A method for constructing a clothing dataset based on multi-view and structured attributes and application thereof

Abstract

This invention discloses a method for constructing a clothing dataset based on multiple views and structured attributes, and its application. The method includes the following steps: collecting raw data from a publicly available fashion platform and performing quality filtering on the raw data; mining the structured attributes of the filtered text data to obtain a set of structured text attributes for the products; mining multi-view images of the products from the filtered image data to obtain a set of individual product images and a set of suit images; aligning the structured text attributes of the products with the individual product images and suit images to obtain a clothing dataset containing multiple views and structured text attributes. This invention solves the problems of existing methods neglecting the structured representation of attributes and having a single view, providing an efficient data foundation for research and applications in the fashion field.

Description

Technical Field

[0001] This invention relates to the intersection of computer vision and natural language processing, and in particular to a method for constructing a clothing dataset based on multiple views and structured attributes, and its application. Background Technology

[0002] With the rapid development of e-commerce and artificial intelligence technologies, the fusion analysis of clothing images and text data has become increasingly important in applications such as product recommendation, image and text retrieval, and virtual try-on. To support these applications, industry and academia have built several clothing image and text datasets, such as FashionGen, Fashion200K, and Fashion-MNIST.

[0003] However, existing datasets still have shortcomings in the following aspects:

[0004] Limited image perspective: Many datasets contain clothing images that only include a frontal view, lacking multi-view images such as side and back views, which limits the model's learning and understanding of the overall features of the clothing.

[0005] Lack of structured attributes: Although some datasets provide attribute labels, these labels are often not detailed enough and lack a comprehensive description of multi-dimensional attributes such as clothing material, color, and style, which affects the model's performance in fine-grained tasks.

[0006] Unstructured text descriptions: The text descriptions in existing datasets are mostly unstructured natural language, lacking a unified format and standard, which increases the difficulty of model processing and understanding.

[0007] Inconsistent data quality: During the data collection and labeling process, there may be problems such as mislabeling and omissions, which affect the overall quality of the dataset and the training effect of the model.

[0008] Therefore, there is an urgent need for a method for constructing clothing datasets that can effectively integrate multi-view images and structured attributes, and its application, to overcome the shortcomings of existing technologies and provide more comprehensive, richer, and easier-to-use data support for clothing-related tasks. Summary of the Invention

[0009] In view of this, the present invention aims to solve the problems of existing clothing datasets, such as single data source, lack of structured information, difficulty in aligning multimodal data, and lack of multi-view data. It provides a method for constructing clothing datasets based on multi-view and structured attributes and its application, in order to build a more comprehensive, richer, and easier-to-use clothing dataset, thereby improving the application effect of clothing-related tasks.

[0010] The present invention solves the technical problem by adopting the following technical solution:

[0011] A method for constructing a clothing dataset based on multiple views and structured attributes includes the following steps:

[0012] Step S1: Collect raw data from public fashion platforms and perform quality filtering on the raw data;

[0013] Step S2: Min the structured attributes of the filtered text data to obtain the set of structured text attributes of the product.

[0014] Step S3: Min the multi-view images of the products from the filtered image data to obtain a set of individual product images and a set of package images;

[0015] Step S4: Align the structured text attributes of the products with the individual product images and suit images to obtain a clothing dataset containing multiple views and structured text attributes.

[0016] Furthermore, in step S1, the raw data includes unstructured text descriptions and a collection of multi-view images.

[0017] Furthermore, step S1 involves quality filtering of the raw data, including:

[0018] Text filtering: Removes invisible characters, special symbols, and invalid content;

[0019] Image filtering: Remove blank or redundant images that are irrelevant to the clothing.

[0020] Furthermore, the method for mining the structured attributes of the filtered text data in step S2 includes:

[0021] Step S21: Min the basic attributes of the filtered text data to obtain a structured set of basic attributes.

[0022] Step S22: Perform exclusive attribute mining on the textualized product description to obtain a structured set of exclusive attributes;

[0023] Step S23: Merge the structured basic attribute set and the structured exclusive attribute set to obtain the product's structured text attribute set.

[0024] Furthermore, the method for performing basic attribute mining on the filtered text data in step S21 is as follows:

[0025] The text data is parsed based on preset attribute matching rules to identify and extract structured basic attributes; the attribute matching rules include regular expression matching methods, which are used to locate target attribute fields from text content or its associated information.

[0026] Furthermore, step S22 involves mining specific attributes of the textualized product description, including:

[0027] Based on the preset instruction template, the large language model is invoked to perform semantic recognition and parsing of product descriptions, thereby extracting a structured set of exclusive attributes. The instruction template is used to constrain the output format of the large language model so that it can generate attribute results that meet the requirements of a specific domain as expected.

[0028] Furthermore, the method for mining exclusive attributes is as follows:

[0029] First, the product description is processed in parallel using two large language models with low computational resource requirements to obtain two candidate sets of structured specific attributes.

[0030] Subsequently, a larger and more powerful language model is used to fuse, filter, and optimize the candidate set of structured proprietary attributes, thereby generating the final set of structured proprietary attributes.

[0031] Furthermore, step S4 aligns the structured text attributes of the product with the individual product image and the set image, including:

[0032] The multimodal feature alignment mechanism associates and maps the structured attributes obtained from text mining with the visual features of the corresponding products. The multimodal feature alignment mechanism includes feature encoding, representation space transformation and similarity calculation processes, which are used to realize the correspondence between text attributes and image features in a unified embedding space.

[0033] Furthermore, the alignment method is as follows:

[0034] First, a pre-trained image coding model is used to extract visual feature vectors of individual product images and package images; then, a language coding model is used to extract semantic representations of structured text attributes; subsequently, the two are fused through a cross-modal alignment network or mapping function to establish a one-to-one correspondence between text attributes and image features, so as to achieve a unified representation of product attributes and visual presentation.

[0035] An application of a clothing dataset constructed using a multi-view and structured attribute-based method is presented, applicable to the following tasks:

[0036] Cross-modal retrieval task: matching images with text queries or retrieving corresponding text descriptions from images;

[0037] Clothing classification and recognition task: Fine-grained category prediction by combining text and image modalities;

[0038] Pairing compatibility task: Evaluate the visual and semantic matching of multi-product combinations.

[0039] The present invention discloses a method for constructing a clothing dataset based on multiple views and structured attributes, and its application, which has the following beneficial effects:

[0040] (1) Significantly improved data comprehensiveness and richness: By integrating multi-view images (single item images and suit images) and structured text attributes (basic attributes and specific attributes), this invention overcomes the problems of single data sources and incomplete information in existing datasets, and constructs a clothing dataset that contains richer and more comprehensive visual and semantic information. This provides a broader data foundation for clothing-related tasks.

[0041] (2) Enhanced structured attribute mining capability: By combining preset matching rules and a large language model to mine basic and specific attributes, this invention can efficiently and accurately transform unstructured text into machine-understandable structured data. In particular, the introduction of a large language model for specific attribute mining greatly enhances the ability to understand complex and personalized clothing descriptions, making the semantic information of the dataset more detailed and in-depth.

[0042] (3) High accuracy of multimodal data alignment: This invention accurately aligns the structured text attributes of the product with the multi-view image, constructing a high-quality multimodal input pair. This alignment method ensures the semantic consistency between the image and the text, effectively solving the problem of difficult multimodal data alignment in the prior art, and providing reliable data support for subsequent image-text joint modeling.

[0043] (4) Wide range of applications: The clothing dataset constructed in this invention can effectively support a variety of complex clothing-related tasks, such as cross-modal retrieval, clothing classification and recognition, and matching compatibility evaluation. Because the dataset contains richer and more accurate multimodal information, the model trained on this dataset can achieve higher accuracy and better performance in these tasks, providing users with more intelligent and personalized services.

[0044] (5) Strict data quality control: Text filtering and image filtering mechanisms were introduced during the data collection stage, which effectively eliminated low-quality, redundant or irrelevant data, ensuring the purity and effectiveness of the dataset from the source, and reducing the difficulty of subsequent processing and the training cost of the model.

[0045] In summary, this invention provides an innovative and efficient method for constructing clothing datasets. The datasets constructed by this invention outperform existing technologies in terms of data quality, information richness, structure, and multimodal alignment, providing a solid data foundation for promoting the development of artificial intelligence technology in the clothing field. Attached Figure Description

[0046] Figure 1 This is a flowchart of the method of the present invention.

[0047] Figure 2 This is a schematic diagram illustrating the classification of attributes specific to this invention.

[0048] Figure 3 This is a schematic diagram of the instruction template of the present invention.

[0049] Figure 4 This is an example diagram of the construction method of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of 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 some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] This invention provides a method for constructing a structured clothing dataset based on image and text information and its application. It aims to extract rich and fine-grained structured attributes of clothing by collecting clothing images and text descriptions from e-commerce platforms, and by using a large language model and preset rules. This results in the construction of a clothing dataset with multi-view and multi-modal support, which is suitable for downstream tasks such as matching recommendations, style transfer, and classification.

[0052] This invention provides a method for constructing a clothing dataset based on multiple views and structured attributes, comprising the following steps:

[0053] Step S1: Collect raw data from public fashion platforms and perform quality filtering on the raw data; the raw data includes unstructured text descriptions and multi-view image sets.

[0054] Quality filtering of the raw data includes:

[0055] Text filtering: Removes invisible characters, special symbols, and invalid content;

[0056] Image filtering: Remove blank or redundant images that are irrelevant to the clothing.

[0057] In this step, to ensure data quality and validity, the collected raw data undergoes rigorous quality filtering. Specifically, text filtering includes removing invisible characters, special symbols, and invalid content to ensure the clarity and accuracy of the text descriptions. Image filtering aims to eliminate blank or redundant images unrelated to clothing, ensuring the validity and relevance of the image data. This step forms the basis for subsequent data processing and dataset construction, effectively avoiding the negative impact of low-quality data on the final dataset.

[0058] Step S2: Min the structured attributes of the filtered text data to obtain the set of structured text attributes of the product.

[0059] Methods for mining structured attributes of filtered text data include:

[0060] Step S21: Min the basic attributes of the filtered text data to obtain a structured set of basic attributes.

[0061] The method for performing basic attribute mining on the filtered text data is as follows:

[0062] This invention parses text data based on preset attribute matching rules to identify and extract structured basic attributes. These basic attributes typically refer to general clothing attributes such as color, gender, and type. The attribute matching rules include regular expression matching, used to locate target attribute fields from text content or its associated information. This method enables the automatic identification and extraction of basic attributes such as gender, color, size, and material. This invention ensures the accuracy and efficiency of basic attribute extraction by identifying and extracting these structured basic attributes through preset matching rules.

[0063] Step S22: Perform exclusive attribute mining on the textualized product description to obtain a structured set of exclusive attributes; exclusive attributes are those more descriptive and personalized attributes that cannot be obtained through simple matching rules, such as clothing style, design details, and applicable occasions.

[0064] Mining specific attributes from textualized product descriptions, including:

[0065] Based on the preset instruction template, the large language model is invoked to perform semantic recognition and parsing of product descriptions, thereby extracting a structured set of exclusive attributes. The instruction template is used to constrain the output format of the large language model so that it can generate attribute results that meet the requirements of a specific domain as expected.

[0066] The method for discovering exclusive attributes is as follows:

[0067] First, the product description is processed in parallel using two large language models with low computational resource requirements to obtain two candidate sets of structured specific attributes.

[0068] Subsequently, a larger and more powerful language model is used to fuse, filter, and optimize the candidate set of structured proprietary attributes, thereby generating the final set of structured proprietary attributes.

[0069] This invention combines instruction templates with a large language model to identify product descriptions, thereby extracting a structured set of specific attributes. This fully utilizes the semantic understanding capabilities of the large language model, overcoming the shortcomings of traditional methods in handling complex and diverse text descriptions, and enabling the dataset to contain richer and more granular clothing attribute information.

[0070] Step S23: Merge the structured basic attribute set and the structured specific attribute set to obtain the product's structured text attribute set. By merging structured attributes from different sources, this invention constructs a comprehensive and hierarchical clothing text attribute system, laying a solid foundation for subsequent image alignment and multimodal modeling.

[0071] This step aims to transform unstructured text descriptions into structured attributes that machines can understand and utilize.

[0072] Step S3: Mining multi-view images of products from the filtered image data to obtain a set of individual product images and a set of package images.

[0073] This step focuses on in-depth mining of image data, aiming to obtain multi-view information about clothing. Through a multimodal image-text model, it identifies and extracts individual images of the same garment (i.e., independent displays of the garment) and sets (i.e., combinations of the garment with other items). Acquiring multi-view images greatly enriches the visual information of the dataset, enabling the model to understand the characteristics of clothing from different angles. This is crucial for tasks such as fine-grained clothing recognition and outfit recommendations.

[0074] Step S4 involves aligning the structured text attributes of the products with individual product images and suit images to obtain a clothing dataset containing multiple views and structured text attributes. The fusion of structured text attribute representation and multiple views significantly enhances the model's ability to capture fashion knowledge.

[0075] Align the structured text attributes of the product with individual product images and set images, including:

[0076] The multimodal feature alignment mechanism associates and maps the structured attributes obtained from text mining with the visual features of the corresponding products. The multimodal feature alignment mechanism includes processes such as feature encoding, representation space transformation and similarity calculation, which are used to realize the correspondence between text attributes and image features in a unified embedding space.

[0077] The alignment method is as follows:

[0078] First, a pre-trained image coding model is used to extract visual feature vectors of individual product images and package images; then, a language coding model is used to extract semantic representations of structured text attributes; subsequently, the two are fused through a cross-modal alignment network or mapping function to establish a one-to-one correspondence between text attributes and image features, so as to achieve a unified representation of product attributes and visual presentation.

[0079] This step is crucial for achieving multimodal data fusion. This invention precisely matches the image information of the same product with its structured textual attribute information to construct multimodal input pairs. This alignment method ensures semantic consistency between images and text, providing high-quality training data for subsequent image-text joint modeling. Through this alignment, the dataset not only contains rich visual and textual information, but more importantly, clear relationships are established between these information, enabling the model to better understand the complex relationship between the visual features and semantic attributes of clothing.

[0080] This invention also provides an application of the dataset constructed by the method for building clothing datasets based on multiple views and structured attributes, applicable to the following tasks:

[0081] Cross-modal retrieval task: Users can match relevant clothing images through text queries, or retrieve corresponding text descriptions through images, greatly improving the flexibility and accuracy of retrieval.

[0082] Clothing classification and recognition task: Combine text and image modalities to perform fine-grained category prediction, making clothing classification more accurate and able to identify more specific clothing categories and attributes.

[0083] Outfit compatibility task: Evaluate the visual and semantic matching of multiple product combinations to provide users with smarter and more personalized clothing matching suggestions.

[0084] like Figure 1 As shown, this embodiment provides a method for constructing a clothing dataset based on multiple views and structured attributes, including the following steps:

[0085] Step S1 involves collecting raw data from a publicly available fashion platform and performing quality filtering on the raw data; the raw data includes unstructured text descriptions and a collection of multi-view images.

[0086] First, image and text information of clothing products are automatically collected from multiple e-commerce platforms (such as SSENSE), including product titles, descriptions, brands, prices, category information, and multi-view images. To improve the efficiency and stability of large-scale data collection, this module adopts a distributed crawler architecture, deployed on multiple collection server nodes. Tasks are sharded according to a URL list, and task distribution, status tracking, and load balancing are achieved through a scheduling system (such as a message queue-based scheduler).

[0087] Each crawler node independently processes a portion of the product pages, parsing the webpage structure and extracting product image links and text content. It employs mechanisms such as breakpoint resumption, exception retries, and dynamic User-Agent and proxy IP pools to ensure stable operation under high concurrency. The data collection tasks support scheduling by platform, category, and time, enabling large-scale, high-coverage acquisition of raw data. Collected image files are downloaded, renamed, and stored in a distributed file system, while text information is formatted into structured JSON format as input for subsequent data cleaning and structured processing modules.

[0088] After the data crawling is completed, the crawled results undergo standardization processing, including missing item removal, character cleanup, and sentence completion. For text content cleaning, this module performs character-level cleanup on product titles and descriptions. Character-level processing includes removing invisible characters (such as whitespace characters like \x00A0 in Unicode), emojis, hyperlinks (URLs), and other special symbols to prevent interference with the structured model's recognition of semantic structure. For image data cleaning, this module uses a combination of image link verification and image content detection to remove invalid images. This includes removing images with invalid links, images with a size of 0, images with incorrect formatting, or images that cannot be downloaded, as well as images identified by image features as blank or unrelated to clothing (such as solid color images, website placeholder images, etc.), thereby ensuring that the image content is authentic, complete, and usable for model training.

[0089] Meanwhile, this module performs an integrity check on each sample record. If key fields (such as name, description, image_urls) are missing, or the total number of images is less than a set threshold (such as less than 1 main image), the record will be marked as an invalid sample and filtered.

[0090] After cleaning, all samples will be converted into a standardized structural format (such as UTF-8 encoded JSON) for use by the subsequent structured attribute extraction module.

[0091] Step S2 involves mining the structured attributes of the filtered text data to obtain a set of structured text attributes.

[0092] First, based on statistical analysis of commonly used descriptive corpora from e-commerce platforms, an attribute keyword dictionary and regular expression matching template are constructed. The product name and description text are parsed to initially extract basic attributes such as gender, category, and designer, resulting in a structured set of basic attributes. For gender, the module identifies the gender of the product by recognizing common gender indicators (such as "men" and "women"). For clothing category identification, a multi-level category dictionary is constructed, matching corresponding words in the product name or description to achieve a step-by-step mapping from general categories to subcategories. Designer or brand information typically appears in the product title, category path, or at the beginning of the description. The system extracts standardized brand names through a combination of location context and brand dictionary judgment, which can be further mapped to a unified designer ID to support downstream modeling. In practical applications, this module also features attribute conflict detection and confidence scoring mechanisms to handle multiple candidate results, fuzzy matching, or semantic repetition, improving the robustness and accuracy of attribute extraction. All extracted results will undergo format standardization and be aligned with a standardized vocabulary, outputting attribute fields with a unified structure, laying the foundation for subsequent structured attribute generation and fusion.

[0093] Then, specific attribute mining is performed on the textualized product descriptions to obtain a structured set of specific attributes. Attribute candidates are generated in parallel using multiple large language models, and the results are then aggregated using a fusion module to obtain the final structured attributes.

[0094] To ensure consistency in structured output, this module uses a unified instruction template (see...). Figure 3 The system guides models to generate formatted outputs with attribute types, hierarchical structures, and annotations. Each model, after parsing the text, provides a set of candidate attributes with structural semantics. These candidates may differ in field coverage, naming conventions, or numerical representations. To address this, the system includes a fusion module, typically based on a more powerful language understanding model (such as GPT-4o), which comprehensively judges and semantically merges the candidate results generated by all models. During fusion, high-frequency consensus attributes are prioritized, while logically conflicting or invalid items are removed. Attribute fields can cover six major structural categories (see...). Figure 2The system includes single-object, multi-object, single-value, multi-value, feature-based, and nested attributes. The fusion process involves not only comparing and deduplicating attribute content, but also semantic similarity assessment, confidence determination of conflicting attributes, unified field naming, and standardized structural format. The system prioritizes retaining high-confidence candidates based on multi-model consensus, while filtering out logically contradictory, context-deficient, or semantically ambiguous fields. The final output structured attribute set not only possesses rich and broad field coverage but also exhibits strong robustness in terms of structural consistency and model understanding, making it directly usable for model training or attribute annotation tasks. This module's multi-model collaboration and fusion strategy significantly improves the ability to extract complex descriptions, long texts, or fine-grained attributes, and is one of the key steps in achieving high-quality structured clothing annotation.

[0095] Step S3 involves mining multi-view images of the products from the filtered image data to obtain a set of individual product images and a set of package images.

[0096] This module combines multiple dimensions, including the image's display order on the page, file path, naming characteristics, image size, and content features, to achieve automated view classification. In practical applications, product images often include various types such as front views, side views, detail images, and model outfit images. Front views are usually placed at the beginning of the page, with a relatively clean background, containing only a single garment, centered composition, and clear details. Outfit images, on the other hand, are often located later in the display order and may include elements such as models, full-body shots, and accessory combinations, resulting in complex compositions and diverse backgrounds. The system performs initial image screening by statistically analyzing image order patterns and combining them with image dimensional features (such as aspect ratio, area proportion, and background color distribution). It then uses a lightweight image classification model or human detection tool to detect and confirm human features within the image content, further improving the accuracy of the classification.

[0097] The segmentation results are presented in a structured manner, with the main view and accompanying images categorized into different image group fields, forming a clear view label structure. For example... Figure 4 As shown, the field view represents the main view, and the field matching view represents the matching chart. This structure not only facilitates downstream models in learning product details and matching contexts separately, but can also be directly applied to multimodal tasks such as clothing matching recommendations, multi-view matching, and style transfer, providing fundamental support for visual-language joint modeling.

[0098] Step S4 aligns the structured text attributes of the products with the individual product images and outfit images to obtain an apparel dataset containing multiple views and structured attributes.

[0099] Based on the product's unique identifier or other associated fields, the extracted structured text attributes such as gender, color, category, and material are associated with the corresponding product's individual and set images, respectively, to construct a one-to-one or one-to-many mapping relationship, thereby forming a unified clothing dataset that simultaneously contains multi-view image information and structured attribute information.

[0100] To verify the effectiveness of the structured clothing dataset constructed in this invention in visual language tasks, it was applied to several mainstream downstream tasks, including: clothing matching recommendation, structured attribute classification, image-text matching, and image-text retrieval. Based on the dataset's multiple views and rich structured attributes, the understanding ability and generalization performance of the task models can be effectively improved.

[0101] The structured clothing data construction method proposed in this invention has the advantages of clear data organization, low annotation cost, and strong adaptability, and has broad application prospects in the field of visual language understanding.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a garment dataset based on multi-view and structured attributes, characterized in that, The method comprises the following steps: Step S1, collecting original data from a public fashion platform and performing quality filtering on the original data; Step S2, performing mining of structured attributes on the filtered text data to obtain a structured text attribute set of the commodity; Step S3, performing mining of commodity multi-view images on the filtered image data to obtain a single product image set and a set of matching images; Step S4, aligning the structured text attributes of the commodity with the single product images and matching images to obtain a clothing data set containing multi-view and structured text attributes; The method for mining structured attributes from the filtered text data in step S2 comprises: Step S21, mining basic attributes from the filtered text data to obtain a structured basic attribute set; Step S22, mining exclusive attributes from the text-based commodity description to obtain a structured exclusive attribute set; Step S23, merging the structured basic attribute set and the structured exclusive attribute set to obtain a structured text attribute set of the commodity; The method for mining basic attributes from the filtered text data in step S21 comprises: Based on a preset attribute matching rule, the text data is parsed to identify and extract structured basic attributes; the attribute matching rule includes a regular expression matching method for locating target attribute fields from text content or associated information; The method for mining exclusive attributes from the text-based commodity description in step S22 comprises: Based on a preset instruction template, a large language model is called to perform semantic recognition and analysis on the commodity description, thereby extracting a structured exclusive attribute set; the instruction template is used to constrain the output form of the large language model, so that it can generate attribute results that meet the requirements of a specific field as expected; The method for mining exclusive attributes comprises: First, two large language models with low computational resource requirements are used to perform parallel processing on the commodity description to obtain two candidate structured exclusive attribute sets; Then, a large language model with larger parameters and stronger capabilities is used to fuse, screen and optimize the candidate structured exclusive attribute sets to generate a final structured exclusive attribute set.

2. The method of claim 1, wherein, In step S1, the original data includes unstructured text descriptions and multi-view image sets.

3. The method of claim 2, wherein, In step S1, the quality filtering on the original data comprises: Text filtering: removing invisible characters, special symbols and invalid content; Image filtering: removing blank or redundant images unrelated to clothing.

4. The method of claim 1, wherein, In step S4, aligning the structured text attributes of the commodity with the single product images and matching images comprises: Through a multi-modal feature alignment mechanism, the structured attributes obtained based on text mining are associated and mapped with the visual features of the corresponding commodity; the multi-modal feature alignment mechanism includes feature encoding, representation space transformation and similarity calculation process, which is used to realize the corresponding relationship between text attributes and image features in a unified embedding space.

5. The method of claim 4, wherein, The alignment method comprises: First, the visual feature vectors of single product images and set images are extracted using a pre-trained image encoding model; then the semantic representation of structured text attributes is extracted using a language encoding model; subsequently, the two are fused through a cross-modal alignment network or mapping function, thereby establishing a one-to-one correspondence between text attributes and image features, to achieve unified representation of product attributes and visual forms.

6. The application of the dataset constructed by the method of constructing a clothing dataset based on multi-view and structured attributes according to any one of claims 1-5, characterized in that, Applications include: Cross-modal retrieval task: matching images through text queries or retrieving corresponding text descriptions through images; Clothing classification and identification task: fine-grained class prediction combining text and image modalities; Compatibility task: evaluating the visual and semantic matching degree of multiple commodity combinations.

Images