Associated Style Number (FK), Image Type, Storage Path. A one-to-many relationship is established between the style number and the product information table via foreign keys, supporting multiple styles and multiple images.

[0073] Index optimization: Create a composite index on the style tag and inventory quantity fields to speed up the retrieval process in step S3.

[0074] Synchronization mechanism: Full synchronization: Every day at 2 AM, during the system's idle period, all styles are traversed to update inventory.

[0075] Incremental synchronization: Supports manual triggering, comparing only changed fields.

[0076] Resume interrupted data crawling: If the crawler is interrupted during the crawling process, the last processed item number (Checkpoint) is recorded. The crawler will resume from the breakpoint the next time it starts, avoiding duplicate data writing and ensuring data consistency (ACID principle). Example 6: Access Control and Security

[0077] This embodiment receives an outfit generation request sent by a user terminal and parses the style constraints and user identification in the request. The method for user identification includes: Obtain user authorization identifier through third-party instant messaging platform API; The authorization identifier is compared with the user information table to determine the user's permission level; If the access level is ordinary user, then the search scope in step S3 is limited to publicly available products, and the export permission for the scheme in step S6 is also limited. If the permission level is administrator, then full data retrieval and permission management functions are enabled.

[0078] Identity verification: Integrates with DingTalk Open Platform OAuth2.0 protocol. User clicks login -> redirects to authorization -> returns code -> obtains User Info.

[0079] Access Control (RBAC model): Regular users: can only see the schemes they created, and cannot export the full data.

[0080] Administrator: Has full access to data and has the "permission assignment" function.

[0081] Log auditing: All operations are written to the "Operation Log Table", which includes user ID, operation type, IP address, and timestamp, for security auditing and fault location. Example 7: Load Balancing and Elastic Scaling

[0082] An AI-powered personalized outfit matching system based on dynamic inventory matching, used to implement the aforementioned method, includes: The data acquisition layer is configured with distributed crawler nodes to acquire product data and images from multiple platforms and execute cleaning logic. The data storage layer includes an analytical database used to store product information tables, image material tables, user information tables, outfit scheme tables, and operation log tables. The AI ​​engine layer integrates image feature extraction models, image enhancement models, and multimedia generation models; The business logic layer is used to perform inventory confidence calculation, product matching and filtering, and inventory status synchronization. The interactive display layer provides a multi-terminal compatible interface for receiving user requests and displaying outfit ideas.

[0083] This system includes a configuration mechanism layer for load balancing and elastic scaling. Specifically, it includes: SLB Configuration: Configure Layer 4 TCP protocol forwarding and set the maximum number of connections threshold to 1000.

[0084] Elastic Scaling: Monitors cloud server CPU utilization. If it exceeds 80% for 5 consecutive minutes, it automatically triggers Elastic Compute Service (ESC) scaling to add instances.

[0085] Objective: To ensure that the search response time is ≤1 second and the image processing time is ≤3 seconds under high concurrency, thus meeting the performance testing requirements. Example 8: Video Generation and Multimedia Composition

[0086] This embodiment extends the method disclosed herein by implementing a method for video generation and multimedia synthesis after a personalized outfit plan is generated.

[0087] Image-to-video: Input 3-5 standardized outfit images, and the AI ​​model automatically calculates the interpolation frames between the images to generate a continuous video of 15-60 seconds.

[0088] Audio synchronization: Automatically matches tracks from the background music library to the video rhythm and adjusts the volume balance.

[0089] Subtitle addition: OCR recognizes product style numbers and prices, automatically generates floating subtitles, and the font size adapts to the screen resolution.

[0090] Preview mechanism: Generate a low-bitrate temporary file for users to preview quickly. After confirming that there are no errors, convert it into a high-definition file and store it on the cloud storage server (OSS), and update the storage address in the "Outfit Scheme Table". Example 9: Feedback Optimization Closed Loop

[0091] This embodiment focuses on the system feedback mechanism, collecting user behavior data related to the generated outfit schemes, including clicks, favorites, shares, and purchase conversions. Based on this behavior data, the weight coefficients ω1 and ω2 in step S4 are adjusted. If the user conversion rate is lower than a preset threshold, the inventory confidence weight is automatically reduced, the image feature similarity weight is increased, and the image feature extraction model is retrained, as illustrated in Figure 6. This demonstrates how user behavior data adjusts the model weights in reverse. To achieve system self-evolution, a feedback optimization module is introduced.

[0092] Data collection: Collect user behavior data on generated outfit suggestions, including clicks, favorites, shares, and purchases.

[0093] Weight adjustment: If the user conversion rate (purchase / generation) is lower than the preset threshold (e.g., 5%), the system will automatically analyze the reasons.

[0094] If the issue is related to inventory (user clicks but item is out of stock), then reduce the inventory confidence weight ω2 and increase the image feature similarity weight ω1, prioritizing products with high visual matching, and then confirming inventory through other methods.

[0095] If the problem is visual, the image feature extraction model is retrained, and the proportion of negative samples (images that the user has not clicked) is increased during training.

[0096] Theoretical basis: Based on the reward mechanism in reinforcement learning, user behavior is used as a reward signal to optimize the policy network. Example 10: This is illustrated in Figure 7. It shows a schematic diagram of the logical framework for a practical application. Users access the interaction design layer through multiple terminals. After initiating an operation request, the interaction design layer standardizes and transmits the request to the business function layer. The business function layer calls upon the resources (data, server computing power) of the data support layer to complete the core processing. The operations management layer monitors the operational status of each layer throughout the process to ensure smooth workflow. Finally, the processing result is fed back to the user through the interaction design layer, forming a closed loop of "request-processing-feedback".

[0097] Specific details of interlayer coordination (1) User layer ↔ Interaction design layer • Coordination Logic: When users select mobile, web, or tablet based on their usage scenario, the interaction design layer automatically adapts to the screen size and operation method (touchscreen / mouse) to present a unified style of functional interface; when users input operation commands (such as clicking "search by image" or uploading an image), the interaction design layer converts the commands into standardized interface requests (such as JSON format) and passes them to the business function layer, while simultaneously receiving the results returned by the business function layer (such as a list of matched products or processed images) and displaying them to the user in a visual way.

[0098] Example: A user uploads an outfit picture on their mobile phone. The interaction design layer adapts to the touch screen upload logic, compresses the image to a suitable size (without affecting recognition accuracy), generates a request containing "image data, operation type (image search), and user ID", and passes it to the business function layer. After processing, the matching inventory product images, style numbers, and inventory status are displayed on the mobile phone screen in a card layout.

[0099] (2) Interaction design layer ↔ Business function layer • Coordination Logic: The request passed by the interaction design layer carries "operation type, user ID, and core data (such as image, product number)". The business function layer routes the request to the corresponding module based on the operation type (e.g., "image search" is routed to the data search module, "face redraw" is routed to the image processing module). After the module completes its processing, it returns standardized results (e.g., product list, processed media file, operation status). The business function layer encapsulates the results (e.g., adds permission verification results, data format conversion) and passes them to the interaction design layer. If the user needs to supplement information during the processing (e.g., triggering a login request when not logged in), the business function layer returns an instruction "information needs to be supplemented", and the interaction design layer guides the user to complete the corresponding operation (e.g., redirecting to DingTalk login).

[0100] Example: When a user initiates a "face redraw" operation, the interaction design layer passes a request containing "image data, user ID, and operation type (face redraw)". The business function layer routes the request to the image processing module, calls the AI ​​model to process the image, and returns "the address of the redrawn image and the operation success status" after processing. The business function layer converts the address into a preview link that the interaction design layer can recognize and passes it to the front-end interface, where the user can directly view or download it.

[0101] (3) Business Function Layer ↔ Data Support Layer • Coordination Logic: When the business function layer executes core operations, it initiates data query, write, and update requests to the data support layer (e.g., the data search module needs to query product information, and the material mapping management module needs to associate images with style numbers); the data support layer responds to the requests through database queries and server computing power scheduling (e.g., AI model operation relies on ESC GPU resources), returning the required data or execution results; the user management module of the data support layer verifies the identity of the "user ID" passed by the business function layer (associating it with DingTalk account information), returns the "permission level," and the business function layer accesses the data according to the permission control function (e.g., ordinary users cannot modify inventory data, while administrators can update it in batches).

[0102] Example: After receiving an "image search" request, the data search module sends a request to the data support layer to "extract image feature values ​​and query similar feature data in the image material table". The data support layer calls the index function of the ADS database (an inverted index built based on image feature values) to quickly match the top 10 similarity product data. At the same time, it queries the product information table to obtain the corresponding style number and inventory status, and returns it to the data search module. The module integrates the data and then passes it to the interaction design layer.

[0103] (4) Data support layer ↔ Operations management layer • Coordination Logic: The operations management layer obtains the real-time operational status of the data support layer (such as server load rate, database connection count, and crawler success rate) through monitoring interfaces. If an anomaly occurs (such as excessive load or crawling failure), the operations management layer triggers an early warning mechanism (such as notifying technical personnel via SMS) and performs emergency handling (such as initiating ESC for elastic scaling or adjusting crawler rules). The operations management layer initiates data maintenance commands (such as batch updating inventory data or optimizing database indexes), and the data support layer executes the corresponding operations and returns the execution results (such as the number of updated records or the optimization completion status). The data support layer periodically synchronizes log data (such as user operation logs and fault logs) to the operations management layer. The operations management layer performs data analysis based on the logs (such as frequently used functions and high-incidence fault points) to provide a basis for iterative optimization.

[0104] Example: The operations management team monitors that the ESC server load rate exceeds 80% during a certain period, triggering an elastic scaling command. After receiving the command, the data support layer automatically adds one ESC instance. The SLB load balancing module adjusts the traffic distribution rules, forwarding some requests to the new instance to reduce the load on the original instance. At the same time, it returns feedback to the operations management team that "scaling was successful and the current load rate is 60%".

[0105] Operation and maintenance support and iteration This embodiment provides a guarantee for the long-term operation of the supplementary system.

[0106] Monitoring: 7x12 hours monitoring of system operation status (server load, database connection count, interface response time).

[0107] Rule updates: The crawler rules are updated quarterly to adapt to changes in the data format of the target platform (such as changes in HTML structure).

[0108] Model optimization: AI model parameters are optimized every six months, and the accuracy of functions is improved based on user feedback on image processing effects and solution generation satisfaction.

[0109] The methods described above can be implemented using software and necessary general-purpose hardware platforms. While hardware can also be used, the former is often a better implementation method. Based on this understanding, the technical solutions of this disclosure, or the parts that contribute to the related technology, can be embodied in the form of software products. The limitations of the hardware structure platform should not be construed as limiting the implementation of the methods disclosed herein. Therefore, all embodiments of this method are applicable to the electronic device and storage medium, and can achieve the same or similar beneficial effects.

[0110] Although the present disclosure has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, such modifications or improvements made without departing from the spirit of the present disclosure are all within the scope of protection claimed by the present disclosure.

Claims

1. A method for generating AI-powered personalized outfit matching solutions based on dynamic inventory matching, characterized in that, Includes the following steps: S1: Through the distributed data acquisition module, basic product information and related image materials are crawled from multiple preset clothing platforms. After data cleaning and deduplication, the data is stored in the analytical database. S2: Receive the outfit generation request sent by the user terminal, and parse the style constraints and user identity identifier in the request; S3: Retrieve a set of candidate products that meet the style constraints from the analytical database and obtain the real-time inventory status of each candidate product; S4: Based on the real-time inventory status, calculate the inventory confidence score of each candidate product, and combine it with image feature similarity to select the target product combination; S5: Utilizes the AI ​​image processing engine to enhance the image materials corresponding to the target product combination, generating standardized outfit materials; S6: Based on standardized outfit materials, synthesize personalized outfit schemes, write the scheme storage address and associated style number set into the scheme database, and trigger an inventory status update command at the same time.

2. The method according to claim 1, characterized in that, The data cleaning and deduplication operations in step S1 specifically include: Use regular expressions to extract product item numbers as unique identifiers from crawled data; The image materials are subjected to resolution detection, and images with a resolution lower than the preset threshold of 720P are discarded, while high-definition main images and detail images are retained; Using the product style number as an index, compare it with existing records in the database. If the style number exists, update the inventory quantity and shelf status. If the style number does not exist, add a new record and record the source of the crawling.

3. The method according to claim 2, characterized in that, The formula for calculating the inventory confidence score in step S4 is as follows: C inv =α*log(1+N stock )+β*e -λ*ΔT Among them, C inv N represents the inventory confidence score. stock ΔT represents the current inventory quantity, ΔT represents the time interval since the last inventory synchronization, α and β are weighting coefficients and α+β=1, and λ is the time decay factor; The specific steps for selecting the target product combination are: calculating the cosine similarity S between the user preference vector and the product feature vector. pref Calculate the final matching score S total =ω1*S pref +ω2*C inv Where ω1 is the user preference weight; ω2 is the inventory confidence weight; C inv Assign a confidence score to the inventory. Select S total The top-ranked products form a combination.

4. The method according to claim 3, characterized in that, The enhancement process in step S5 includes at least one of the following: Background removal processing based on semantic segmentation technology uses a deep learning model to identify the foreground clothing area and replace the background pixels with a preset solid color background. Facial redrawing based on generative adversarial networks can repair facial occlusion or optimize expressions in images containing models, while keeping the clothing texture unchanged. Watermark removal based on texture analysis identifies high-frequency noise watermark regions in images and performs pixel repair.

5. The method according to claim 1, characterized in that, Step S6 is followed by a multi-terminal adaptation step: Monitor the device type identifier of the user terminal and parse the user agent string; If the device type is mobile, a compressed preview file will be generated, and the interface will use a single-column flow layout, limiting complex editing functions; If the device type is web-based or tablet-based, the original resolution file will be generated. The interface adopts a multi-column layout and provides batch processing and data statistics functions. For tablets, the screen orientation sensor signal is detected, supporting automatic switching between portrait and landscape layouts.

6. The method according to claim 1, characterized in that, The specific steps in step S2, such as parsing the user identity identifier, include: Obtain user authorization identifier through third-party instant messaging platform API; The authorization identifier is compared with the user information table to determine the user's permission level; If the access level is ordinary user, then the search scope in step S3 is limited to public products, and the scheme export permission in step S6 is also limited. If the permission level is administrator, then full data retrieval and permission management functions are enabled.

7. The method according to claim 2, characterized in that, Step S1 also includes an inventory synchronization mechanism: It supports manual synchronization and scheduled automatic synchronization. Scheduled synchronization can be set to perform a full synchronization at a fixed time every day at midnight. During synchronization, compare existing data and only update change fields such as inventory quantity and shelf status, generating operation logs; If synchronization fails, an error log is recorded and an alarm notification is triggered, supporting breakpoint resumption.

8. An AI-based personalized outfit matching system for generating outfit solutions based on dynamic inventory matching, used to implement the method described in any one of claims 1-7, characterized in that, include: The data acquisition layer is configured with distributed crawler nodes to acquire product data and images from multiple platforms and execute cleaning logic. The data storage layer includes an analytical database used to store product information tables, image material tables, user information tables, outfit scheme tables, and operation log tables. The AI ​​engine layer integrates image feature extraction models, image enhancement models, and multimedia generation models; The business logic layer is used to perform inventory confidence calculation, product matching and filtering, and inventory status synchronization. The interactive display layer provides a multi-terminal compatible interface for receiving user requests and displaying outfit ideas.

9. The system according to claim 8, characterized in that, The specific structure of the data storage layer includes: The product information table includes fields for style number primary key, brand, style tag, size range, and inventory quantity. The image resource table contains an image ID primary key, a related product number foreign key, an image type, and a storage path field. It establishes a one-to-many relationship with the product information table through the related product number foreign key. The outfit scheme table includes a scheme ID primary key, a creator ID foreign key, and a set of related style numbers field. It is linked to the user information table through the creator ID foreign key. The analytical database is pre-partitioned to store basic product information, image materials, and user operation logs, and indexes are created to accelerate queries.

10. The system according to claim 8, characterized in that, The system also includes a feedback optimization module: Collect user behavior data related to generated outfit suggestions, including clicks, favorites, shares, and conversions to purchases; The weighting coefficients ω1 and ω2 in step S4 are adjusted based on the operational behavior data; If the user conversion rate is lower than the preset threshold, the inventory confidence weight will be automatically reduced, the image feature similarity weight will be increased, and the image feature extraction model will be retrained.

Images