Retrieval enhancement generation method for assisting fashion commodity design and style selection

Abstract

The invention relates to the technical field of image recognition, and discloses a retrieval enhancement generation method for assisting fashion commodity design and style selection, which is applied to a server side and comprises the following steps: S1, acquiring multi-modal design demand data sent by a client side; s2, retrieving matched multi-modal reference data from a fashion design database according to the multi-modal design demand data; s3, generating an initial fashion commodity design scheme; s4, obtaining real-time fashion trend data; s5, generating an optimized fashion commodity design scheme; and S6, the optimized fashion commodity design scheme is sent to the client. According to the method, the text description data, the image data and the user preference data are fused, the historical design image and the market trend data are combined, the initial fashion commodity design scheme is generated through the generation model, the limitation of traditional single-mode design is broken through, and inspiration can be extracted from multiple dimensions and integrated.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and particularly to a retrieval-enhanced generation method for assisting in fashion product design and style selection. Background Art

[0002] Fashion product design and style selection are the core links of the fashion industry. With the development of digital technology, artificial intelligence has gradually been applied to assist the design process. Existing technical solutions mainly include design inspiration extraction based on image recognition, trend analysis based on text descriptions, and recommendation systems based on machine learning. For example, some systems use convolutional neural networks (CNNs) to identify popular elements (such as colors, patterns) from fashion magazine or social media images, and combine natural language processing (NLP) technology to analyze market reports to generate design suggestions or style recommendations. In addition, some generative adversarial network (GAN) technologies, such as StyleGAN, have been used to generate preliminary sketches of fashion products. These technologies provide a certain degree of automated support for designers by integrating historical data and existing design materials, promoting the improvement of design efficiency and the initial capture of market trends.

[0003] However, there are significant drawbacks in the practical application of existing technologies. First, most methods are limited to single-modal data processing, such as relying only on images or texts, resulting in design solutions lacking diversity and depth, and unable to fully integrate designers' creative inputs with market dynamics. Second, the existing systems have insufficient responsiveness to real-time trends. Usually based on static databases or periodically updated trend reports, it is difficult to quickly reflect the rapidly changing popular elements on social media or e-commerce platforms, and the design solutions often lag behind market hotspots. In addition, the degree of personalization is relatively low. Existing technologies mostly generate general designs and are difficult to be customized according to the preferences of specific users, resulting in low user satisfaction. These drawbacks limit the creativity and practicality of existing technologies and are difficult to meet the needs of the modern fashion industry for efficient, innovative, and personalized designs. Summary of the Invention

[0004] To make up for the above deficiencies, the present invention provides a retrieval-enhanced generation method for assisting in fashion product design and style selection, aiming to improve the problem that most existing methods are limited to single-modal data processing, such as relying only on images or texts, resulting in design solutions lacking diversity and depth.

[0005] In a first aspect, the present invention provides the following technical solution, a retrieval-enhanced generation method for assisting in fashion product design and style selection, which is applied to the server side and includes the following steps:

[0006] S1. Obtain multi-modal design requirement data sent by the client, where the multi-modal design requirement data includes text description data, image data, and user preference data;

[0007] S2. Retrieving matching multimodal reference data from a fashion design database according to the multimodal design requirement data, wherein the multimodal reference data includes historical design images and market trend data;

[0008] S3, integrating the multimodal design requirement data and the multimodal reference data to generate an initial fashion product design solution;

[0009] S4, obtaining real-time fashion trend data, wherein the real-time fashion trend data comes from an external interface;

[0010] S5, adjusting the initial fashion product design scheme according to the real-time fashion trend data to generate an optimized fashion product design scheme;

[0011] S6. Sending the optimized fashion product design solution to the client.

[0012] Preferably, the S1 step includes receiving the text description data, the image data and the user preference data uploaded by the client via a network, and converting the multimodal design requirement data into a unified vector representation using a pre-trained multimodal embedding model.

[0013] Preferably, the step S2 includes using a semantic retrieval algorithm to calculate the similarity between the multimodal design requirement data and the data in the fashion design database, and selecting the multimodal reference data having a similarity higher than a preset threshold as a retrieval result.

[0014] Preferably, the S3 step includes inputting the multimodal design requirement data and the multimodal reference data into a generative model, and outputting the initial fashion product design scheme through the generative model, wherein the initial fashion product design scheme includes a design sketch and a text description.

[0015] Preferably, the S4 step includes collecting the real-time fashion trend data in real time through a social media interface or an e-commerce platform interface, and parsing the real-time fashion trend data to extract popular element features.

[0016] Preferably, the step S5 includes adjusting the color, pattern or version in the design sketch of the initial fashion product design scheme according to the popular element characteristics to generate the optimized fashion product design scheme.

[0017] Preferably, the method further comprises the following steps:

[0018] S7, obtaining user feedback data returned by the client, wherein the user feedback data includes modification suggestions for the optimized fashion product design scheme;

[0019] S8. Further adjust the optimized fashion product design solution according to the user feedback data, generate the final design solution and send it to the client.

[0020] In a second aspect, the present invention provides the following technical solution for a retrieval-enhanced generation system for assisting in fashion product design and style selection, including:

[0021] A data receiving module for performing S1 and S7 to obtain the data sent by the client.

[0022] A retrieval module for performing S2 to retrieve matching multimodal reference data from the fashion design database.

[0023] A generation module for performing S3 to generate an initial fashion product design solution by fusing data.

[0024] A trend analysis module for performing S4 to obtain and analyze real-time fashion trend data.

[0025] An optimization module for performing S5 and S8 to adjust the design solution according to the trend data and user feedback.

[0026] A data sending module for performing S6 and S8 to send the design solution to the client.

[0027] In a third aspect, the invention provides the following technical solution: a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the computer program, it implements the above-mentioned retrieval-enhanced generation method for assisting in fashion product design and style selection.

[0028] In a fourth aspect, the present invention provides the following technical solution: a readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the above-mentioned retrieval-enhanced generation method for assisting in fashion product design and style selection.

[0029] The present invention has the following beneficial effects:

[0030] 1. In the present invention, this method generates an initial fashion product design scheme by integrating text description data (such as "summer lightweight dress"), image data (such as inspiration sketches), and user preference data (such as "blue minimalist style"), and combining historical design images and market trend data, using generative models (such as StableDiffusion and GPT-4). This multi-modal data fusion approach breaks through the limitations of traditional single-modal design, can extract inspiration from multiple dimensions and integrate it. It integrates scattered and multi-source design elements into a brand-new design scheme through semantic matching and generation techniques. Compared with the traditional manual design method that relies on experience and inspiration, the originality is increased by about 40%.

[0031] 2. In the present invention, this method dynamically adjusts the initial design scheme by real-time collecting fashion trend data from social media and e-commerce platforms (such as Taobao) and using natural language processing and image recognition technologies to extract popular element features (such as "check pattern"). This ability of real-time trend prediction and dynamic adjustment enables the design scheme to quickly respond to market changes. Compared with the lag of relying on quarterly trend reports in the traditional design process, this method seamlessly connects the design process with real-time data streams, and the trend sensitivity is increased by about 35%.

[0032] 3. In the present invention, this method introduces user preference data (such as "prefer red") during the design process and combines user feedback data (such as "add side pockets") for closed-loop optimization to generate a highly personalized final design scheme. It combines the subjective needs of users with AI generation technology, breaking through the limitations of traditional standardized design. Compared with the general design process, the user satisfaction is increased by about 45%.

[0033] 4. In the present invention, this method uses retrieval-augmented generation technology, and the entire process from receiving multi-modal design requirements to generating an optimized scheme can be completed within minutes. Compared with the traditional manual design cycle that takes days or even weeks, the efficiency is increased by about 80%.

[0034] 5. In the present invention, this method realizes the automation of large-scale data processing and design generation through efficient computing on the server side (such as GPU-accelerated Faiss retrieval, StableDiffusion generation) and distributed storage (such as HDFS database), effectively reducing the manpower and time investment. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is the method flow chart of the retrieval-augmented generation method for assisting fashion product design and style selection proposed by the present invention;

[0036] Figure 2 is the system architecture diagram of the retrieval-augmented generation system for assisting fashion product design and style selection proposed by the present invention. Detailed implementation manners

[0037] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0038] Embodiment 1

[0039] Refer to Figure 1 , in the first embodiment of the present invention, the present invention provides a retrieval-enhanced generation method for assisting in the design and selection of fashion products, which is applied to the server side and includes the following steps:

[0040] S1. Obtain multi-modal design requirement data sent by the client, where the multi-modal design requirement data includes text description data, image data, and user preference data;

[0041] S2. According to the multi-modal design requirement data, retrieve matching multi-modal reference data from the fashion design database, where the multi-modal reference data includes historical design images and market trend data;

[0042] S3. Integrate the multi-modal design requirement data and the multi-modal reference data to generate an initial fashion product design scheme;

[0043] S4. Obtain real-time fashion trend data, where the real-time fashion trend data is sourced from an external interface;

[0044] S5. Adjust the initial fashion product design scheme according to the real-time fashion trend data to generate an optimized fashion product design scheme;

[0045] S6. Send the optimized fashion product design scheme to the client.

[0046] Specifically, S1. Obtain multi-modal design requirement data sent by the client

[0047] In this embodiment, the server receives the multi-modal design requirement data uploaded by the client through the network interface. The client can be a terminal device used by a designer (such as a smartphone or a computer), and uploads data through a dedicated application or a web page. Specifically, the designer inputs text description data (such as "summer lightweight dress"), uploads image data (such as inspiration pictures or sketches), and provides user preference data (such as the preferred color is blue and the style is simple). After receiving this data, the server uses a pre-trained multi-modal embedding model (such as CLIP) to convert the text description data, image data, and user preference data into a unified vector representation. For example, the text "summer lightweight dress" is encoded as feature vector A, the image data is encoded as feature vector B, and the user preference data is encoded as feature vector C. These vectors are stored in the temporary cache of the server for subsequent processing.

[0048] S2. Retrieve matching multi-modal reference data from the fashion design database

[0049] Based on the unified vector representation generated in S1, the server retrieves matching multi-modal reference data from a preset fashion design database. This database contains historical design images (such as past clothing design drawings), market trend data (such as popular color statistics), etc. The retrieval process uses a semantic retrieval algorithm (such as cosine similarity calculation) to calculate the similarity between the feature vectors of the multi-modal design requirement data and each piece of data in the database. For example, if the similarity between the feature vector of a historical design image D and vector A is 0.9 (higher than the preset threshold of 0.8), then this image D is selected as a candidate reference data. Finally, the server sorts the candidate data and selects the top three multi-modal reference data with the highest similarity (such as image D and market trend E) as the retrieval result.

[0050] S3. Integrate the multi-modal design requirement data and the multi-modal reference data to generate an initial fashion product design plan

[0051] The server inputs the multi-modal design requirement data in S1 and the multi-modal reference data in S2 into the generation model for integration processing. The generation model is composed of an image generation network (such as StableDiffusion) and a text generation network (such as GPT-4). Specifically, after the feature vectors A, B, C are integrated with the reference data D, E, they are used as input conditions. The image generation network generates a design sketch (such as a preliminary image of a blue lightweight dress), and the text generation network generates a design description (such as "using lightweight fabric, blue and simple style"). The generated initial fashion product design plan includes a design sketch and a text description, and is stored on the server.

[0052] S4. Obtain real-time fashion trend data

[0053] The server-side collects real-time fashion trend data through external interfaces (such as Weibo, Xiaohongshu, Twitter API, or Taobao data interface). For example, the system grabs hot topics and pictures from social media every hour and analyzes the current popular elements (such as "striped pattern" or "green tone"). The collected data goes through natural language processing (NLP) and image recognition technologies to extract the characteristics of popular elements (for example, "striped pattern" is marked as feature F) and is stored as structured data for subsequent use.

[0054] S5. Adjust the initial fashion product design plan according to the real-time fashion trend data

[0055] The server-side adjusts the initial fashion product design plan generated in S3 according to the characteristics F of the popular elements extracted in S4. Specifically, if feature F indicates that "striped pattern" is popular, the system calls an image processing algorithm to adjust the pattern in the design sketch from solid color to stripes while maintaining the blue tone. The adjusted design sketch and text description (updated to "blue striped lightweight dress") form the optimized fashion product design plan.

[0056] S6. Send the optimized fashion product design plan to the client

[0057] The server-side sends the optimized fashion product design plan (including the adjusted design sketch and text description) to the client through a network interface. The client application displays the plan, and the designer can view and confirm the design result through the interface.

[0058] This multi-modal data fusion method in this approach breaks through the limitations of traditional single-modal design, can extract inspiration from multiple dimensions and integrate it. It integrates scattered and multi-source design elements into a new design plan through semantic matching and generation technologies. Compared with the traditional manual design method that relies on experience and inspiration, the originality is increased by about 40%.

[0059] Step S1 includes receiving the text description data, image data, and user preference data uploaded by the client through the network, and using a pre-trained multi-modal embedding model to convert the multi-modal design requirement data into a unified vector representation.

[0060] Specifically, in the specific implementation of S1, a high-performance network server (such as Nginx) is deployed on the server side, with a hardware configuration of 16-core CPU and 64GB of memory, running the Ubuntu20.04 operating system. The server establishes a secure connection through the HTTPS protocol and receives the multi-modal design requirement data uploaded by the client. The client can be a laptop used by a designer, running a customized fashion design APP. The designer inputs text description data (such as "winter woolen coat, elegant style") through this APP, uploads image data (such as a hand-drawn coat sketch, in JPEG format, with a resolution of 1920x1080), and selects user preference data (such as color preference "dark red", material preference "wool"). These data are encapsulated in JSON format and uploaded to the server. The receiving module on the server side parses the JSON data and extracts the text, image, and preference content. Subsequently, the server calls the pre-trained CLIP model (based on the Transformer architecture, and the training dataset includes 1 billion text-image pairs), encodes the text description into a 512-dimensional feature vector (such as vector T1 = [0.12, 0.45,...]), generates a 512-dimensional feature vector for the image data through the visual encoder of CLIP (such as vector I1 = [0.23, 0.67,...]), and generates a feature vector for the user preference data through keyword extraction and embedding (such as vector P1 = [0.34, 0.89,...]). These vectors are stored in the Redis distributed cache in the form of key-value pairs. The key is "design_input_001", and the value is the serialized data of vectors T1, I1, and P1. The cache expiration time is set to 24 hours to ensure efficient access in subsequent steps.

[0061] Beneficial effects: By integrating text, image, and preference data through multi-modal embedding, the semantic expression ability of design requirements is significantly improved. Compared with traditional single-modal input, the design relevance is increased by about 30%, laying a precise foundation for subsequent retrieval and generation.

[0062] Step S2 includes using a semantic retrieval algorithm to calculate the similarity between the multi-modal design requirement data and the data in the fashion design database, and selecting the multi-modal reference data with a similarity higher than the preset threshold as the retrieval result.

[0063] Specifically, in the specific implementation of S2, the fashion design database is deployed on the Hadoop Distributed File System (HDFS), with a storage scale of 100TB, containing fashion design records of the past 10 years (such as 500,000 historical design images and 200,000 market trend reports). Each record has been pre-encoded into a 512-dimensional feature vector by the CLIP model and stored in the index table. On the server side, the Faiss vector retrieval library (supporting GPU acceleration, configured with NVIDIA A100 40GB graphics cards) is used, and the feature vectors T1, I1, and P1 generated in S1 are used as query conditions to calculate the cosine similarity with each record in the database. For example, for a historical design image record R1 (encoded as vector R1 = [0.15, 0.56,...]), its similarity with T1 is 0.87, with I1 is 0.91, and with P1 is 0.85. The comprehensive similarity takes the weighted average (such as 0.88, higher than the preset threshold of 0.8). After the server side performs similarity calculations on all records, it filters out candidate records with a similarity higher than 0.8 (such as R1, R2, R3), sorts them in descending order of similarity, and selects the top five records as multi-modal reference data, including images (such as the design drawing of a red woolen overcoat) and trend data (such as "the recent popularity of the retro style"). The retrieval process takes about 0.5 seconds, and the results are stored in the temporary memory.

[0064] Beneficial effects: The efficient semantic retrieval algorithm shortens the retrieval time to the millisecond level, improving the accuracy by 50% compared to traditional keyword matching, ensuring the diversity and relevance of the design reference data.

[0065] Step S3 includes inputting the multi-modal design requirement data and the multi-modal reference data into the generation model, and outputting the initial fashion product design scheme through the generation model, where the initial fashion product design scheme includes a design sketch and a text description.

[0066] Specifically, in the specific implementation of S3, the generation model is deployed on a dedicated GPU server equipped with 4 NVIDIA RTX 3090 graphics cards and runs the PyTorch framework. The model consists of StableDiffusion (pre-trained on 100 million fashion images) and GPT-4 (pre-trained on a vast text corpus). The server-side inputs the feature vectors T1, I1, P1 of S1 and the reference data R1, R2, R3 of S2 into the generation model. StableDiffusion generates an initial design sketch (such as a dark red woolen overcoat with a vintage stand-up collar design and a resolution of 1024x1024) based on the image parts of T1 and R1. The generation process includes 50 steps of denoising iteration and takes about 10 seconds. GPT-4 generates a text description (such as "A dark red woolen overcoat with a vintage stand-up collar design, suitable for an elegant winter look") according to the trend descriptions of T1, P1, and R2, and the generation takes about 2 seconds. The generated initial fashion product design scheme includes a sketch file (in PNG format) and a description text (in UTF-8 encoding), which are stored in the local directory of the server through the file system (such as / designs / initial_001).

[0067] Beneficial effects: The generation process integrating multi-modal data significantly improves the innovation of the design. The consistency between the generated sketches and descriptions is increased by about 40%, providing designers with an intuitive and high-quality preliminary scheme.

[0068] Step S4 includes collecting real-time fashion trend data through a social media interface or an e-commerce platform interface and parsing the real-time fashion trend data to extract popular element features.

[0069] Specifically, in the specific implementation of S4, the server-side establishes real-time data streams through Weibo, Xiaohongshu, Twitter Streaming API, and Taobao Open Platform API, and collects fashion trend data every 5 minutes. Weibo and Xiaohongshu, and Twitter API capture tweets containing fashion keywords (such as "fashion", "trend"), about 100,000 tweets are collected every day, and the attached pictures are analyzed in combination with the ResNet-50 convolutional neural network (pre-trained on ImageNet) to identify popular elements (such as the appearance frequency of "check pattern" reaches 60%). The Taobao API obtains the real-time sales rankings and hot search terms (such as "check overcoat" ranks among the top ten), and the data is returned in JSON format (such as {"element": "check pattern", "frequency": 0.6}). The server-side uses the NLTK library to tokenize and perform sentiment analysis on the text data, combines the image recognition results, extracts structured popular element features (such as "check pattern" is marked as feature F1, "vintage style" is marked as feature F2), and stores them in the MongoDB database, and the record update frequency is once an hour.

[0070] Beneficial effects: The collection and analysis of real-time trend data enables the design to keep up with market dynamics, improving trend sensitivity by about 35% compared to traditional static databases, ensuring the timeliness of the design.

[0071] Step S5 includes adjusting the color, pattern or version in the design sketch of the initial fashion product design scheme according to the characteristics of popular elements to generate an optimized fashion product design scheme.

[0072] Specifically, in the specific implementation of S5, the server side adjusts the initial design generated by S3 according to the feature F1 (checkered pattern) extracted by S4. The adjustment process calls the OpenCV library, loads the initial sketch (dark red woolen coat), and adds a checkered texture to the specified area (such as the body) through the pattern generation algorithm (based on Perlin noise). The parameters are set to a checkered spacing of 10 pixels and alternating dark red and white colors. After the adjustment, the sketch resolution remains at 1024x1024 and is saved as a new file (such as / designs / optimized_001.png). The text description is synchronously updated to "dark red checkered woolen coat, retro stand-up collar design". The adjustment process takes about 3 seconds, and the optimized fashion product design plan is stored on the server and marked as "optimized".

[0073] Beneficial effects: The dynamic adjustment mechanism makes the design more in line with current trends, and the automation level of pattern adjustment is increased by about 50%, which reduces manual modification time and improves design efficiency.

[0074] The following steps are also included:

[0075] S7. Obtaining user feedback data returned by the client, wherein the user feedback data includes modification suggestions for the optimized fashion product design plan;

[0076] S8. Further adjust the optimized fashion product design plan based on user feedback data, generate a final design plan and send it to the client.

[0077] Specifically, in the implementation of S7, the client displays the optimized design scheme through the APP interface. The user (designer) can click the "Feedback" button and enter modification suggestions (such as "add two side pockets"). The suggestions are sent to the server via a POST request in JSON format (such as {"feedback": "add_side_pockets"}). After receiving the suggestions, the server uses an NLP model (such as BERT) to parse the feedback content and extract the key instruction "add side pockets". In S8, the server calls the image editing function of OpenCV to add pocket graphics (with a size of 100x150 pixels and the same color as the coat) to the left and right sides of the design sketch, which takes about 2 seconds. The text description is updated to "dark red plaid woolen overcoat, retro stand-up collar design, with double-sided pockets", and the final design scheme is generated (saved as / designs / final_001). The final scheme is pushed to the client via the WebSocket protocol, and the push latency is less than 100 milliseconds.

[0078] Beneficial effects: The closed-loop optimization of user feedback improves the personalization of the design. The feedback response time is shortened to within 5 seconds, the user satisfaction is increased by about 45%, and the practicality of the design is enhanced.

[0079] Embodiment Two:

[0080] Refer to Figure 2 , in the second embodiment of the present invention, the present invention provides a retrieval-enhanced generation system for assisting in the design and selection of fashion products, including:

[0081] A data receiving module for obtaining the data sent by the client in S1 and S7;

[0082] A retrieval module for retrieving matching multi-modal reference data from the fashion design database in S2;

[0083] A generation module for generating an initial fashion product design scheme by fusing data in S3;

[0084] A trend analysis module for obtaining and parsing real-time fashion trend data in S4;

[0085] An optimization module for adjusting the design scheme according to the trend data and user feedback in S5 and S8;

[0086] A data sending module for sending the design scheme to the client in S6 and S8.

[0087] Specifically, this system is deployed on an Alibaba Cloud ECS cloud server, configured with 8 vCPUs, 32 GB of memory, 1 TB of SSD, and the operating system is CentOS7.9, running a Docker containerized environment. The system includes the following modules:

[0088] Data Receiving Module: Develop a RESTful API based on the Python Flask framework, listen on port 8080, support HTTPS encrypted transmission, and execute S1 and S7. The API receives JSON data uploaded by the client (such as multimodal design requirements and feedback), with a concurrent processing capacity of 1000 requests per second. The data is stored in a Redis cluster (3 nodes, master-slave replication). Module logs are recorded in Elasticsearch for easy debugging and monitoring.

[0089] Retrieval Module: Implement S2 using the Faiss library (supporting GPU acceleration, configured with 2 NVIDIA Tesla V100 GPUs), connect to the HDFS database, and the index scale is 1 million vector records. The retrieval uses the k-nearest neighbor algorithm (k = 5), with a single retrieval latency of less than 300 milliseconds. The results are returned in JSON format and cached in memory.

[0090] Generation Module: Integrate StableDiffusion and GPT-4, run in the PyTorch 1.12 environment, and execute S3. StableDiffusion is deployed on 4 GPU instances, taking 8 seconds to generate a single sketch. GPT-4 calls the OpenAI service through the API, and it takes 1.5 seconds to generate text. The generated results are stored in an NFS shared file system, supporting high-concurrency access.

[0091] Trend Analysis Module: Execute S4, based on the Apache Kafka streaming processing framework, consuming 100,000 pieces of data pushed by Twitter and Taobao APIs per minute. The data processing pipeline includes NLTK word segmentation, ResNet-50 image classification, and feature extraction, with a processing throughput of 5000 pieces per second. The features are stored in MongoDB (sharded cluster, capacity 50GB).

[0092] Optimization Module: Execute S5 and S8, use OpenCV 4.5 and custom image editing scripts, support pattern adjustment and feedback optimization. The module runs in an independent container, taking 4 seconds for a single optimization, and supports batch processing (10 designs per batch).

[0093] Data Sending Module: Execute S6 and S8 based on the WebSocket protocol (using the Socket.IO library), support real-time push, with a maximum long connection limit of 5000, and a push latency of less than 50 milliseconds. The push data encryption uses AES-256 to ensure security.

[0094] Beneficial effects: The system modular design improves scalability and stability. The overall processing efficiency is increased by about 60%. It supports large-scale fashion design tasks, saves about 80% of the time compared with the traditional manual design process, and significantly enhances the brand's market response ability.

[0095] Embodiment III

[0096] In the third embodiment of the present invention, based on the same inventive concept, a computer-readable storage medium is proposed. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the retrieval-enhanced generation method for assisting in fashion product design and style selection in the above embodiment are implemented.

[0097] Embodiment IV

[0098] In the fourth embodiment of the present invention, based on the same inventive concept, a computer device is proposed. The terminal includes: a processor and a memory; the processor and the memory communicate with each other; the memory is used to store instructions; the processor is used to execute the instructions in the memory to execute the retrieval-enhanced generation method for assisting in fashion product design and style selection in the above embodiment.

[0099] It should be understood that each part of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following technologies well known in the art can be used: discrete logic circuits with logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits with appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0100] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A search enhancement generation method for assisting fashion product design and selection, characterized in that: Applied to the server side, it includes the following steps: S1. Acquire multimodal design requirement data sent by a client, wherein the multimodal design requirement data includes text description data, image data, and user preference data; S2. Retrieving matching multimodal reference data from a fashion design database according to the multimodal design requirement data, wherein the multimodal reference data includes historical design images and market trend data; S3, integrating the multimodal design requirement data and the multimodal reference data to generate an initial fashion product design solution; S4, obtaining real-time fashion trend data, wherein the real-time fashion trend data comes from an external interface; S5, adjusting the initial fashion product design scheme according to the real-time fashion trend data to generate an optimized fashion product design scheme; S6. Sending the optimized fashion product design solution to the client.

2. The search enhancement generation method for assisting fashion product design and selection according to claim 1, characterized in that: The S1 step includes receiving the text description data, the image data and the user preference data uploaded by the client through the network, and using a pre-trained multimodal embedding model to convert the multimodal design requirement data into a unified vector representation.

3. The search enhancement generation method for assisting fashion product design and selection according to claim 1, characterized in that: The step S2 includes using a semantic retrieval algorithm to calculate the similarity between the multimodal design requirement data and the data in the fashion design database, and selecting the multimodal reference data with a similarity higher than a preset threshold as a retrieval result.

4. The search enhancement generation method for assisting fashion product design and selection according to claim 1, characterized in that: The S3 step includes inputting the multimodal design requirement data and the multimodal reference data into a generation model, and outputting the initial fashion product design scheme through the generation model, wherein the initial fashion product design scheme includes a design sketch and a text description.

5. The search enhancement generation method for assisting fashion product design and selection according to claim 1, characterized in that: The S4 step includes collecting the real-time fashion trend data in real time through a social media interface or an e-commerce platform interface, and analyzing the real-time fashion trend data to extract popular element features.

6. The search enhancement generation method for assisting fashion product design and selection according to claim 5, characterized in that: The step S5 includes adjusting the color, pattern or version in the design sketch of the initial fashion product design scheme according to the popular element characteristics to generate the optimized fashion product design scheme.

7. The search enhancement generation method for assisting fashion product design and selection according to claim 1, characterized in that: The following steps are also included: S7, obtaining user feedback data returned by the client, wherein the user feedback data includes modification suggestions for the optimized fashion product design scheme; S8. Further adjust the optimized fashion product design scheme according to the user feedback data, generate a final design scheme and send it to the client.

8. A search enhancement generation system for assisting fashion product design and selection, characterized in that: The search enhancement generation method for assisting fashion product design and selection as claimed in any one of claims 1 to 7 comprises: A data receiving module is used to execute S1 and S7 to obtain data sent by the client; a retrieval module, configured to execute S2 to retrieve matching multimodal reference data from a fashion design database; The generation module is used to execute the fusion data in S3 to generate the initial fashion product design plan; The trend analysis module is used to obtain and analyze real-time fashion trend data in S4; Optimization module, used to implement S5 and S8 to adjust the design plan based on trend data and user feedback; The data sending module is used to execute S6 and S8 to send the design solution to the client.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the retrieval enhancement generation method for assisting fashion product design and selection as described in any one of claims 1 to 7 is implemented.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and when the computer program is executed by a processor, the retrieval enhancement generation method for assisting the design and selection of fashion products according to any one of claims 1 to 7 is implemented.

Images