Optimization scheme generation method and device for user-defined article, and medium

Abstract

The invention provides an optimization scheme generation method and device for a user-defined article and a medium, and relates to the technical field of optimization scheme generation for the user-defined article, and the method comprises the steps: obtaining a plurality of article parts corresponding to an original article picture and a reference article picture; obtaining a plurality of candidate feature vectors corresponding to the original article picture; judging whether the material combination corresponding to each candidate feature vector is coordinated or not, and determining the candidate feature vector of which the judgment result is coordinated as an intermediate feature vector; obtaining a common article part feature vector CG; obtaining a first similarity; determining an undetermined feature vector from the intermediate feature vector corresponding to the first similarity greater than a first preset similarity threshold; pushing an article optimization scheme corresponding to the target feature vector to the user; according to the method, the generated optimization scheme can be ensured to have supply chain feasibility and production landing performance, and the aesthetic quality, the user intention integrating degree and the actual performability of the user-defined article optimization scheme are comprehensively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of generating optimization scheme of self-defined articles, and particularly relates to a method, device and medium for generating optimization scheme of self-defined articles of a user. BACKGROUND

[0002] Under the trend of individualized consumption, users often want to obtain unique articles through self-defined design, for example, combining the style of one piece of clothing with the material style of another piece of clothing. However, the existing methods for generating self-defined scheme usually have obvious limitations: firstly, the system relies on fixed matching rules, lacks intelligent and refined judgment of the coordination of multiple materials in multi-dimensional attributes such as texture and style, and is easy to cause the visual effect of the recommended combination to be jarring; secondly, when the user provides pictures of original articles and reference articles at the same time, the existing technology is difficult to accurately compare and integrate the intentions at the part level, often only performs overall style imitation or single attribute extraction, resulting in that the generated scheme does not meet the deep expectations of the user; thirdly, the existing methods mostly focus on design effects, and less consider actual production constraints such as material supply and process feasibility, which may lead to the scheme being unable to be implemented or the cost being too high. Therefore, there is an urgent need for a method for generating optimization scheme which can intelligently evaluate the coordination of materials, accurately understand and integrate the reference intentions of the user, and take into account the feasibility of the supply chain, so as to provide a self-defined solution for articles which truly meets the individualized needs of the user and is executable. SUMMARY

[0003] In view of the above technical problems, the technical scheme adopted by the present application is as follows: According to a first aspect of the present application, a method for generating optimization scheme of self-defined articles of a user is provided, comprising the following steps: S100, if the article pictures uploaded by the user are original article pictures and reference article pictures, segmenting each article part in the original article pictures and the reference article pictures to obtain a plurality of article parts corresponding to the original article pictures and the reference article pictures; S200, obtaining a plurality of candidate feature vectors corresponding to the original article pictures according to each preset attribute feature of each selectable material corresponding to each article part of the original article pictures; S300, inputting each candidate feature vector into a coordination judgment model corresponding to the original article to judge whether the material combination corresponding to each candidate feature vector is coordinated, and determining the candidate feature vector with the judgment result of coordination as an intermediate feature vector; S400, obtaining a common article part feature vector corresponding to each intermediate feature vector and a common article part feature vector CG corresponding to the reference article pictures; the common article part feature vector is obtained according to each preset attribute feature of each selectable material of the common article part corresponding to the original article pictures and the reference article pictures; S500, obtain the first similarity between the common item component feature vector corresponding to each intermediate feature vector and the CG; S600, determine the intermediate feature vector corresponding to the first similarity that is greater than the first preset similarity threshold as the undetermined feature vector; S700 determines the target feature vector from the candidate feature vectors based on the provider of each material corresponding to each candidate feature vector, and pushes the item optimization plan corresponding to the target feature vector to the user.

[0004] According to another aspect of this application, a non-transitory computer-readable storage medium is also provided, wherein at least one instruction or at least one program is stored in the storage medium, and the at least one instruction or at least one program is loaded and executed by a processor to implement the above-described method for generating an optimized scheme for user-defined items.

[0005] According to another aspect of this application, an electronic device is also provided, including a processor and the aforementioned non-transitory computer-readable storage medium.

[0006] The present invention has at least the following beneficial effects: The method for generating optimized solutions for user-defined items in this invention effectively solves the problem of material mismatch caused by relying on fixed rules in the prior art by introducing a coordination judgment model to intelligently screen material combinations. By extracting and calculating the similarity of feature vectors of the original item and the reference item on common components, it achieves accurate understanding and deep integration of the user's reference intent at the component level, overcoming the limitations of existing solutions that only imitate the overall style or extract a single attribute. At the same time, the material provider (supplier) factor is incorporated into the final solution decision to ensure that the generated optimized solution has supply chain feasibility and production implementation, comprehensively improving the aesthetic quality, user intent matching, and actual executability of the optimized solution for user-defined items. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 A flowchart illustrating a method for generating optimization schemes for user-defined items, as provided in an embodiment of the present invention. Detailed Implementation

[0009] 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, and 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.

[0010] It should be noted that, based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Furthermore, this device and / or practice the method can be implemented using other structures and / or functionalities besides one or more of the aspects set forth herein.

[0011] The following will refer to Figure 1 The flowchart shown illustrates a method for generating optimization schemes for user-defined items, introducing such a method.

[0012] S100: If the item images uploaded by the user are an original item image and a reference item image, then each item component in the original item image and the reference item image is segmented to obtain several item components corresponding to the original item image and the reference item image.

[0013] In this embodiment, a semantic segmentation model based on deep learning can be used to segment item components. First, a dedicated component segmentation model is constructed and trained according to the target item category (such as clothing or furniture). Taking clothing as an example, the training data needs to contain a large number of labeled clothing images, where each pixel is labeled with its corresponding component category, such as "left sleeve," "right sleeve," "collar," "front," and "pocket." The model typically uses an encoder-decoder structure (such as U-Net or DeepLabV3+). Its encoder (such as ResNet or EfficientNet) is responsible for extracting deep features from the image, and the decoder is responsible for progressively upsampling the feature map and restoring spatial details. Finally, it outputs a segmentation mask map of the same size as the input image, where the value of each pixel is the label of its component category.

[0014] After a user uploads an image of the original item and an image of a reference item, the system inputs both images into the trained segmentation model. The model outputs the corresponding segmentation results, namely, the "part mask images" of each of the two images. Based on the mask images, the system extracts the pixel regions belonging to the same part label, thereby obtaining several independent image blocks of the item parts corresponding to the original item and the reference item.

[0015] For example: A user uploads an original image of a "standard shirt" and a reference image of a "silk dress shirt". The segmentation model identifies the components in the original image as: [standard collar, long sleeves, front placket, pockets]; and the components in the reference image as: [wing collar, French cuffs, front placket, decorative buttons].

[0016] The original item image is the base image of the item the user wants to modify, optimize, or redesign. It represents the basic structure, functional form, and core components the user wants to retain. The reference item image is an image of an item provided by the user as inspiration for an ideal style, texture, or design detail. It expresses the aesthetic direction, feel, or certain characteristics the user hopes the final optimized solution will present. The entire optimization process is carried out without changing the basic framework of the original item's components. This ensures that the optimized result is controllable and feasible in terms of functionality, fit, and manufacturability. The role of the reference item is to provide a set of aesthetic attribute targets. The system does not directly copy its structure (e.g., it doesn't forcibly add parts of a coat to a shirt), but rather learns its aesthetic characteristics (such as the sheen of silk) and seeks materials that can achieve similar effects on the original item's structure.

[0017] This step transforms unstructured image information into structured, independently processable component units, providing precise operational objects for subsequent material property analysis and replacement of each component. It forms the basis for the entire solution to achieve component-level optimization and integration.

[0018] S200: Based on each preset attribute feature of each optional material corresponding to each part of the original item image, obtain several candidate feature vectors corresponding to the original item image.

[0019] The preset attribute features include the physical attribute features and aesthetic attribute features of the material; the physical attribute features include at least one of material type, thickness, and elastic modulus, and the aesthetic attribute features include at least one of texture, gloss, and color space value.

[0020] This step is based on a structured materials database. In the database, each item component corresponds to several selectable materials, and each material (such as "pure cotton twill," "silk satin," and "calfskin") has a predefined and quantified set of multi-dimensional "preset attribute characteristics." These characteristics can be categorized as follows: Physical properties: such as material type (encoding vector), weight (numerical value), elongation (numerical value), stiffness and flexibility (numerical value).

[0021] Aesthetic attributes include texture (vectors obtained through Gabor filters or deep learning texture descriptors), gloss (numerical value), primary color (Lab color space value), and pattern style (style encoding vectors such as "checkered" or "floral" patterns).

[0022] For each component (e.g., "left sleeve") obtained from the segmentation of the original item image, the system selects a "list of optional materials" from the material database based on the component type and the user's possible preferences (or default settings). Then, the system performs material combination enumeration: for each component of the original item, it selects one material from its list of optional materials to form a complete "material combination scheme." For each material combination scheme, the system concatenates or weights the preset attribute feature vectors of all selected materials for all components according to a fixed component order, generating a global, high-dimensional "candidate feature vector." This vector uniquely represents the overall attributes of this material combination.

[0023] For example, if there are 3 material options for the "collar", 4 material options for the "sleeves", and 5 material options for the "body", then theoretically there can be 3×4×5=60 material combinations, which will generate 60 candidate feature vectors.

[0024] In this embodiment, to ensure that all candidate feature vectors have a unified mathematical expression form to facilitate subsequent model processing and comparison, the system adopts a predefined global component index and vector padding mechanism to solve the dimensionality unification problem, as follows: The system predefines a complete and standardized global component list for a certain type of item (such as "shirt" or "chair") and assigns a fixed index position to each component.

[0025] For example, for the "Shirt" class, the global list might be: [index 0: collar, index 1: left sleeve, index 2: right sleeve, index 3: front placket, index 4: back panel, index 5: pocket, ...].

[0026] After segmenting the original item image, the system matches the identified components with a global component list. For components present in the original item, their corresponding index positions are marked as "valid"; for components that do not exist (such as a shirt without pockets), their corresponding index positions are marked as "missing" or "empty". This forms a binary existence mask of the same length as the global list.

[0027] The system initializes a zero-valued or special null-valued marker vector with a length equal to (total number of global parts) × (material attribute feature dimension of each part). When generating candidate feature vectors for a specific material combination, the system only fills the corresponding index interval of the global list with the attribute feature vectors of the selected materials for each part. For parts that do not exist in the original item (masked as "empty"), their corresponding index intervals retain the initialized zero or null values. Ultimately, each candidate feature vector is a fixed-length vector, where some segments contain meaningful material attribute data, and the remaining segments are placeholders.

[0028] This step abstracts physical materials into computable data vectors and systematically enumerates all possible material replacement schemes, providing a comprehensive candidate solution space for subsequent intelligent screening and decision-making.

[0029] Furthermore, step S200 includes the following steps: S210, for any item component RA in the original item image, select a material from the optional materials corresponding to RA.

[0030] For each item component (denoted as component RA) obtained by segmenting the original item image using S100, the system selects a material from its associated set of optional materials.

[0031] By exhaustively selecting all possible combinations of components, the system generates all possible material combinations. Each complete combination constitutes a candidate material combination.

[0032] S220 combines the preset attribute features corresponding to the materials selected for each component to generate a candidate feature vector representing the overall material combination; wherein, the combination method is vector concatenation or weighted fusion based on component weights.

[0033] For each candidate material combination generated by S210, the system needs to construct a fixed-dimensional mathematical representation for it, namely the candidate feature vector.

[0034] The vector is constructed as follows: a. Initialize a uniform template vector: The system pre-creates a fixed-length zero-value (or null-value) vector template based on the global component index list for this item category. The total dimension of this vector is determined by (total number of components) × (material attribute feature dimension of each component).

[0035] b. Feature extraction and filling: For each component in the current candidate material combination, extract its preset attribute features (such as physical and aesthetic attribute vectors) from the selected material.

[0036] c. Combining and Filling: Fill or set the attribute feature vector of this component into a specific dimension of the unified template vector corresponding to the global index of this component.

[0037] Vector concatenation: Directly concatenate the feature sub-vectors of all components according to the global index order to form a long vector. For components that do not exist in the original item, their corresponding dimensions are filled with zero vectors.

[0038] Ultimately, each candidate material combination is transformed into a candidate feature vector with uniform dimensions, thus providing a standardized input for all subsequent vector-based computations and model processing.

[0039] Step S200, through the systematic enumeration in S210, ensures the completeness of the solution search; through the standardized vector construction in S220, especially combined with the global component indexing and filling mechanism, it solves the problem of inconsistent input dimensions caused by different component compositions, providing technical feasibility assurance for downstream coordination model judgment and accurate intent similarity comparison. The vector concatenation method preserves the complete semantic structure of the components, while the weighted fusion method provides a more compact feature representation, which can be selected and applied according to the overall system architecture.

[0040] S300: Input each candidate feature vector into the coordination judgment model corresponding to the original item to determine whether the material combination corresponding to each candidate feature vector is coordinated, and determine the candidate feature vector that is coordinated as the intermediate feature vector.

[0041] Furthermore, the coordination judgment model is the CLIP model.

[0042] In this embodiment, the coordination judgment model can be trained using the following method: 1. Training data construction: Positive Samples (Harmony): Collect a large number of images of finished products with recognized excellent design and aesthetic harmony (such as those from designer brands or award-winning works). For each image, through professional annotation or reverse engineering, determine the materials used in each component, and generate a corresponding material combination feature vector (i.e., the vector form defined in S200) based on a material database. In this way, each harmonious product image is equipped with a feature vector "describing its constituent materials".

[0043] Negative samples (inconsistency): This is crucial for teaching the model to identify "inconsistency." Negative samples are primarily synthesized artificially in two ways: Random Combination: Randomly select materials for each component of the item from the material library (ensuring that these materials rarely or never appear together in the actual design), generate their feature vectors, and pair them with a visually discordant image of the item generated by digital rendering of these materials, or simply label it as "incongruous".

[0044] Style Conflict Combination: Intentionally select materials with conflicting styles, eras, or textures to combine (such as pairing "silk" with "industrial studded leather") to generate feature vectors and corresponding conflict style renderings.

[0045] 2. Model architecture modification and training objectives: Architecture: The dual encoder structure of CLIP is retained, but key modifications are made.

[0046] Image Encoder: Directly uses a CLIP pre-trained visual encoder (such as ViT) and freezes its parameters or performs lightweight fine-tuning. Its function is to encode the input object image into a fixed-dimensional visual feature vector V. I .

[0047] Text encoder replacement: The CLIP's original text encoder is replaced with a completely new material feature encoder (e.g., a multilayer perceptron, MLP). This encoder takes a material combination feature vector as input and outputs a vector that is analogous to the visual feature vector V. I Feature vectors t in the same semantic space m .

[0048] Training objective (contrastive learning): The training objective is to enable the model to learn to narrow the gap between harmonious "image-text pairs" and widen the gap between inharmonious "image-text pairs".

[0049] A training batch contains N pairs of (image, material vector) samples, including both positive and negative samples. The cosine similarity between all image features and all material features within the batch is calculated, forming an N×N similarity matrix.

[0050] Loss function: Use contrastive loss (such as InfoNCE Loss). For any image, its paired positive sample material vector should be identified as the best-matching description (high similarity), while all other unpaired material vectors within the batch (which may describe other items or incongruous combinations) should be judged as mismatched (low similarity). Conversely, the same applies to material vectors.

[0051] 3. Model output and consistency assessment: Once training is complete, the model will have cross-modal matching capabilities.

[0052] Input: A candidate feature vector generated by S200.

[0053] Inference process: The vector is input into the trained material feature encoder to obtain its semantic space projection. The model internally (or through an additional lightweight classification head) evaluates: "If an item is made of the material described by the semantic space projection, its visual features should be highly similar to the visual features of a 'coordinated item'." Finally, a coordination score is output (i.e., the degree of matching between the semantic space projection and the model's internal "coordinated" visual concept prototype).

[0054] Decision: By comparing the score with a preset threshold, it can be determined whether the candidate material combination is "coordinated".

[0055] The beneficial effects of the coordination judgment model trained in the above manner are: Knowledge transfer and efficient learning: It fully utilizes the powerful general visual concept understanding capabilities (such as texture, shape, and style) learned by the CLIP pre-trained model from massive amounts of Internet data. It can quickly obtain professional-level aesthetic judgment by only making relatively minor adjustments on the "coherence" data in a specific domain, avoiding the massive data and computational costs required to train a model from scratch.

[0056] It achieves precise alignment across modalities: it successfully maps abstract, structured "material property vectors" into a semantic space common to visual features, enabling quantitative comparison between "material description" and "overall visual perception".

[0057] It ensures the objectivity and consistency of the judgment: The model's judgment is based on data-driven principles, avoiding the subjectivity and limitations of manual rules, and can stably and efficiently process massive candidate combinations. This is the core technical guarantee for achieving full-solution automation.

[0058] S400, obtain the common item component feature vector corresponding to each intermediate feature vector and the common item component feature vector CG corresponding to the reference item image; the common item component feature vector is obtained based on each preset attribute feature of each optional material of the common item component corresponding to the original item image and the reference item image.

[0059] Furthermore, step S400 includes the following steps: S410, Obtain the common item components between the original item image and the reference item image.

[0060] Input: The original item image and the reference item image are divided into a list of parts by S100 and their corresponding semantic tags (such as "left sleeve" and "stand-up collar").

[0061] Matching method: The system performs precise matching or similarity matching based on the semantic tags of the components (such as using word vectors to calculate tag similarity) to find components with the same or highly similar semantics in two images and identify them as "shared item components".

[0062] Example: Although the "standard collar" of the original shirt and the "wing collar" of the reference shirt have different shapes, they both belong to the category of "collar" and can be identified as common components.

[0063] Output: A clear "list of common parts", such as ['collar', 'front'].

[0064] S420, for any intermediate feature vector ZA, extract the preset attribute features corresponding to the materials belonging to the common item components in the material combination corresponding to ZA, and construct the common item component feature vector corresponding to ZA.

[0065] Input: Any intermediate feature vector ZA that is determined to be "harmonious" in S300. This vector is a complete and unified vector that represents a material combination scheme for all components of the original item.

[0066] Extraction process: a. Based on the list of common components obtained in S410 and the global component indexing rules used in S200, the system determines the dimension segment interval corresponding to each common component in the unified vector ZA.

[0067] b. The system precisely slices these dimensions from the ZA, and each dimension contains all the preset property characteristics of the material selected for the component.

[0068] c. Reassemble or combine these sliced ​​feature vectors according to the same component order as S410 to form a new vector with a relatively smaller dimension. This new vector is the common item component feature vector corresponding to ZA. It only describes the material properties of the common components and is unrelated to the non-common components of the original item (such as unique pockets) and the non-common components of the reference item (such as unique decorations).

[0069] Example: Suppose ZA is a vector describing the entire shirt, with common components ['collar', 'front placket']. This step will extract the material features belonging to the collar and front placket from ZA to form a new vector specifically describing the "collar and front placket material".

[0070] S430, extract the visual features GA belonging to the common item parts in the reference item image, and map the GA to the same feature space as the preset attribute features to form CG.

[0071] For each common component identified by S410, the system uses a pre-trained visual encoder (such as a CNN) to process the image patch of that component in the reference image and extract its appearance visual feature vector. These features describe visual information such as the component's texture, color, and gloss.

[0072] The extracted visual features (GA) and material property features are usually not in the same mathematical space and cannot be directly compared. Therefore, a mapping model (which could be a neural network, such as a multilayer perceptron (MLP)) is needed to non-linearly map the GA to a feature space that is exactly the same as the material's "pre-defined property features". This mapping model can be trained in advance, and its goal is to make the mapped vector understandable by subsequent similarity calculations and reconciliation models, just like the material property vectors.

[0073] The model is trained using a material-image pairing dataset. This dataset contains sample images of various materials and their corresponding normalized material property feature vectors. The training objective is to enable the mapping model to learn to predict (or approximate) the normalized property features of materials from their images.

[0074] The vector obtained after mapping is the common item component feature vector CG corresponding to the reference item image. CG has the exact same dimension and physical meaning as the vector extracted from S420.

[0075] The above steps are the core of achieving intent quantification. Through feature deconstruction and cross-modal mapping, it transforms the user's vague reference intent (an image) into a clear, quantifiable mathematical target (CG vector). Using this target as a benchmark, it accurately measures all coordination solutions that pass the initial aesthetic screening, ensuring that the final recommended solution highly matches the user's source of inspiration in key details.

[0076] S500, obtain the first similarity between the common item component feature vector corresponding to each intermediate feature vector and the CG.

[0077] For each intermediate feature vector corresponding to a common component feature vector, calculate its "first similarity" with the common component feature vector CG of the reference image. This can be calculated using cosine similarity or the reciprocal of the Euclidean distance to measure their closeness in the material property feature space.

[0078] S600, determine the intermediate feature vector corresponding to the first similarity that is greater than the first preset similarity threshold as the undetermined feature vector.

[0079] In this step, a "first preset similarity threshold" (e.g., 0.85) is set. The system filters out all intermediate feature vectors with a first similarity greater than this threshold and identifies them as "feature vectors to be determined." This means that these material combinations are not only harmonious internally, but also that the material style of their common components highly matches the intent of the reference images provided by the user.

[0080] For example, the reference image shows that the collar and body of the garment have a "high-gloss silk texture". If the feature vectors of common parts (collar, body) of a feature vector to be determined have a high similarity to CG, it means that the scheme has selected materials with gloss and texture very close to silk for these parts.

[0081] The first preset similarity threshold can be set using the following method: Collect a pre-labeled "reference image-coordination scheme" pairing test set covering multiple item categories. For each pairing in the test set, calculate the similarity between its common component feature vector and the reference image feature vector CG, obtaining a set of similarity score distributions.

[0082] Set the threshold at a higher quantile of the score distribution. For example, you could take the 75th percentile or mean of all pair similarity scores plus a standard deviation. This means that only candidates that match the reference intent better than most historical "good solutions" will advance to the next round of screening.

[0083] For example, on the clothing category test set, the similarity scores of the "good solutions" were mainly distributed between 0.6 and 0.95, with an average of 0.8. The first preset similarity threshold can be initially set to 0.85 to ensure that the selected solutions belong to the higher matching range.

[0084] This step transforms the user's abstract "reference intent" into a quantifiable similarity metric, and achieves precise control through threshold screening. This ensures that the final shortlisted solutions are not only aesthetically consistent but also highly aligned with the user's desired style inspiration, effectively solving the problem of the solution being out of touch with the user's intent.

[0085] S700 determines the target feature vector from the candidate feature vectors based on the provider of each material corresponding to each candidate feature vector, and pushes the item optimization plan corresponding to the target feature vector to the user.

[0086] Furthermore, step S700 includes the following steps: S710, obtain the total number of different material providers in the material combination corresponding to each undetermined feature vector, and obtain the list of total providers NUM = (NUM1, NUM2, ..., NUM...). i ..., NUM n ), i=1,2,…,n; NUM i Let be the total number of different material providers in the material combination corresponding to the i-th undetermined feature vector; n is the number of undetermined feature vectors.

[0087] For each "undetermined feature vector" (representing a material combination scheme that is both harmonious and in line with the user's style intent) selected in S600, the system queries the "provider" (supplier) associated with each specific material.

[0088] The total number of different providers in this scheme is denoted as NUM. i (For the i-th option). This value directly reflects the number of suppliers that need to be coordinated to produce this option, and is one of the core quantitative indicators of supply chain management complexity.

[0089] At the same time, the system counts the quantity of different material types used in the scheme, denoted as C. iHere, "material type" refers to a category based on the physical and aesthetic properties of the material (such as "silk satin," "lambskin," and "denim"), and is independent of the supplier. Multiple components may use the same material type (but may come from different suppliers). Ultimately, the system obtains a list of supply chain metrics, NUM.

[0090] S720, iterate through NUM, if NUM i / C i If < γ, then the i-th undetermined feature vector is determined as the target feature vector; C i γ represents the number of material types corresponding to the i-th undetermined feature vector; γ is the pre-proportional threshold.

[0091] The system calculates the "provider concentration" ratio, or NUM, for each option. i / C i This ratio represents the average number of independent suppliers for each type of material.

[0092] Set a pre-proportional threshold γ (usually 0 < γ ≤ 1). The system iterates through all schemes and filters out those that satisfy NUM. i / C i The scheme with <γ is identified and determined as the target feature vector.

[0093] NUM i / C i =1: This means that each type of material is supplied by exactly one independent supplier, the supply chain structure is clear but the number of suppliers is as many as the types of materials.

[0094] NUM i / C i <1: This means that at least one supplier provides multiple material types, the supply chain is integrated, and the total number of suppliers is less than the total number of material types.

[0095] NUM i / C i >1: This means that at least one type of material is supplied by multiple suppliers (for example, the same "cotton fabric" is supplied by two suppliers), which usually increases the difficulty of supply chain coordination and quality control risks.

[0096] Conditional NUM i / C i <γ aims to screen for solutions with high supplier integration and a more streamlined supply chain structure. The smaller the γ value, the stricter the screening criteria and the higher the required supplier integration (one supplier handles the supply of more material types).

[0097] Compared to simply choosing the option with the fewest total number of suppliers, this method, by introducing a "provider concentration" ratio, can more intelligently assess the actual management efficiency of the supply chain. A solution with a slightly larger number of suppliers but high concentration (one core supplier provides most of the materials) may be easier to manage and coordinate than a solution with fewer suppliers but operating independently. This reflects a shift in thinking from simply pursuing the minimization of quantity to pursuing the optimization of structure.

[0098] The pre-proportional threshold γ is not a fixed value; its appropriate setting can be determined by referring to historical data (such as the average supplier concentration of past successful orders), business strategy (whether the supply chain is more concentrated or more diversified), or through A / B testing. This provides flexibility for adapting the solution to different product categories.

[0099] When multiple solutions meet the threshold condition, NUM can be further selected. i The smallest option, or when pushing out the solution, should explain the supply chain advantages of each option to the user. This rule provides a clear, data-driven explanation of "why this option is chosen" (e.g., "this option uses fewer core suppliers for all materials, and is expected to have higher production coordination efficiency and more stable delivery").

[0100] This step deeply couples design feasibility with production feasibility by quantitatively evaluating the supply chain structure of material combination solutions. It focuses not only on "whether it can be made," but also on "whether it can be made efficiently, stably, and economically." By introducing a supplier concentration threshold for screening, the system can automatically recommend solutions that excel in aesthetics, intent matching, and supply chain practicality. This significantly improves the actual success rate of recommended solutions, reduces users' subsequent production coordination costs, and enhances the platform's professionalism and user trust. This marks a key upgrade in solution generation from a "design tool" to a "one-stop production solution."

[0101] Furthermore, prior to step S100, the method further includes the following steps: S010, if the item image uploaded by the user is the original item image, then each item component in the original item image is segmented to obtain several item components corresponding to the original item image.

[0102] This step is completely consistent with the processing of the original item image in S100. The system calls the same set of deep learning-based component semantic segmentation models (such as U-Net and Mask R-CNN). The model receives the original item image uploaded by the user and outputs pixel-level segmentation masks, thereby deconstructing the image into several independent, semantically meaningful item component image blocks.

[0103] Example: A user uploads a picture of an old sofa, and the system segments it into components such as [seat cushion, backrest, armrests, and sofa legs]. Reusing the S100 segmentation model ensures consistency in system input processing and reduces maintenance costs.

[0104] S020, extract attribute features for each item component corresponding to the original item image to obtain the original item feature vector XL1 corresponding to the original item image.

[0105] For each part image patch obtained in S010, a pre-trained convolutional neural network (CNN) encoder is used to extract its visual feature vector. This vector captures the part's texture, color, gloss, and other appearance information. The extracted visual feature vector is then input into a visual feature-to-material property space mapping model (this model is the same or isomorphic model as the mapping model used to generate CG in S430). This model is trained to map visual features into a semantic space isomorphic to the "preset property features" in the material database.

[0106] The attribute feature vectors obtained after mapping all components are combined according to the same global component index and combination rules (such as splicing) as S200 to finally generate a complete "original item feature vector XL1". XL1 mathematically accurately represents the comprehensive properties of the materials of the current components of the original item.

[0107] Example: An old sofa XL1 might be represented as a combined vector of: [(seat cushion: worn velvet feature), (backrest: same velvet feature), (armrest: wood feature), (sofa legs: metal feature)].

[0108] The generation of XL1 enables precise quantification of the original item's current state, providing an objective and calculable benchmark for subsequent quantification of the "degree of change".

[0109] S030, based on each preset attribute feature of each optional material corresponding to each part of the original item image, obtain several candidate feature vectors corresponding to the original item image.

[0110] The system enumerates all possible combinations from its optional material library for each component of the original item, generating a "candidate feature vector" with uniform dimensions for each combination. The method is the same as S200.

[0111] S040, each candidate feature vector is input into the coordination judgment model corresponding to the original item to determine whether the material combination corresponding to each candidate feature vector is coordinated, and the candidate feature vector with the judgment result of coordination is determined as the intermediate feature vector.

[0112] All candidate feature vectors are input into the trained "coordination judgment model" (i.e., a CLIP-based fine-tuned model). This model outputs a coordination score for each vector. The system selects all "coordinated" solutions based on a preset threshold and determines their feature vectors as "intermediate feature vectors." This step completely reuses the core algorithm module from the above embodiments, ensuring the basic quality (coordination) of the generated solutions.

[0113] S050, obtain the second similarity between each intermediate feature vector and XL1.

[0114] For each "intermediate feature vector", the system calculates its second similarity with XL1 in the feature space. Cosine similarity can be used for calculation.

[0115] The second similarity ranges from [-1, 1]. The closer the second similarity is to 1, the more similar the optimized solution is to the original item in terms of material properties, and the more conservative the changes are; the lower the value, the greater the degree of innovation.

[0116] Example: One solution changes the sofa fabric from "velvet" to "corduroy," which has similar visual properties, and its second similarity with XL1 may be as high as 0.92; while another solution changes the fabric to "leather," which has a completely different luster, and its second similarity may only be 0.45.

[0117] This step transforms the abstract concept of "how much change there is" into a precisely measurable numerical value, providing a data foundation for intelligent filtering based on user preferences.

[0118] S060, the intermediate feature vector corresponding to the second similarity in the similarity interval is determined as the undetermined feature vector, and then proceed to S700; the similarity interval is obtained through user input.

[0119] The system provides users with intuitive interactive controls (e.g., a slider labeled “Conservative Fine-tuning”, “Style Innovation”, “Bold Reshaping”, or direct input of a numerical range) to allow users to specify a desired “similarity interval” (e.g., [0.7, 0.9]).

[0120] The system iterates through all intermediate feature vectors and compares their second similarity with the user-defined interval. All intermediate feature vectors whose second similarity falls within the specified interval are identified as "pending feature vectors" that will proceed to the final decision-making stage.

[0121] The selected set of undetermined feature vectors will directly jump to step S700 of the main process for final selection and recommendation based on the supply chain provider.

[0122] Example: If a user selects "moderate innovation," the corresponding system range is [0.5, 0.75]. Only coordination solutions with a second similarity within this range (i.e., moderate variation) will proceed to the final supplier selection stage.

[0123] This step empowers users with direct control over the direction and intensity of optimization, achieving a perfect blend of AI-generated results and human guidance. The system is no longer a black-box, random recommendation, but rather a collaborative design partner guided by user intent, significantly enhancing user engagement, controllability, and ultimate satisfaction.

[0124] The S010-S060 supplementary processes, together with the original S100-S700 main processes, constitute a complete, flexible, and user-friendly item optimization solution generation system. This supplementary process effectively addresses users' optimization needs when there are no clear reference targets. By extracting the original item feature vector XL1 and introducing user-defined similarity intervals, it successfully transforms the fuzzy task of "autonomous optimization" into a clear and computable problem of "finding the optimal coordination solution within a user-specified range of variation." It not only fully utilizes and reuses all the core technologies and data modules of the main process (segmentation model, coordination model, supply chain database), maintaining a simple and efficient system architecture, but also deeply integrates user preferences into the algorithm process through innovative interactive design, realizing a shift from "the system recommends what users see" to "users guide the system to explore what." Ultimately, all solutions undergo unified supply chain feasibility (S700) checks, ensuring end-to-end quality from idea to implementation, significantly improving the practical value and commercial conversion efficiency of the solutions.

[0125] Furthermore, although the steps of the method in this disclosure are described in a specific order in the accompanying drawings, this does not require or imply that the steps must be performed in that specific order, or that all the steps shown must be performed to achieve the desired result. Additional or alternative steps may be omitted, multiple steps may be combined into one step, and / or a step may be broken down into multiple steps.

[0126] Embodiments of the present invention also provide a non-transitory computer-readable storage medium that can be disposed in an electronic device to store at least one instruction or at least one program related to implementing a method in the method embodiments, wherein the at least one instruction or the at least one program is loaded and executed by the processor to implement the method provided in the above embodiments.

[0127] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0128] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting programs for use by or in conjunction with an instruction execution system, apparatus, or device.

[0129] The program code contained on the readable medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0130] Program code for performing the operations of this application can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0131] Embodiments of the present invention also provide an electronic device, including a processor and the aforementioned non-transitory computer-readable storage medium.

[0132] The electronic device is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments in this application.

[0133] Electronic devices are manifested in the form of general-purpose computing devices. Components of an electronic device may include, but are not limited to: at least one processor, at least one memory, and a bus connecting different system components (including memory and processor).

[0134] The memory stores program code that can be executed by the processor, causing the processor to perform the steps in the various embodiments described in this specification.

[0135] The memory may include readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory, and may further include read-only memory (ROM).

[0136] The memory may also include programs / utilities having a set (at least one) of program modules, including but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0137] A bus can represent one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus that uses any of the various bus structures.

[0138] Electronic devices can also communicate with one or more external devices (e.g., keyboards, pointing devices, Bluetooth devices, etc.), one or more devices that enable user interaction with the electronic device, and / or any device that enables the electronic device to communicate with one or more other computing devices (e.g., routers, modems, etc.). This communication can be achieved through input / output (I / O) interfaces. Furthermore, electronic devices can communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via network adapters. The network adapter communicates with other modules of the electronic device via a bus. It should be understood that other hardware and / or software modules can be used in conjunction with the electronic device, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0139] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, terminal device, or network device, etc.) to execute the methods according to the embodiments of this disclosure.

[0140] Embodiments of the present invention also provide a computer program product including program code, which, when the program product is run on an electronic device, causes the electronic device to perform the steps of the methods described above in various exemplary embodiments of the present invention.

[0141] While specific embodiments of the invention have been described in detail by way of examples, those skilled in the art should understand that the examples are for illustrative purposes only and are not intended to limit the scope of the invention. Those skilled in the art should also understand that various modifications can be made to the embodiments without departing from the scope and spirit of the invention.

Claims

1. A method for generating an optimization scheme of a user-defined article, characterized by, The method comprises the following steps: S100, if the uploaded item picture is an original item picture and a reference item picture, segmenting each item part in the original item picture and the reference item picture to obtain a plurality of item parts corresponding to the original item picture and the reference item picture; S200, obtaining a plurality of candidate feature vectors corresponding to the original item picture according to each preset attribute feature of each selectable material corresponding to each item part of the original item picture; S300, inputting each candidate feature vector into a coordination judgment model corresponding to the original item to determine whether the material combination corresponding to each candidate feature vector is coordinated, and determining the candidate feature vector with a judgment result of coordination as an intermediate feature vector; S400, obtaining a common item part feature vector corresponding to each intermediate feature vector and a common item part feature vector CG corresponding to the reference item picture; the common item part feature vector is obtained according to each preset attribute feature of each selectable material of the common item part corresponding to the original item picture and the reference item picture; S500, obtaining a first similarity between the common item part feature vector corresponding to each intermediate feature vector and CG; S600, determining an undetermined feature vector corresponding to the intermediate feature vector with a first similarity greater than a first preset similarity threshold; S700, determining a target feature vector from the undetermined feature vector according to a provider of each material corresponding to each undetermined feature vector, and pushing an item optimization scheme corresponding to the target feature vector to the user.

2. The method of claim 1, wherein, Step S200 comprises the following steps: S210, selecting a material from the selectable materials corresponding to RA for any item part RA in the original item picture; S220, combining the preset attribute features corresponding to the selected material for each part to generate a candidate feature vector representing the overall material combination; wherein the combination mode is vector splicing.

3. The method of claim 1, wherein, Step S400 comprises the following steps: S410, obtaining a common item part between the original item picture and the reference item picture; S420, for any intermediate feature vector ZA, extracting the preset attribute features corresponding to the materials belonging to the common item part in the material combination corresponding to ZA to form a common item part feature vector corresponding to ZA; S430, extracting the visual features GA belonging to the common item part in the reference item picture, and mapping GA to the same feature space as the preset attribute features to form CG.

4. The method of claim 1, wherein, Step S700 comprises the following steps: S710, obtaining the total number of different material providers in the material combination corresponding to each pending eigenvector, to obtain a provider total number list NUM= (NUM1, NUM2, …, NUM i , …, NUM n ), i=1, 2, …, n; NUM i is the total number of different material providers in the material combination corresponding to the ith pending eigenvector; n is the number of pending eigenvectors; S720, traversing NUM, if NUM i / C i < γ, the ith pending eigenvector is determined as the target eigenvector; C i is the number of material types corresponding to the ith pending eigenvector; and γ is a pre-proportion threshold.

5. The method of claim 1, wherein, Before step S100, the method further comprises the following steps: S010, if the uploaded item picture is an original item picture, segmenting each item part in the original item picture to obtain a plurality of item parts corresponding to the original item picture; S020, performing attribute feature extraction on each item part corresponding to the original item picture to obtain an original item feature vector XL1 corresponding to the original item picture; S030, obtaining a plurality of candidate feature vectors corresponding to the original item picture according to each preset attribute feature of each selectable material corresponding to each item part of the original item picture; S040, inputting each candidate feature vector to a coordination judgment model corresponding to the original item to judge whether the material combination corresponding to each candidate feature vector is coordinated, and determining the candidate feature vector with a judgment result of coordination as an intermediate feature vector; S050, obtaining a second similarity between each intermediate feature vector and XL1; S060, determining the intermediate feature vector corresponding to the second similarity belonging to the similarity interval as a pending feature vector, and entering S700; the similarity interval is obtained through user input.

6. The method of claim 1-5, wherein, The preset attribute features include physical attribute features and aesthetic attribute features of the material; the physical attribute features include at least one of a material type, a thickness, and an elastic modulus, and the aesthetic attribute features include at least one of a texture, a glossiness, and a color space value.

7. The method of claim 1, wherein, The coordination judgment model is a CLIP model. 8.A non-transitory computer readable storage medium having stored therein at least one instruction or at least one piece of program, characterized in that, The at least one instruction or the at least one program is loaded and executed by the processor to implement the optimization scheme generation method of the user-defined item according to any one of claims 1-7.

9. An electronic device, comprising: The non-transitory computer-readable storage medium of claim 8 is included in a processor. The non-transitory computer-readable storage medium of claim 8 is included in a processor.

Images