A Method and Equipment for Garment Processing Information Recognition Based on a Data Engineering Operating System

Abstract

This specification discloses a method and device for identifying garment processing information based on a data engineering operating system. The method includes: preprocessing the original garment image uploaded by the client to obtain a non-deformable garment image and extracting image features from the non-deformable garment image; encoding the inspiration tags uploaded by the client into semantic vectors, and filtering candidate attribute subsets based on the similarity between the semantic vectors and the tag vectors in a pre-set multi-dimensional attribute database; aligning the image features and tag semantic vectors based on a cross-modal attention mechanism to generate a joint embedding representation vector, and inputting the joint embedding representation vector and the candidate attribute subsets into a pre-set recognition model to identify and output initial garment processing information; performing process completion on the initial garment processing information according to a pre-set process constraint specification library, and detecting the completed initial garment processing information based on a pre-set manufacturability detection strategy, and synchronizing the detected garment processing information to the client.

Description

Technical Field

[0001] This specification relates to the field of automated production technology, and in particular to a method and device for identifying garment processing information based on a data engineering operating system. Background Technology

[0002] With the continuous development of technology, the apparel industry is transforming towards digitalization and intelligentization. Traditional methods of apparel processing information identification and processing often rely on manual labor. However, due to differences in human experience, there are significant variations in the technical solutions for the same design, easily leading to significant design misunderstandings. This results in inefficient and error-prone information identification and transmission processes, making it difficult to meet rapidly changing market demands. Enterprises urgently need an efficient and accurate method to process apparel processing information to improve production efficiency, reduce costs, enhance product quality, and strengthen their competitiveness in the market.

[0003] Currently, the methods for acquiring garment processing information generally rely on image processing software and CAD software. These simple image analysis tools and pre-defined basic rules extract basic visual features from garment images, such as color histograms and texture features, and then match them with a pre-established simple feature library to identify garment information. While this method reduces reliance on manual labor, it cannot comprehensively and accurately describe the complex information of garments based solely on simple features like color and texture. It struggles to effectively identify deeper information such as style and craftsmanship. Furthermore, it lacks semantic understanding of garment information, failing to grasp the inherent relationships and semantic meanings between garment style, craftsmanship, and other information. This hinders more specific information analysis and processing, and it also lacks adaptive and intelligent learning capabilities, making it difficult to accurately identify the pattern and craftsmanship information of garments with innovative designs. Summary of the Invention

[0004] To address the aforementioned technical problems, one or more embodiments of this specification provide a method and device for identifying garment processing information based on a data engineering operating system.

[0005] One or more embodiments of this specification employ the following technical solutions:

[0006] This specification provides one or more embodiments of a method for identifying garment processing information based on a data engineering operating system. The method is executed based on the data engineering operating system and includes:

[0007] The original clothing image uploaded by the client is preprocessed to obtain a non-deformable clothing image, and the image features of the non-deformable clothing image are extracted.

[0008] The inspiration tags uploaded by the client are encoded into semantic vectors, and a subset of candidate attributes is selected based on the similarity between the semantic vectors and the tag vectors in the pre-set multi-dimensional attribute database; wherein, the attributes include: style, craftsmanship, fabric, and pattern.

[0009] Image features and label semantic vectors are aligned based on a cross-modal attention mechanism to generate a joint embedding representation vector. The joint embedding representation vector and the candidate attribute subset are then input into a preset recognition model to identify and output initial garment processing information.

[0010] The initial garment processing information is supplemented according to the pre-set process constraint specification library, and the supplemented initial garment processing information is tested according to the pre-set manufacturable detection strategy. The garment processing information that passes the test is synchronized to the client.

[0011] Optionally, in one or more embodiments of this specification, the original clothing image uploaded by the client is preprocessed to obtain a non-deformable clothing image, specifically including:

[0012] Based on the pixel grayscale values ​​of the original clothing image, the texture distribution features corresponding to the original clothing image are determined according to the pixel grayscale values, and the first image region distribution information and multiple second image distribution information of the original clothing image are determined based on the texture distribution features; wherein, the texture of the first image region is lower than the texture of the second image region;

[0013] The current clothing type is determined based on the original clothing image uploaded by the client, and a preset standard template is obtained based on the current clothing type; wherein, the current clothing type includes: top, pants, and skirt;

[0014] Based on a convolutional neural network architecture, the clothing key points of the original clothing image are extracted, and a homography matrix corresponding to the original clothing image is constructed according to the correspondence between the clothing key points and the preset standard template.

[0015] The original clothing image is subjected to perspective transformation based on the homography matrix to obtain an initial non-deformable clothing image;

[0016] The second image region corresponding to the distribution information of the second image is extracted from the initial non-mutant clothing image by using an image mask, and the second image region is smoothed by a preset bilateral filter to obtain a filtered second image region. The initial non-mutant clothing image is then corrected based on the filtered second image region to obtain a non-mutant clothing image.

[0017] Optionally, in one or more embodiments of this specification, the inspiration tags uploaded by the client are encoded into semantic vectors, and a subset of candidate attributes is selected based on the similarity between the semantic vectors and the tag vectors in a pre-set multi-dimensional attribute database. Specifically, this includes:

[0018] Based on a pre-set thesaurus, the inspiration tags are replaced to obtain standard inspiration tags, and the standard inspiration tags are encoded to obtain the semantic vectors corresponding to each standard inspiration tag;

[0019] Read the vector representations of each label in the preset multi-dimensional attribute database, and calculate the similarity between the semantic vector and the vectors of each label in each dimension attribute database in turn, so as to determine the similarity matrix between the semantic vector and each dimension attribute database.

[0020] The similarity matrix is ​​used to determine the similarity relationship between the standard inspiration tags and the attribute databases of each dimension, and the attribute databases of each dimension are filtered according to the similarity relationship to obtain a subset of candidate attributes.

[0021] Optionally, in one or more embodiments of this specification, the step of aligning image features and label semantic vectors based on a cross-modal attention mechanism to generate a joint embedding representation vector specifically includes:

[0022] Obtain the feature vector corresponding to the image features, and map the feature vector and the label semantic vector to the recognition dimension corresponding to the preset recognition model based on linear transformation;

[0023] The degree of correlation between each element in the feature vector and each element in the label semantic vector is determined based on the dot product between each element in the feature vector and each element in the label semantic vector.

[0024] The cross-modal attention weight matrix is ​​determined based on the degree of correlation between each element in the feature vector and each element in the label semantic vector;

[0025] The feature vector and the label semantic vector are adjusted based on the cross-modal attention weight matrix to obtain the adjusted feature vector and the adjusted label semantic vector.

[0026] The adjusted feature vector and the adjusted label semantic vector are weighted and summed to obtain the joint embedding representation vector.

[0027] Optionally, in one or more embodiments of this specification, before inputting the joint embedding representation vector and the candidate attribute subset into a preset recognition model, the method further includes:

[0028] Call the historical clothing database corresponding to the data engineering operating system to use the attribute-annotated clothing images in the historical database as the training dataset;

[0029] Randomly crop the historical clothing images in the training dataset at different scales to obtain local historical clothing images, and use the local historical clothing images as new training data to expand the training dataset to obtain the first training dataset;

[0030] Gamma transformation is performed on each historical clothing image in the expanded training dataset to further expand the first training dataset based on the transformed historical clothing images, thereby obtaining a second training dataset.

[0031] Based on the noise parameters corresponding to each fabric texture, a corresponding fabric texture image is created, and the pixel values ​​of the fabric texture image are added to the pixel values ​​of each historical clothing image in the second training dataset to obtain the expanded third training dataset.

[0032] The visual neural network model is trained based on the third training data to obtain a visual neural network model that meets the requirements as a preset recognition model.

[0033] Optionally, in one or more embodiments of this specification, the step of completing the initial garment processing information according to a preset process constraint specification library specifically includes:

[0034] Extract the entity information of the initial garment processing information, and obtain the compliance features in the pre-set process constraint specification library corresponding to the initial garment processing information based on the entity association relationship between the entity information and the entity corresponding to the pre-set process constraint specification library.

[0035] Based on the aforementioned compliance features, a pre-set garment process knowledge graph is traversed to obtain process nodes associated with each compliance feature, thereby generating candidate process chains.

[0036] Based on the selected candidate process chains, the missing processes in the initial garment processing information are determined, and the initial garment processing information is completed based on the process nodes corresponding to the missing processes.

[0037] Optionally, in one or more embodiments of this specification, the step of detecting the completed initial garment processing information based on a preset manufacturable detection strategy, and synchronizing the detected garment processing information to the client, specifically includes:

[0038] The garment processing information is processed using an intelligent coding engine to obtain the structured parameters corresponding to the garment processing information; wherein, the structured parameters are BOM data;

[0039] The structured parameters are transformed based on the corresponding geometric constraint mapping to obtain CAD constraint conditions corresponding to the garment processing information;

[0040] The constraint solver is invoked to verify the CAD constraints and obtain the test results; wherein, the verification includes: process feasibility verification, process conflict verification and garment closure verification;

[0041] If the test result is that the test is passed, the garment processing information that has passed the test will be synchronized to the client.

[0042] Optionally, in one or more embodiments of this specification, after the obtained garment processing information that has passed the inspection is synchronized to the client, the method further includes:

[0043] Obtain the garment piece area corresponding to the garment processing information, and sort the garment piece areas in descending order based on a greedy strategy to obtain the initial layout for material laying.

[0044] The fabric utilization rate is determined by calculating the ratio of the total area of ​​all garment pieces to the total area occupied by the initial layout of the fabric. The fabric utilization rate is used as a fitness function, and the initial layout of the fabric is optimized based on the fitness function to obtain the optimal layout of the fabric.

[0045] The fabric utilization rate corresponding to the optimal fabric layout is compared with the preset utilization rate threshold. If it is less than the preset utilization rate threshold, the simulated annealing strategy is triggered to locally optimize the optimal fabric layout and obtain the current fabric layout.

[0046] Receive order specification information uploaded by the client, and compare the order specification information with the BOM data corresponding to the garment processing information to determine whether there are updated parameters;

[0047] If so, iteratively update the current material layout to obtain the material layout to be produced;

[0048] The production layout, order specifications, garment processing information, and BOM data are encapsulated to obtain a data packet to be transmitted, which is then sent to the factory for production integration.

[0049] This specification provides one or more embodiments of a garment processing information recognition device based on a data engineering operating system. The device includes:

[0050] At least one processor; and,

[0051] A memory communicatively connected to the at least one processor; wherein,

[0052] The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to:

[0053] The original clothing image uploaded by the client is preprocessed to obtain a non-deformable clothing image, and the image features of the non-deformable clothing image are extracted.

[0054] The inspiration tags uploaded by the client are encoded into semantic vectors, and a subset of candidate attributes is selected based on the similarity between the semantic vectors and the tag vectors in the pre-set multi-dimensional attribute database; wherein, the attributes include: style, craftsmanship, fabric, and pattern.

[0055] Image features and label semantic vectors are aligned based on a cross-modal attention mechanism to generate a joint embedding representation vector. The joint embedding representation vector and the candidate attribute subset are then input into a preset recognition model to identify and output initial garment processing information.

[0056] The initial garment processing information is supplemented according to the pre-set process constraint specification library, and the supplemented initial garment processing information is tested according to the pre-set manufacturable detection strategy. The garment processing information that passes the test is synchronized to the client.

[0057] Optionally, in one or more embodiments of this specification, the inspiration tags uploaded by the client are encoded into semantic vectors, and a subset of candidate attributes is selected based on the similarity between the semantic vectors and the tag vectors in a pre-set multi-dimensional attribute database. Specifically, this includes:

[0058] Based on a pre-set thesaurus, the inspiration tags are replaced to obtain standard inspiration tags, and the standard inspiration tags are encoded to obtain the semantic vectors corresponding to each standard inspiration tag;

[0059] Read the vector representations of each label in the preset multi-dimensional attribute database, and calculate the similarity between the semantic vector and the vectors of each label in each dimension attribute database in turn, so as to determine the similarity matrix between the semantic vector and each dimension attribute database.

[0060] The similarity matrix is ​​used to determine the similarity relationship between the standard inspiration tags and the attribute databases of each dimension, and the attribute databases of each dimension are filtered according to the similarity relationship to obtain a subset of candidate attributes.

[0061] The above-described at least one technical solution adopted in the embodiments of this specification can achieve the following beneficial effects:

[0062] Preprocessing the original garment images to obtain deformation-free garment images eliminates the interference of image deformation on feature extraction, making the extracted image features more realistically reflect the garment itself, thus providing a more reliable foundation for subsequent information recognition. Encoding inspiration tags into semantic vectors and filtering candidate attribute subsets enables precise matching of relevant candidate attribute subsets based on the customer's personalized inspiration and needs, thereby narrowing the query scope for subsequent recognition and matching. Aligning image features with tag semantic vectors based on a cross-modal attention mechanism can capture the complex semantic relationships between images and text. The generated joint embedding representation vector integrates multimodal information, obtaining garment-related information from multiple perspectives, no longer limited to a single modality, and can more comprehensively describe the characteristics and attributes of the garment. By performing process completion based on a pre-set process constraint specification library, gaps in process information are filled, making garment processing information more complete and solving the problem that traditional garment processing information processing methods cannot quickly and accurately respond to customers' personalized customization needs. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments or prior art of this specification, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0064] Figure 1 A schematic diagram of a method for identifying garment processing information based on a data engineering operating system, provided in the embodiments of this specification;

[0065] Figure 2 A schematic diagram of an upload interface for a client provided in an embodiment of this specification;

[0066] Figure 3 This is a schematic diagram of a client-side interface for synchronously acquiring garment processing information, provided as an embodiment of this specification.

[0067] Figure 4 This is a schematic diagram of a client-side interface for uploading order specification information, provided as an embodiment of this specification.

[0068] Figure 5 This is a schematic diagram of the structure of a garment processing information recognition device based on a data engineering operating system, provided as an embodiment of this specification. Detailed Implementation

[0069] This specification provides an embodiment of a method and device for identifying garment processing information based on a data engineering operating system.

[0070] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0071] like Figure 1 As shown, this specification provides a flowchart illustrating a method for identifying garment processing information based on a data engineering operating system in one or more embodiments. Figure 1 As can be seen from one or more embodiments of this specification, a method for identifying garment processing information based on a data engineering operating system includes the following steps:

[0072] S101: Preprocess the original clothing image uploaded by the client to obtain a non-deformable clothing image, and extract the image features of the non-deformable clothing image.

[0073] Original clothing images may be deformed due to shooting angle, clothing wearing condition, or other factors, potentially distorting patterns, lines, and other features. To avoid inaccurate feature extraction caused by deformation affecting the judgment of the true information of the clothing, this embodiment of the specification preprocesses the original clothing image uploaded by the client after acquisition. This preprocessing eliminates the deformation in the image, resulting in a deformation-free clothing image. This allows the image features extracted from the deformation-free clothing image to more accurately reflect the true attributes of the clothing. This process ensures a unified standard for feature extraction of various clothing images uploaded by different clients, facilitating subsequent fusion processing with other modal information and improving the accuracy of clothing processing information recognition.

[0074] Specifically, in one or more embodiments of this specification, the original clothing image uploaded by the client is preprocessed to obtain a non-deformable clothing image, specifically including:

[0075] During the shooting process, the original clothing image often undergoes varying degrees of deformation due to factors such as shooting angle, clothing wearing condition, and wrinkles. The texture in the original clothing image can also interfere with subsequent analysis and processing due to the deformation. Therefore, in order to clarify the image region distribution information of different textures, the embodiments of this specification will determine the texture distribution features corresponding to the original clothing image based on the pixel grayscale values, and determine the first image region distribution information and multiple second image distribution information of the original clothing image based on the texture distribution features; wherein, the texture of the first image region is lower than the texture of the second image region.

[0076] Then, based on the original clothing image uploaded by the client, the current clothing type is determined, and a preset standard template is obtained based on this clothing type. It should be noted that the current clothing type includes: tops, pants, and skirts. Obtaining the corresponding preset standard template allows for appropriate processing methods to be used for different types of clothing, better handling the differences between different clothing types and improving the accuracy and adaptability of the processing. Next, the clothing key points of the original clothing image are extracted using a convolutional neural network architecture. Based on the correspondence between the clothing key points and the preset standard template, a homography matrix corresponding to the original clothing image is constructed. Perspective transformation is then performed on the original clothing image based on the homography matrix to obtain an initial, deformation-free clothing image. In other words, using a convolutional neural network architecture to extract clothing key points and constructing a homography matrix based on their correspondence with the preset standard template can accurately describe the geometric transformation relationship between the original clothing image and the standard template, providing a reliable basis for perspective transformation and effectively correcting image deformation. To address interference from multi-textured regions and improve subsequent recognition accuracy, a second image region corresponding to the distribution information of the second image is extracted from the initial non-deformable clothing image using an image mask. This second image region is then smoothed using a preset bilateral filter to obtain a filtered second image region. The initial non-deformable clothing image is then corrected based on this filtered second image region to obtain the final non-deformable clothing image. This process first obtains the initial non-deformable clothing image through perspective transformation, and then performs smoothing filtering and correction on specific second image regions. This step-by-step processing method can further optimize image quality, reduce noise and irregular textures in the image, and make the final non-deformable clothing image more clearly and accurately represent the true shape of the clothing.

[0077] S102: Encode the inspiration tags uploaded by the client into semantic vectors, and filter candidate attribute subsets based on the similarity between the semantic vectors and the tag vectors in the pre-set multi-dimensional attribute database; wherein, the attributes include: style, craftsmanship, fabric, and pattern.

[0078] In the embodiments of this specification, after obtaining image features based on the above steps, since searching in a database containing a large amount of data would consume a lot of analysis costs, in order to narrow the search range and improve accuracy, images uploaded by the client, such as... Figure 2The inspiration tags shown are encoded, and then the inspiration tags uploaded by the client are encoded into semantic vectors. By converting the originally complex and diverse text tags, which are difficult to process directly, into a digital vector form that computers can understand and process, the data is quantified and standardized, facilitating subsequent calculations and analysis. Furthermore, by calculating the similarity between the semantic vectors and the tag vectors in a pre-built multi-dimensional attribute database, the association between inspiration tags and attribute tags such as style, craftsmanship, fabric, and pattern in the database can be mined from a semantic level. In this way, based on the semantic connotation of the inspiration tags, a subset of candidate attributes with high relevance can be selected, providing targeted information for further identification and improving the utilization efficiency of the database.

[0079] Specifically, in one or more embodiments of this specification, the inspiration tags uploaded by the client are encoded into semantic vectors, and a subset of candidate attributes is selected based on the similarity between the semantic vectors and the tag vectors in a pre-set multi-dimensional attribute database. This specifically includes:

[0080] First, based on a pre-built thesaurus, the inspiration tags uploaded by the client are replaced. The aim is to unify inspiration tags with different expressions but similar or identical meanings into a standardized form, reducing interference caused by the diversity of language expressions. For example, if "long coat" and "long coat" are synonyms, by replacing them with thesaurus, they are all unified as "long coat," thus obtaining a standard inspiration tag. Then, the standard inspiration tags are encoded. This may use word vector technology from natural language processing to convert each standard inspiration tag into a corresponding semantic vector. It can be understood that a semantic vector is a numerical representation that can reflect the semantic features and semantic relationships of the tags in a vector space, facilitating subsequent similarity calculations.

[0081] Then, the vector representations of each tag in the pre-set multi-dimensional attribute database are read to calculate the similarity between the semantic vector of each standard inspiration tag and the tag vectors in each dimension attribute database. There are various methods for calculating similarity, such as cosine similarity and Euclidean distance, which will not be elaborated here. Through the above calculations, a similarity matrix between the semantic vector and each dimension attribute database is determined. It should be noted that the rows of this matrix can represent standard inspiration tags, the columns can represent tags in each dimension attribute database, and the element values ​​in the matrix are the similarity scores between the corresponding standard inspiration tags and attribute database tags. This presents the similarity relationship between them intuitively in matrix form. Then, based on the obtained similarity matrix, the similarity relationship between the standard inspiration tags and each dimension attribute database is analyzed and determined. For example, a similarity threshold can be set; when the similarity score is higher than this threshold, the standard inspiration tag and the corresponding attribute database tag are considered to have a strong similarity. Based on the similarity relationship, each dimension attribute database is filtered, retaining attribute tags with high similarity to the standard inspiration tags, and these tags are formed into a candidate attribute subset. This subset of candidate attributes includes multi-dimensional attribute information related to the semantics of the inspiration tag, such as style, craftsmanship, fabric, and pattern.

[0082] S103: Align image features and label semantic vectors based on cross-modal attention mechanism to generate joint embedding representation vector, and input the joint embedding representation vector and the candidate attribute subset into a preset recognition model to identify and output initial garment processing information.

[0083] In steps S101-S102 above, the original clothing images uploaded by the client have been preprocessed, and image features of the non-deformable clothing images have been extracted. The inspiration tags uploaded by the client have also been encoded to obtain semantic vectors. Since images and text belong to different modalities and their information differs, this embodiment uses a cross-modal attention mechanism to automatically learn the relationship between image features and label semantic vectors. This avoids the problems of low accuracy and difficulty in recognizing complex clothing when relying solely on single image features. The embodiments in this specification automatically learn the relationship between image features and label semantic vectors through a cross-modal attention mechanism, and then fuse the image features and label semantic vectors to generate a joint embedding representation vector. This vector integrates information from both images and text, enabling a more comprehensive and accurate representation of the clothing's features and semantics. Furthermore, by calculating the similarity between the inspiration tag semantic vector and the tag vectors in a pre-set multi-dimensional attribute database, a subset of candidate attributes has been selected. This subset includes attribute information related to the inspiration tag, such as style, craftsmanship, fabric, and pattern. Therefore, the joint embedding representation vector and the subset of candidate attributes are provided as input to the pre-set recognition model. The model analyzes and processes this input information, using its internal parameters and structure to learn the mapping relationship between the input information and garment processing information. It then identifies and outputs initial garment processing information, which may include specific garment style details, required processing techniques, applicable fabric types, pattern sizes, etc., providing a foundation and guidance for subsequent garment processing and production.

[0084] Specifically, in one or more embodiments of this specification, a joint embedding representation vector is generated by aligning image features and label semantic vectors based on a cross-modal attention mechanism, which can be achieved through the following process:

[0085] First, the feature vectors corresponding to the clothing image features obtained after preprocessing and feature extraction are acquired. Simultaneously, the previously encoded label semantic vectors are also obtained. Since the image feature vectors and label semantic vectors may have different dimensions, but subsequent operations need to be performed in the same dimension, they are mapped to the recognition dimension corresponding to the pre-set recognition model using a linear transformation. This is done to enable comparison and fusion of these two different modal vectors in the same space, laying the foundation for subsequent calculations and processing. For example, if the image feature vector is originally 512-dimensional and the label semantic vector is 768-dimensional, they are both converted to the 1024-dimensional dimension required by the recognition model through a linear transformation. Then, based on the dot product between each element in the feature vector and each element in the label semantic vector (i.e., by calculating the sum of the products of corresponding elements of the two vectors), the correlation between each element in the feature vector and each element in the label semantic vector is determined. Then, based on the correlation between each element in the feature vector and each element in the label semantic vector, a cross-modal attention weight matrix is ​​determined. The feature vectors and label semantic vectors are then adjusted based on the cross-modal attention weight matrix. Specifically, each element of the feature vector is multiplied by the corresponding weight value in the attention weight matrix to obtain the adjusted feature vector. Similarly, the same operation is performed on the label semantic vector to obtain the adjusted label semantic vector. This method allows elements more closely related to other modalities to receive greater weight during the fusion process, thus highlighting this important information. The adjusted feature vector and the adjusted label semantic vector are then weighted and summed to obtain the final joint embedding representation vector. This joint embedding representation vector integrates information from image features and label semantic vectors, and highlights the correlation between them through an attention mechanism. It can more effectively represent the comprehensive information of clothing, providing higher-quality input data for subsequent input into the pre-built recognition model for accurate initial clothing processing information recognition.

[0086] Furthermore, in one or more embodiments of this specification, in order to improve the accuracy of process feature recognition, before inputting the joint embedding representation vector and the candidate attribute subset into the preset recognition model, the method further includes the following process:

[0087] The system accesses the historical clothing database corresponding to the data engineering operating system and retrieves attribute-annotated clothing images from this database as the initial training dataset. For example, clothing images in a given scene might be labeled with attributes including style, silhouette, craftsmanship, fabric color, and material. Then, to enable the model to learn the features of clothing images in different local contexts and improve its adaptability to various clothing image variations, the historical clothing images in the training dataset are randomly cropped at different scales. This method yields various localized historical clothing images, each containing different parts and features of the original image. These localized historical clothing images are added as new training data to the original training dataset, forming the first training dataset.

[0088] Then, gamma transforms are applied to each historical garment image in the expanded first training dataset to further expand the first training dataset, resulting in a second training dataset. This further enriches the diversity of the training data, enabling the model to learn garment image features under different contrasts and brightness levels, and enhancing the model's robustness to changes in image lighting. Corresponding fabric texture images are also created based on the noise parameters associated with each fabric texture. These noise parameters describe the characteristics and distribution of different fabric textures, and the fabric texture images generated using these parameters have a texture effect similar to actual fabrics. The pixel values ​​of the fabric texture images are added to the pixel values ​​of each historical garment image in the second training dataset to obtain the expanded third training dataset. This operation simulates fabric texture interference that may occur in actual garment production and photography, enabling the model to learn how to handle images with fabric texture noise during training, improving the model's recognition ability in complex real-world scenarios. The expanded third training dataset is used to train the visual neural network model. During the training process, the model continuously learns the image features and corresponding attribute annotation information in the third training dataset and adjusts its own parameters so that the model can accurately map image features to the corresponding clothing attributes. After multiple iterations of training, when the model's performance reaches a certain standard, a visual neural network model that meets the requirements can be obtained as a pre-set recognition model.

[0089] S104: Complete the initial garment processing information according to the preset process constraint specification library, and test the completed initial garment processing information based on the preset manufacturable detection strategy, and synchronize the garment processing information that has passed the test to the client.

[0090] Initial garment processing information is derived through the processing and identification of multiple sources, including garment image features and tag semantic vectors. However, this information may be incomplete and cannot be directly used to guide actual production. Therefore, the initial garment processing information is supplemented with a pre-set process constraint specification library. Then, based on a pre-set manufacturability detection strategy, the supplemented initial garment processing information is tested to determine its feasibility. If the test passes, the tested garment processing information is synchronized to the client. Figure 3 As shown. It should be noted that the pre-set process constraint specification library is a database containing a large amount of garment process knowledge and standard rules. It stores the physical / aesthetic constraint relationships between styles, fabrics and processes, such as the constraint relationship: "lockstitch is prohibited on silk fabrics".

[0091] Specifically, in one or more embodiments of this specification, the initial garment processing information is supplemented according to a pre-set process constraint specification library, which includes the following process:

[0092] The initial garment processing information contains various types of entity information, such as garment style (e.g., shirt, dress), fabric (e.g., pure cotton, silk), and fit (e.g., fitted, loose). This entity information forms the basis for subsequent process completion, so these key entity contents can be extracted from the initial garment processing information. Each entity is then associated with a series of specific process features and requirements. For example, "pure cotton fabric" may be associated with "washing process" and "softener treatment process." Therefore, based on the entity association relationship between the entity information and the entities in the pre-set process constraint specification library, the compliance features corresponding to the initial garment processing information in the pre-set process constraint specification library can be obtained. Based on the compliance features, the pre-set garment process knowledge graph is traversed to obtain the process nodes associated with each compliance feature, generating candidate process chains. Then, based on the selected candidate process chains, the missing processes in the initial garment processing information are determined, and the initial garment processing information is completed based on the process nodes corresponding to the missing processes. For example, if the initial garment processing information only mentions cutting and sewing, but the candidate process chain also includes ironing and quality inspection, then ironing and quality inspection are the missing processes. Based on the process nodes corresponding to the missing processes, this missing process information is added to the initial garment processing information. The added information includes detailed details such as specific process requirements, operating procedures, and required equipment, making the initial garment processing information more complete and providing comprehensive guidance for actual garment production.

[0093] Specifically, in one or more embodiments of this specification, the initial garment processing information after completion is detected based on a preset manufacturable detection strategy, and the detected garment processing information is synchronized to the client. The specific process includes the following steps:

[0094] First, the garment processing information is processed using an intelligent grading engine to obtain the structured parameters corresponding to the garment processing information. For example, in a certain application scenario, the intelligent grading engine will automatically generate specification data and automatically generate a BOM for structured data output. Then, the structured parameters are transformed based on the corresponding geometric constraint mapping to obtain the CAD constraint conditions corresponding to the garment processing information. For example, it can be converted into CAD constraint conditions in the form of CAD geometric constraint equations based on the process of "\text{curled edge area}=\text{sleeve edge offset inward 30mm}\quad\Rightarrow\quad\text{CAD constraint: Offset(sleeve_edge,-30mm)}". Then, the constraint solver is called to verify the CAD constraint conditions and obtain the detection results; the verification includes: process feasibility verification, process conflict verification, and garment piece closure verification. Specifically, the constraint solver can be called to verify geometric feasibility: output a conflict report (such as "curled edge exceeds garment piece boundary") and recommend parameter adjustment ranges. For garment piece closure verification, a topological approach can be used, calculating the Euler characteristic of the garment piece's outline edges. If the formula χ = V - E + F = 1 (the standard value for a closed outline), where V (number of vertices), E (number of edges), and F (number of faces) are satisfied, the garment is considered closed. Otherwise, contour repair based on Moving Least Squares (MLS) is triggered. Alternatively, a boundary cycle traversal approach can be used, employing depth-first search to detect boundary connectivity. If unclosed cycles exist, the gap locations are marked, and repair paths are generated. If the detection result is successful, the garment processing information is synchronized to the client.

[0095] Furthermore, in one or more embodiments of this specification, after the garment processing information that has passed the inspection is synchronized to the client, the method further includes the following process:

[0096] After the garment processing information that has passed inspection is synchronized to the client, the area of ​​each garment piece corresponding to that information is first obtained. Then, the garment piece areas are sorted in descending order based on a greedy strategy, that is, the garment pieces are arranged in descending order of area. Based on the initial layout, the ratio of the total area of ​​all garment pieces to the total area occupied by the initial layout area is calculated; this ratio is the fabric utilization rate. The fabric utilization rate is used as a fitness function, and the initial layout is optimized based on the fitness function to obtain the optimal layout. Then, the fabric utilization rate corresponding to the optimal layout is compared with a preset utilization rate threshold. If it is less than the preset utilization rate threshold, a simulated annealing strategy is triggered to locally optimize the optimal layout and obtain the current layout. Then, the client uploads data such as... Figure 4The order specifications shown are compared with the corresponding BOM data for garment processing to determine if there are any updated parameters. For example, if the user's waist size needs to be increased, the corresponding BOM data needs to be updated. In other words, when a customer selects a custom-made order, the BOM data, CAD pattern, and layout are recalculated. If the comparison determines that updated parameters exist, the current layout needs to be iteratively updated. This means that the position and rotation angle of the garment pieces are readjusted according to the updated parameters to obtain a layout for production that meets the new specifications. Through iterative updates, it is ensured that the layout always matches the customer's order specifications. Then, the layout for production, order specifications, garment processing information, and BOM data are packaged into a data packet to be transmitted. This data packet is sent to the factory for production integration. After receiving the data packet, the factory can produce the garment according to the information contained therein, ensuring that the produced garment meets the customer's requirements and quality standards. During this process, the order specifications uploaded by the client are received, compared with the BOM data to determine updated parameters, and the BOM data, CAD pattern, and layout are recalculated when the customer selects a custom-made order. This approach allows for rapid response to personalized customer needs, ensuring that produced garments meet specific requirements and improving customer satisfaction. Furthermore, by iteratively updating the current pattern layout and readjusting the position and rotation angle of garment pieces, the layout consistently matches the customer's order specifications. This dynamic adjustment method adapts to various changes that may occur during production, guaranteeing production continuity and accuracy. The pattern layout, order specifications, garment processing information, and BOM data are then packaged into a data package and sent to the factory. The factory can then produce garments based on the detailed information in the data package, avoiding information transmission errors or omissions, ensuring that the produced garments meet customer requirements and quality standards, and reducing production errors and rework costs.

[0097] like Figure 5 As shown, this specification provides a structural schematic diagram of a garment processing information identification device based on a data engineering operating system in one or more embodiments. Figure 5 As shown in one or more embodiments of this specification, a garment processing information identification device based on a data engineering operating system includes:

[0098] At least one processor; and,

[0099] A memory communicatively connected to the at least one processor; wherein,

[0100] The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to:

[0101] The original clothing image uploaded by the client is preprocessed to obtain a non-deformable clothing image, and the image features of the non-deformable clothing image are extracted.

[0102] The inspiration tags uploaded by the client are encoded into semantic vectors, and a subset of candidate attributes is selected based on the similarity between the semantic vectors and the tag vectors in the pre-set multi-dimensional attribute database; wherein, the attributes include: style, craftsmanship, fabric, and pattern.

[0103] Image features and label semantic vectors are aligned based on a cross-modal attention mechanism to generate a joint embedding representation vector. The joint embedding representation vector and the candidate attribute subset are then input into a preset recognition model to identify and output initial garment processing information.

[0104] The initial garment processing information is supplemented according to the pre-set process constraint specification library, and the supplemented initial garment processing information is tested according to the pre-set manufacturable detection strategy. The garment processing information that passes the test is synchronized to the client.

[0105] Furthermore, in one or more embodiments of this specification, the inspiration tags uploaded by the client are encoded into semantic vectors, and a subset of candidate attributes is selected based on the similarity between the semantic vectors and the tag vectors in a pre-set multi-dimensional attribute database. Specifically, this includes:

[0106] Based on a pre-set thesaurus, the inspiration tags are replaced to obtain standard inspiration tags, and the standard inspiration tags are encoded to obtain the semantic vectors corresponding to each standard inspiration tag;

[0107] Read the vector representations of each label in the preset multi-dimensional attribute database, and calculate the similarity between the semantic vector and the vectors of each label in each dimension attribute database in turn, so as to determine the similarity matrix between the semantic vector and each dimension attribute database.

[0108] The similarity matrix is ​​used to determine the similarity relationship between the standard inspiration tags and the attribute databases of each dimension, and the attribute databases of each dimension are filtered according to the similarity relationship to obtain a subset of candidate attributes.

[0109] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0110] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0111] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. A method for identifying garment processing information based on a data engineering operating system, characterized in that, The method is executed based on a data engineering operating system, and the method includes: The original clothing image uploaded by the client is preprocessed to obtain a non-deformable clothing image, and the image features of the non-deformable clothing image are extracted. The inspiration tags uploaded by the client are encoded into semantic vectors, and a subset of candidate attributes is selected based on the similarity between the semantic vectors and the tag vectors in the pre-set multi-dimensional attribute database; wherein, the attributes include: style, craftsmanship, fabric, and pattern. Image features and label semantic vectors are aligned based on a cross-modal attention mechanism to generate a joint embedding representation vector. The joint embedding representation vector and the candidate attribute subset are then input into a preset recognition model to identify and output initial garment processing information. The initial garment processing information is supplemented according to the pre-set process constraint specification library, and the supplemented initial garment processing information is tested according to the pre-set manufacturable testing strategy. The garment processing information that passes the test is synchronized to the client. The process completion of the initial garment processing information based on the pre-set process constraint specification library specifically includes: Extract the entity information of the initial garment processing information, and obtain the compliance features in the pre-set process constraint specification library corresponding to the initial garment processing information based on the entity association relationship between the entity information and the entity corresponding to the pre-set process constraint specification library. Based on the aforementioned compliance features, a pre-set garment process knowledge graph is traversed to obtain process nodes associated with each compliance feature, thereby generating candidate process chains. Based on the selected candidate process chains, the missing processes in the initial garment processing information are determined, and the initial garment processing information is completed based on the process nodes corresponding to the missing processes. The initial garment processing information is inspected based on a pre-defined manufacturable inspection strategy. Garment processing information that passes the inspection is then synchronized to the client. Specifically, this includes: The garment processing information is processed using an intelligent coding engine to obtain the structured parameters corresponding to the garment processing information; wherein, the structured parameters are BOM data; The structured parameters are transformed based on the corresponding geometric constraint mapping to obtain CAD constraint conditions corresponding to the garment processing information; The constraint solver is invoked to verify the CAD constraints and obtain the test results; wherein, the verification includes: process feasibility verification, process conflict verification and garment closure verification; If the test result is that the test is passed, the garment processing information that has passed the test will be synchronized to the client.

2. The method for identifying garment processing information based on a data engineering operating system according to claim 1, characterized in that, The original clothing image uploaded by the client is preprocessed to obtain a non-deformable clothing image, specifically including: Based on the pixel grayscale values ​​of the original clothing image, the texture distribution features corresponding to the original clothing image are determined according to the pixel grayscale values, and the first image region distribution information and multiple second image distribution information of the original clothing image are determined based on the texture distribution features; wherein, the texture of the first image region is lower than the texture of the second image region; The current clothing type is determined based on the original clothing image uploaded by the client, and a preset standard template is obtained based on the current clothing type; wherein, the current clothing type includes: top, pants, and skirt; Based on a convolutional neural network architecture, the clothing key points of the original clothing image are extracted, and a homography matrix corresponding to the original clothing image is constructed according to the correspondence between the clothing key points and the preset standard template. The original clothing image is subjected to perspective transformation based on the homography matrix to obtain an initial non-distorted clothing image; The second image region corresponding to the distribution information of the second image is extracted from the initial non-mutant clothing image by using an image mask, and the second image region is smoothed by a preset bilateral filter to obtain a filtered second image region. The initial non-mutant clothing image is then corrected based on the filtered second image region to obtain a non-mutant clothing image.

3. The method for identifying garment processing information based on a data engineering operating system according to claim 1, characterized in that, The step of encoding the inspiration tags uploaded by the client into semantic vectors, and then filtering candidate attribute subsets based on the similarity between the semantic vectors and the tag vectors in a pre-set multi-dimensional attribute database, specifically includes: Based on a pre-set thesaurus, the inspiration tags are replaced to obtain standard inspiration tags, and the standard inspiration tags are encoded to obtain the semantic vectors corresponding to each standard inspiration tag; Read the vector representations of each label in the preset multi-dimensional attribute database, and calculate the similarity between the semantic vector and the vectors of each label in each dimension attribute database in turn, so as to determine the similarity matrix between the semantic vector and each dimension attribute database. The similarity matrix is ​​used to determine the similarity relationship between the standard inspiration tags and the attribute databases of each dimension, and the attribute databases of each dimension are filtered according to the similarity relationship to obtain a subset of candidate attributes.

4. The method for identifying garment processing information based on a data engineering operating system according to claim 1, characterized in that, The method of aligning image features and label semantic vectors based on a cross-modal attention mechanism to generate a joint embedding representation vector specifically includes: Obtain the feature vector corresponding to the image features, and map the feature vector and the label semantic vector to the recognition dimension corresponding to the preset recognition model based on linear transformation; The degree of correlation between each element in the feature vector and each element in the label semantic vector is determined based on the dot product between each element in the feature vector and each element in the label semantic vector. The cross-modal attention weight matrix is ​​determined based on the degree of correlation between each element in the feature vector and each element in the label semantic vector; The feature vector and the label semantic vector are adjusted based on the cross-modal attention weight matrix to obtain the adjusted feature vector and the adjusted label semantic vector. The adjusted feature vector and the adjusted label semantic vector are weighted and summed to obtain the joint embedding representation vector.

5. The method for identifying garment processing information based on a data engineering operating system according to claim 1, characterized in that, Before inputting the joint embedding representation vector and the candidate attribute subset into the preset recognition model, the method further includes: Call the historical clothing database corresponding to the data engineering operating system, and use the attribute-annotated clothing images in the historical clothing database as the training dataset; Randomly crop the historical clothing images in the training dataset at different scales to obtain local historical clothing images, and use the local historical clothing images as new training data to expand the training dataset to obtain the first training dataset; Gamma transformation is performed on each historical clothing image in the expanded training dataset to further expand the first training dataset based on the transformed historical clothing images, thereby obtaining a second training dataset. Based on the noise parameters corresponding to each fabric texture, a corresponding fabric texture image is created, and the pixel values ​​of the fabric texture image are added to the pixel values ​​of each historical clothing image in the second training dataset to obtain the expanded third training dataset. The visual neural network model is trained based on the third training data to obtain a visual neural network model that meets the requirements as a preset recognition model.

6. The method for identifying garment processing information based on a data engineering operating system according to claim 1, characterized in that, After the obtained garment processing information that has passed the inspection is synchronized to the client, the method further includes: Obtain the garment piece area corresponding to the garment processing information, and sort the garment piece areas in descending order based on a greedy strategy to obtain the initial layout for material laying. The fabric utilization rate is determined by calculating the ratio of the total area of ​​all garment pieces to the total area occupied by the initial layout of the fabric. The fabric utilization rate is used as a fitness function, and the initial layout of the fabric is optimized based on the fitness function to obtain the optimal layout of the fabric. The fabric utilization rate corresponding to the optimal fabric layout is compared with the preset utilization rate threshold. If it is less than the preset utilization rate threshold, the simulated annealing strategy is triggered to locally optimize the optimal fabric layout and obtain the current fabric layout. Receive order specification information uploaded by the client, and compare the order specification information with the BOM data corresponding to the garment processing information to determine whether there are updated parameters; If so, iteratively update the current material layout to obtain the material layout to be produced; The production layout, order specifications, garment processing information, and BOM data are encapsulated to obtain a data packet to be transmitted, which is then sent to the factory for production integration.

7. A garment processing information recognition device based on a data engineering operating system, characterized in that, The device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to: The original clothing image uploaded by the client is preprocessed to obtain a non-deformable clothing image, and the image features of the non-deformable clothing image are extracted. The inspiration tags uploaded by the client are encoded into semantic vectors, and a subset of candidate attributes is selected based on the similarity between the semantic vectors and the tag vectors in the pre-set multi-dimensional attribute database; wherein, the attributes include: style, craftsmanship, fabric, and pattern. Image features and label semantic vectors are aligned based on a cross-modal attention mechanism to generate a joint embedding representation vector. The joint embedding representation vector and the candidate attribute subset are then input into a preset recognition model to identify and output initial garment processing information. The initial garment processing information is supplemented according to the pre-set process constraint specification library, and the supplemented initial garment processing information is tested according to the pre-set manufacturable testing strategy. The garment processing information that passes the test is synchronized to the client. The process completion of the initial garment processing information based on the pre-set process constraint specification library specifically includes: Extract the entity information of the initial garment processing information, and obtain the compliance features in the pre-set process constraint specification library corresponding to the initial garment processing information based on the entity association relationship between the entity information and the entity corresponding to the pre-set process constraint specification library. Based on the aforementioned compliance features, a pre-set garment process knowledge graph is traversed to obtain process nodes associated with each compliance feature, thereby generating candidate process chains. Based on the selected candidate process chains, the missing processes in the initial garment processing information are determined, and the initial garment processing information is completed based on the process nodes corresponding to the missing processes. The initial garment processing information is inspected based on a pre-defined manufacturable inspection strategy. Garment processing information that passes the inspection is then synchronized to the client. Specifically, this includes: The garment processing information is processed using an intelligent coding engine to obtain the structured parameters corresponding to the garment processing information; wherein, the structured parameters are BOM data; The structured parameters are transformed based on the corresponding geometric constraint mapping to obtain CAD constraint conditions corresponding to the garment processing information; The constraint solver is invoked to verify the CAD constraints and obtain the test results; wherein, the verification includes: process feasibility verification, process conflict verification and garment closure verification; If the test result is that the test is passed, the garment processing information that has passed the test will be synchronized to the client.

8. A garment processing information identification device based on a data engineering operating system according to claim 7, characterized in that, The step of encoding the inspiration tags uploaded by the client into semantic vectors, and then filtering candidate attribute subsets based on the similarity between the semantic vectors and the tag vectors in a pre-set multi-dimensional attribute database, specifically includes: Based on a pre-set thesaurus, the inspiration tags are replaced to obtain standard inspiration tags, and the standard inspiration tags are encoded to obtain the semantic vectors corresponding to each standard inspiration tag; Read the vector representations of each label in the preset multi-dimensional attribute database, and calculate the similarity between the semantic vector and the vectors of each label in each dimension attribute database in turn, so as to determine the similarity matrix between the semantic vector and each dimension attribute database. The similarity matrix is ​​used to determine the similarity relationship between the standard inspiration tags and the attribute databases of each dimension, and the attribute databases of each dimension are filtered according to the similarity relationship to obtain a subset of candidate attributes.

Images