Textile upper intelligent design method and system based on large model

Abstract

This invention discloses a method and system for intelligent design of textile and footwear uppers based on a large model, relating to the field of artificial intelligence technology. The method includes: constructing an image-text multimodal dataset from original textile and footwear upper images; designing and executing a LoRA fine-tuning training workflow based on the multimodal dataset to obtain a textile and footwear upper generation model; generating a creative image of the footwear upper based on user-input text descriptions and evaluating its quality, including extracting traditional image quality features, AI-generated defect features, and quality scoring features obtained using a large multimodal model; weighted fusion of each feature to obtain a fusion score, and introducing a fatal defect correction factor to reduce the fusion score to obtain a final quality score; comparing the final quality score with a preset threshold to prompt or automatically trigger regeneration, outputting a qualified creative image of the footwear upper. This invention achieves efficient generation of textile and footwear upper patterns and a closed-loop iteration with controllable quality, improving design efficiency and shortening the product development cycle.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for intelligent design of textile and footwear uppers based on a large model. Background Technology

[0002] In the traditional textile and footwear manufacturing industry, the creative design of shoe upper patterns mainly relies on designers to complete the work by hand. Designers draw shoe upper patterns based on their design concepts. This creative process is often time-consuming and requires multiple iterations and revisions, which leads to longer product development cycles and increased labor costs for companies.

[0003] With the development of generative artificial intelligence technology, deep learning methods such as diffusion models can automatically generate images based on text descriptions, providing new possibilities for design automation. However, general generative models, trained on large-scale general data, often lack professionalism when directly applied to the textile and footwear pattern field, failing to meet the consistency requirements of actual production in terms of pattern detail and style. Images generated by existing general models may differ from the needs of the textile industry in terms of texture detail and color matching.

[0004] In the specific field of textile and footwear design, there is an urgent need for an intelligent design tool capable of quickly generating footwear patterns that meet industry requirements based on user-provided creative text. The LoRA (Low-Rank Adaptation) fine-tuning technology, which has emerged in recent years, can efficiently and orient the model to specific domain data while preserving existing knowledge of the large model, thus potentially generating images that better fit industry needs. Therefore, applying a large model combined with LoRA fine-tuning to the generation of creative textile and footwear upper designs, and implementing it with an interactive system, becomes a feasible way to improve the efficiency and quality of footwear upper design. Summary of the Invention

[0005] The purpose of this invention is to solve the problems in the prior art.

[0006] The technical solution adopted by this invention to solve its technical problem is: to provide a smart design method for textile footwear uppers based on a large model, comprising the following steps:

[0007] The original image data of textile and footwear uppers are preprocessed, and corresponding text description labels are written for each image to obtain a multimodal dataset for training; wherein, the preprocessing includes standardizing the images and adjusting the image resolution to a uniform pixel;

[0008] Design a LoRA fine-tuning training workflow for generating creative images of textile shoe uppers, and execute the LoRA model training workflow using a multimodal dataset to obtain a textile shoe upper generation model;

[0009] The aforementioned textile shoe upper generation model generates a creative shoe upper image based on the text description input by the user;

[0010] A textile shoe upper generation model is used to intelligently generate corresponding shoe upper creative images based on text descriptions. The shoe upper creative images are then evaluated by image scoring to obtain a final quality score.

[0011] The final quality score is compared with a preset threshold: when the final quality score is less than the preset threshold, the user is prompted to regenerate the shoe upper creative image or the shoe upper creative image is automatically regenerated; when the final quality score is not less than the preset threshold, a qualified shoe upper creative image is output.

[0012] The image scoring evaluation includes:

[0013] The creative image of the shoe upper is scaled to a fixed size to eliminate resolution differences, and the creative image of the shoe upper is normalized to obtain a first image; the first image is converted from RGB space to YCbCr space and Lab space to obtain a second image and a third image respectively; the first image is divided into multiple non-overlapping blocks to obtain several block images; at the same time, the RGB space image is retained for subsequent feature extraction of different types and local defect detection;

[0014] Multi-dimensional traditional image quality features are extracted based on the second image, the third image, and the block image;

[0015] Extract AI defect features from AI-generated images that have unique defects;

[0016] The first image is scored using a multimodal large model to obtain large model evaluation features; wherein, preset prompt words are input into the multimodal large model to make it ignore artistic style differences, focus only on objective quality defects, and output a quality score of 0-100, and the quality score is normalized and used as the large model evaluation features.

[0017] The traditional image quality features, the AI ​​defect features, and the large model evaluation features are normalized, and a hierarchical weighting strategy is used to fuse the features to obtain a fusion score. On this basis, a correction factor for handling fatal defects is introduced to obtain a final quality score, and the final quality score is output. The final quality score is obtained by applying a reduction processing determined by the correction factor to the fusion score.

[0018] Preferably, the traditional image quality features include at least sharpness features, noise and artifact features, contrast and dynamic range features, and color consistency features.

[0019] Preferably, the sharpness feature is obtained from the Y component of the second image and characterized based on the statistics of the Laplacian response and the edge density obtained by the Sobel operator;

[0020] The noise and artifact features are obtained from the block image and are characterized based on the mean standard deviation and mean entropy of the block image, combined with a preset threshold.

[0021] The contrast and dynamic range features are obtained from the Y component of the second image and characterized based on the quantile statistics of the grayscale image.

[0022] The color consistency feature is obtained from the third image and is characterized based on the mean and standard deviation of channel a and channel b, combined with a preset coefficient.

[0023] Preferably, the AI ​​defect features include at least semantic consistency features, detail authenticity features, and generation logic consistency features;

[0024] The semantic consistency feature is obtained based on RGB images and is characterized by at least contour smoothness and region compactness.

[0025] The detailed authenticity features are obtained based on the block image and are characterized based on the number of peaks in the block image and a preset coefficient.

[0026] The generated logic consistency feature is obtained based on the Y component of the second image and is characterized by at least the mean of the light and shadow direction deviation and the occlusion rationality score.

[0027] Preferably, the large model evaluation features are obtained through the following process: the first image is scored using a multimodal large model to obtain an image quality score of 0–100, and the image quality score is normalized to obtain the large model evaluation features; wherein, the preset prompt words are required to be evaluated from at least two dimensions: traditional image quality and AI-generated unique defects.

[0028] Preferably, the process of obtaining the final quality score includes the following steps:

[0029] The traditional image quality features, the AI ​​defect features, and the large model evaluation features are normalized, wherein the feature extrema used for normalization are the feature extrema on a preset high-quality AI-generated image dataset; and the normalization includes linear normalization or inverse normalization to ensure that the features fall within the [0,1] interval. The normalization formula is as follows:

[0030] ;

[0031] in, , For the feature extrema from a high-quality AI-generated image dataset; This represents the value before normalization. Represents a normalized value;

[0032] A hierarchical weighted strategy is used to calculate the fusion score, dividing the features into four layers: traditional image quality layer, AI-specific defect layer, semantic logic layer, and large model scoring layer, with corresponding weights assigned to each. The fusion score is obtained by multiplying the scores of each of the four layers by their respective weights and then summing the results, as shown in the formula below:

[0033] ;

[0034] in, This represents the score of the feature at the i-th layer. Indicates correspondence The weight, Indicates the fusion score;

[0035] A correction factor is introduced to handle fatal defects: when the semantic consistency feature is lower than the first preset threshold or the generation logic consistency feature is lower than the second preset threshold, the correction factor takes a preset value less than 1; otherwise, the correction factor takes 1.

[0036] The final quality score is calculated based on the fusion score and the correction factor. The design image for the shoe upper is deemed unqualified if the final quality score is less than a preset threshold, and the formula is as follows:

[0037] ;

[0038] Among them, the correction factor When a fatal defect is triggered, a preset value less than 1 is used; otherwise, 1 is used.

[0039] Preferably, the method further includes: manually adjusting the shoe upper creative image generated by the large model, and then sending the adjusted shoe upper creative image back into the textile shoe upper generation model for secondary generation or optimization.

[0040] Preferably, it also includes: using a multimodal large model to analyze the currently generated image and existing guide words and parameters, and providing new guide words and parameters based on user needs.

[0041] Preferably, the textile and shoe upper generation model and the image scoring evaluation are encapsulated as an interface; and text input, generation parameter adjustment, task progress monitoring and generation result display are realized through the front-end interface, and the generation progress and results are returned to the front-end through the back-end service calling the interface.

[0042] This invention also provides a large-scale model-based intelligent design system for textile and footwear uppers, comprising:

[0043] The dataset creation module is used to preprocess the original image data of textile and footwear uppers and write corresponding text description labels for each image to obtain a multimodal dataset for training; wherein, the preprocessing includes standardizing the images and adjusting the image resolution to a uniform pixel;

[0044] The generative model training module is used to design a LoRA fine-tuning training workflow for generating creative images of textile and footwear uppers, and to execute the LoRA model training workflow using the multimodal dataset to obtain a textile and footwear upper generation model.

[0045] The shoe upper image generation module is used to generate a creative shoe upper image by calling the textile shoe upper generation model based on the text description input by the user;

[0046] An image scoring module is used to score and evaluate the creative image of the shoe upper. The image scoring module includes: an image preprocessing submodule, a traditional image quality feature extraction submodule, an AI defect feature extraction submodule, a multimodal large model scoring submodule, and a hierarchical weighted fusion and fatal defect correction submodule, and outputs a final quality score. Specifically, the image preprocessing submodule scales the image to a fixed size to eliminate resolution differences, converts the image from RGB space to YCbCr space and Lab space, and performs block processing, while retaining the RGB space image for subsequent feature extraction and local defect detection. The hierarchical weighted fusion and fatal defect correction submodule divides the features into four layers, performs weighted summation to obtain a fusion score, and reduces the fusion score through a correction factor when a fatal defect is triggered to obtain the final quality score.

[0047] The iterative control module is used to compare the final quality score with a preset threshold, and prompt to regenerate the shoe upper creative image or automatically regenerate the shoe upper creative image when the final quality score is less than the preset threshold, and output a qualified shoe upper creative image when the final quality score is not less than the preset threshold.

[0048] The present invention has the following beneficial effects:

[0049] (1) High degree of design automation: It realizes the automatic generation of textile shoe upper creative drawings from text descriptions, reducing the workload of manual drawing and reducing the dependence on designer experience.

[0050] (2) Improve efficiency and shorten cycle: The process of automatically generating patterns by machine only takes tens of seconds to a few minutes, which is much faster than manual design. It can significantly improve design efficiency and shorten the development cycle of products from ideas to samples.

[0051] (3) Professionalism and consistency: LoRA fine-tuning of the large model was performed using a professional image dataset of textile and footwear, so that the style of the generated footwear pattern is more in line with industry standards, ensuring the professionalism of the pattern details and the consistency of style between different outputs.

[0052] (4) Real-time interactive experience: The front-end and back-end collaboration provides a user-friendly interface and a real-time feedback mechanism. Users can view the generation progress and results in real time, and adjust the input as needed to iteratively optimize the design scheme, which improves the interactivity and convenience of the design process.

[0053] (5) Easy to deploy and integrate: The system supports local deployment, and the data does not leave the enterprise, ensuring the security of design data. The system architecture is modular and clear, and can be applied to fields such as textile and footwear manufacturing and apparel design, providing an innovative technical solution for the digital transformation of the traditional textile industry.

[0054] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating the steps of a smart design method for textile and footwear uppers based on a large model, according to an embodiment of the present invention.

[0056] Figure 2 This is a flowchart illustrating an intelligent design method for textile footwear uppers based on a large model, according to an embodiment of the present invention.

[0057] Figure 3 This is a flowchart illustrating the operation of an intelligent design method for textile and footwear uppers based on a large model, according to an embodiment of the present invention.

[0058] Figure 4 This is a flowchart illustrating the execution of a parameter guidance module in an intelligent design method for textile and footwear uppers based on a large model, according to an embodiment of the present invention.

[0059] Figure 5 This is a flowchart illustrating the execution of an image scoring module in an intelligent design method for textile and footwear uppers based on a large model, as described in an embodiment of the present invention.

[0060] Figure 6 This is a visualization process diagram of a smart design method for textile footwear uppers based on a large model, according to an embodiment of the present invention.

[0061] Figure 7 This is a result diagram of a smart design method for textile footwear uppers based on a large model, according to an embodiment of the present invention.

[0062] Figure 8 This is a system structure diagram of a textile and footwear upper intelligent design system based on a large model, according to an embodiment of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0064] See Figure 1 As shown in the figure, an embodiment of the present invention provides a smart design method for textile footwear uppers based on a large model, comprising the following steps:

[0065] S101, preprocess the original image data of textile and shoe uppers, and write corresponding text description labels for each image to obtain a multimodal dataset for training; wherein, the preprocessing includes standardizing the images and adjusting the image resolution to a uniform pixel;

[0066] S102, Design a LoRA fine-tuning training workflow for generating creative images of textile and footwear uppers, and use a multimodal dataset to execute the LoRA model training workflow to obtain a textile and footwear upper generation model;

[0067] S103, using the textile upper generation model to generate a creative image of the upper based on the text description input by the user;

[0068] S104, Using a textile shoe upper generation model, corresponding shoe upper creative images are intelligently generated based on text descriptions. The shoe upper creative images are then evaluated using image scoring to obtain a final quality score.

[0069] S105, compare the final quality score with a preset threshold: when the final quality score is less than the preset threshold, prompt the user to regenerate the shoe upper creative image or automatically regenerate the shoe upper creative image; when the final quality score is not less than the preset threshold, output a qualified shoe upper creative image.

[0070] The image scoring evaluation includes:

[0071] The creative image of the shoe upper is scaled to a fixed size to eliminate resolution differences, and the creative image of the shoe upper is normalized to obtain a first image; the first image is converted from RGB space to YCbCr space and Lab space to obtain a second image and a third image respectively; the first image is divided into multiple non-overlapping blocks to obtain several block images; at the same time, the RGB space image is retained for subsequent feature extraction of different types and local defect detection;

[0072] Multi-dimensional traditional image quality features are extracted based on the second image, the third image, and the block image;

[0073] Extract AI defect features from AI-generated images that have unique defects;

[0074] The first image is scored using a multimodal large model to obtain large model evaluation features; wherein, preset prompt words are input into the multimodal large model to make it ignore artistic style differences, focus only on objective quality defects, and output a quality score of 0-100, and the quality score is normalized and used as the large model evaluation features.

[0075] The traditional image quality features, the AI ​​defect features, and the large model evaluation features are normalized, and a hierarchical weighting strategy is used to fuse the features to obtain a fusion score. On this basis, a correction factor for handling fatal defects is introduced to obtain a final quality score, and the final quality score is output. The final quality score is obtained by applying a reduction processing determined by the correction factor to the fusion score.

[0076] For specific details and procedures, please refer to [link / reference]. Figure 2 As shown, the original image data of textile shoe uppers is first preprocessed to standardize the image resolution to a uniform pixel level and generate corresponding text description labels for each image to construct a multimodal dataset for model training. A LoRA model training workflow is built and executed based on an open-source platform, using the preprocessed professional dataset for training to obtain a dedicated model for textile shoe uppers. Subsequently, in the inference stage, a creative image workflow is built on the open-source platform, integrating the trained LoRA model and image scoring and evaluation process. This enables the intelligent generation of corresponding creative shoe upper images based on text descriptions and the scoring and evaluation of the generated results. The core functionality is encapsulated as an API interface. The front-end is based on the HTML framework, using TailWind and Vue technologies to develop a local application interface and implement user interaction modules. The back-end service is built using the FastAPI framework and utilizes WebSocket service to achieve long-term communication with the front-end application, completing the API calls and service integration. For practical applications, refer to [link to relevant documentation]. Figure 3 As shown, the local application interface implements functions such as text input, parameter adjustment, real-time monitoring of task progress, and display of generated results. Finally, system integration testing and performance optimization of front-end and back-end services and workflow are performed to ensure stable software operation.

[0077] Specifically, the training data used to build the fine-tuned training network is mainly based on professional shoe upper design images provided by relevant companies. In the data preprocessing stage, all images are standardized to a uniform pixel resolution, and a text description file containing key features of the shoe upper pattern is created for each image, ensuring a one-to-one correspondence between image data and text description, thus obtaining the multimodal dataset for training.

[0078] Specifically, the image generation inference workflow is built on an open-source platform. A LoRA-tuned textile and footwear upper-specific model serves as the core model in this workflow, receiving textual descriptions as input and outputting creative footwear upper design images. The backend maintains a persistent connection with the frontend interaction module using WebSocket, enabling real-time push of intermediate progress and final results to the frontend during image generation. The inference workflow encapsulated in the backend service provides an API interface, allowing the frontend to call the model to generate images and obtain result data.

[0079] Specifically, a front-end / back-end separation architecture is adopted. The front-end is based on the HTML framework, combined with TailWind and Vue to develop a local interactive interface. This interface acquires and sends the user-inputted shoe upper design text and generation parameters, and displays the real-time status of the generation process and the final shoe upper design image in a visual manner. The front-end can also have creative... Figure 2 The editing function provides a drawing panel on the front end, allowing manual adjustments to the images generated from the large model. The adjusted images can then be sent back to the textile and shoe upper generation model for secondary generation or optimization to obtain a more suitable creative image for the shoe upper.

[0080] Specifically, embodiments of the present invention may also include a parameter guidance module, see [link to relevant documentation]. Figure 4 As shown, the front end has parameter and guide word suggestion functions based on a multimodal large model. It allows the analysis of the currently generated image and existing guide words and parameters using a multimodal large model, and provides new guide words and parameters based on user needs to help users generate more efficiently and improve output quality.

[0081] Specifically, embodiments of the present invention also include an image scoring and evaluation process, see [link to relevant documentation]. Figure 5 As shown, the image scoring evaluation preprocesses the input AI-generated image by scaling it to a fixed size to eliminate resolution differences and normalizing it to obtain a first image. The first image is then converted from RGB space to YCbCr space and Lab space to obtain a second image and a third image, respectively, while retaining the RGB space for subsequent extraction of different types of features. The first image is then divided into multiple non-overlapping blocks to obtain several block images, preparing for local defect detection and local statistical feature extraction.

[0082] Specifically, in the multi-dimensional feature extraction stage, multi-dimensional traditional image quality features are first extracted, including at least sharpness features, noise and artifact features, contrast and dynamic range features, and color consistency features. Among them, sharpness features are obtained from the Y component of the second image and characterized based on the statistics of the Laplacian response and the edge density obtained by the Sobel operator; noise and artifact features are obtained from the block image and characterized based on the mean of the standard deviation and the mean of the entropy of the block image combined with a preset threshold; contrast and dynamic range features are obtained from the Y component of the second image and characterized based on the quantile statistics of the grayscale image; color consistency features are obtained from the third image and characterized based on the mean and standard deviation of the a channel and the b channel combined with a preset coefficient.

[0083] Specifically, the unique defect features of the AI-generated image are then extracted, including at least semantic consistency features, detail realism features, and generation logic consistency features. Among them, the semantic consistency features are obtained based on the RGB image and are characterized by at least contour smoothness and region compactness; the detail realism features are obtained based on the block image and are characterized based on the peak number of the block image and combined with preset coefficients; the generation logic consistency features are obtained based on the Y component of the second image and are characterized by at least the mean of the light and shadow direction deviation and the occlusion rationality score.

[0084] Specifically, the large model evaluation feature extraction process is executed synchronously. The preprocessed first image is input into the multimodal large model, and preset prompt words are input to make it ignore artistic style differences and only focus on objective quality defects. The evaluation is carried out from at least two dimensions: traditional image quality and AI-generated unique defects, and a quality score of 0-100 is output. The quality score is normalized and used as the large model evaluation feature.

[0085] Specifically, to facilitate the fusion of features with different dimensions, the traditional image quality features, the AI ​​defect features, and the large model evaluation features are all normalized to ensure that the features fall within the [0,1] interval. The feature extrema used for normalization are the feature extrema on a preset high-quality AI-generated image dataset. For features where larger values ​​are better, linear normalization is used, with the formula shown below:

[0086] ;

[0087] in, Let be a certain original feature value to be normalized. These are the normalized eigenvalues; For the minimum and maximum values ​​of this feature from a high-quality AI-generated image dataset. When When dividing by zero, a very small positive number can be added to the denominator to avoid division by zero. For features where smaller is better, reverse normalization can be used to map them to the [0,1] interval and maintain the consistent evaluation direction of "the larger the value, the better the quality".

[0088] Specifically, the final step involves feature weighting fusion and decision correction, employing a hierarchical weighting strategy. Features are divided into four layers: traditional image quality layer, AI-specific defect layer, semantic logic layer, and large model scoring layer, with corresponding weights assigned to each. The fusion score is obtained by multiplying the scores of each of the four layers by their respective weights and then summing the results, as shown in the formula below:

[0089] ;

[0090] in, To achieve a score based on integration, For layer index, These correspond to the traditional image quality layer, the AI-specific defect layer, the semantic logic layer, and the large model scoring layer, respectively. For the first The layer score, which can be obtained by aggregating several normalized features within the layer in a preset manner; For the first Layer weights are non-negative real numbers, and can be preset or determined by parameter tuning through a validation set.

[0091] Specifically, a correction factor is introduced to handle fatal flaws. When the semantic consistency feature is lower than a first preset threshold or the generation logic consistency feature is lower than a second preset threshold, the correction factor is set to a preset value less than 1; otherwise, the correction factor is set to 1. The final quality score is calculated based on the fusion score and the correction factor. The formula is as follows:

[0092] ;

[0093] in, For the final quality score, To achieve a score based on integration, As a correction factor; when a fatal flaw is triggered A preset value less than 1 is used to reduce the fusion score; otherwise... In addition to outputting the final quality score, it can also output feature scores for each dimension to pinpoint specific quality defects in the AI-generated image, thereby assisting users in adjusting prompts or parameters and guiding subsequent iterations.

[0094] Specifically, after scoring, the final quality score is compared with a preset threshold: if the final quality score is less than the preset threshold, the user is prompted to regenerate the shoe upper design image, or the shoe upper design image is automatically regenerated; if the final quality score is not less than the preset threshold, a qualified shoe upper design image is output. The threshold can be determined based on actual business acceptance standards or validation set statistics.

[0095] For details on the display effect of the local application interface, please refer to [link / reference]. Figure 6 and Figure 7 The figures shown are a visualization of the process and a result diagram of a large-model-based intelligent design method for textile and footwear uppers according to an embodiment of the present invention.

[0096] See Figure 8 The diagram shown is a system structure diagram of a large-scale intelligent design system for textile and footwear uppers according to an embodiment of the present invention, including:

[0097] The dataset creation module 801 is used to preprocess the original image data of textile and footwear uppers and write corresponding text description labels for each image to obtain a multimodal dataset for training.

[0098] The generative model training module 802 is used to design a LoRA fine-tuning training workflow for generating creative images of textile and footwear uppers, and to execute the LoRA model training workflow using the multimodal dataset to obtain a textile and footwear upper generation model.

[0099] The shoe upper image generation module 803 is used to generate a creative shoe upper image by calling the textile shoe upper generation model based on the text description input by the user.

[0100] The image scoring module 804 is used to score and evaluate the creative image of the shoe upper. The image scoring module includes an image preprocessing submodule, a traditional image quality feature extraction submodule, an AI defect feature extraction submodule, a multimodal large model scoring submodule, and a hierarchical weighted fusion and fatal defect correction submodule, and outputs a final quality score. Specifically, the image preprocessing submodule is used to scale the image to a fixed size to eliminate resolution differences, convert the image from RGB space to YCbCr space and Lab space and perform block processing, while retaining the RGB space image for subsequent feature extraction and local defect detection. The hierarchical weighted fusion and fatal defect correction submodule is used to divide the features into four layers, perform weighted summation to obtain a fusion score, and reduce the fusion score through a correction factor when a fatal defect is triggered to obtain the final quality score.

[0101] The iteration control module 805 is used to compare the final quality score with a preset threshold, and prompt to regenerate the shoe upper creative image or automatically regenerate the shoe upper creative image when the final quality score is less than the preset threshold, and output a qualified shoe upper creative image when the final quality score is not less than the preset threshold.

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

Claims

1. A smart design method for textile footwear uppers based on a large model, characterized in that, Includes the following steps: The original image data of textile and footwear uppers are preprocessed, and corresponding text description labels are written for each image to obtain a multimodal dataset for training; wherein, the preprocessing includes standardizing the images and adjusting the image resolution to a uniform pixel; Design a LoRA fine-tuning training workflow for generating creative images of textile shoe uppers, and execute the LoRA model training workflow using a multimodal dataset to obtain a textile shoe upper generation model; The aforementioned textile shoe upper generation model generates a creative shoe upper image based on the text description input by the user; A textile shoe upper generation model is used to intelligently generate corresponding shoe upper creative images based on text descriptions. The shoe upper creative images are then evaluated by image scoring to obtain a final quality score. The final quality score is compared with a preset threshold: when the final quality score is less than the preset threshold, the user is prompted to regenerate the shoe upper creative image or the shoe upper creative image is automatically regenerated; when the final quality score is not less than the preset threshold, a qualified shoe upper creative image is output. The image scoring evaluation includes: The creative image of the shoe upper is scaled to a fixed size to eliminate resolution differences, and the creative image of the shoe upper is normalized to obtain a first image; the first image is converted from RGB space to YCbCr space and Lab space to obtain a second image and a third image respectively; the first image is divided into multiple non-overlapping blocks to obtain several block images; at the same time, the RGB space image is retained for subsequent feature extraction of different types and local defect detection; Multi-dimensional traditional image quality features are extracted based on the second image, the third image, and the block image; Extract AI defect features from AI-generated images that have unique defects; The first image is scored using a multimodal large model to obtain large model evaluation features; wherein, preset prompt words are input into the multimodal large model to make it ignore artistic style differences, focus only on objective quality defects, and output a quality score of 0-100, and the quality score is normalized and used as the large model evaluation features. The traditional image quality features, the AI ​​defect features, and the large model evaluation features are normalized, and a hierarchical weighting strategy is used to fuse the features to obtain a fusion score. On this basis, a correction factor for handling fatal defects is introduced to obtain a final quality score, and the final quality score is output. The final quality score is obtained by applying a reduction processing determined by the correction factor to the fusion score.

2. The intelligent design method for textile and footwear uppers based on a large model according to claim 1, characterized in that, The traditional image quality features include at least sharpness features, noise and artifact features, contrast and dynamic range features, and color consistency features.

3. The intelligent design method for textile and footwear uppers based on a large model according to claim 2, characterized in that, The sharpness feature is obtained from the Y component of the second image and is characterized based on the statistics of the Laplacian response and the edge density obtained by the Sobel operator. The noise and artifact features are obtained from the block image and are characterized based on the mean standard deviation and mean entropy of the block image, combined with a preset threshold. The contrast and dynamic range features are obtained from the Y component of the second image and characterized based on the quantile statistics of the grayscale image. The color consistency feature is obtained from the third image and is characterized based on the mean and standard deviation of channel a and channel b, combined with a preset coefficient.

4. The intelligent design method for textile and footwear uppers based on a large model according to claim 1, characterized in that, The AI ​​defect features include at least semantic consistency features, detail authenticity features, and generation logic consistency features; The semantic consistency feature is obtained based on RGB images and is characterized by at least contour smoothness and region compactness. The detailed authenticity features are obtained based on the block image and are characterized based on the number of peaks in the block image and a preset coefficient. The generated logic consistency feature is obtained based on the Y component of the second image and is characterized by at least the mean of the light and shadow direction deviation and the occlusion rationality score.

5. The intelligent design method for textile and footwear uppers based on a large model according to claim 1, characterized in that, The large model evaluation features are obtained through the following process: the first image is scored using a multimodal large model to obtain an image quality score of 0–100, and the image quality score is normalized to obtain the large model evaluation features; wherein, the preset prompt words are required to be evaluated from at least two dimensions: traditional image quality and AI-generated unique defects.

6. The intelligent design method for textile and footwear uppers based on a large model according to claim 1, characterized in that, The process of obtaining the final quality score includes the following steps: The traditional image quality features, the AI ​​defect features, and the large model evaluation features are normalized, wherein the feature extrema used for normalization are the feature extrema on a preset high-quality AI-generated image dataset; and the normalization includes linear normalization or inverse normalization to ensure that the features fall within the [0,1] interval. The normalization formula is as follows: ； in, , For the feature extrema from a high-quality AI-generated image dataset; This represents the value before normalization. Represents a normalized value; A hierarchical weighted strategy is used to calculate the fusion score, dividing the features into four layers: traditional image quality layer, AI-specific defect layer, semantic logic layer, and large model scoring layer, with corresponding weights assigned to each. The fusion score is obtained by multiplying the scores of each of the four layers by their respective weights and then summing the results, as shown in the formula below: ； in, This represents the score of the feature at the i-th layer. Indicates correspondence The weight, Indicates the fusion score; A correction factor is introduced to handle fatal defects: when the semantic consistency feature is lower than the first preset threshold or the generation logic consistency feature is lower than the second preset threshold, the correction factor takes a preset value less than 1; otherwise, the correction factor takes 1. The final quality score is calculated based on the fusion score and the correction factor. The design image for the shoe upper is deemed unqualified if the final quality score is less than a preset threshold, and the formula is as follows: ； Among them, the correction factor When a fatal defect is triggered, a preset value less than 1 is used; otherwise, 1 is used.

7. The intelligent design method for textile and footwear uppers based on a large model according to claim 1, characterized in that, Also includes: The creative image of the shoe upper generated by the large model is manually adjusted, and then the adjusted creative image of the shoe upper is sent back to the textile shoe upper generation model for secondary generation or optimization.

8. The intelligent design method for textile and footwear uppers based on a large model according to claim 1, characterized in that, Also includes: By using a multimodal large model to analyze the currently generated images and existing guide words and parameters, new guide words and parameters can be provided based on user needs.

9. The intelligent design method for textile and footwear uppers based on a large model according to claim 1, characterized in that, The textile and shoe upper generation model and the image scoring evaluation are encapsulated into an interface; and the front-end interface enables text input, generation parameter adjustment, task progress monitoring and generation result display. The back-end service calls the interface and returns the generation progress and results to the front-end.

10. A smart design system for textile footwear uppers based on a large model, characterized in that, include: The dataset creation module is used to preprocess the original image data of textile and footwear uppers and write corresponding text description labels for each image to obtain a multimodal dataset for training; wherein, the preprocessing includes standardizing the images and adjusting the image resolution to a uniform pixel; The generative model training module is used to design a LoRA fine-tuning training workflow for generating creative images of textile and footwear uppers, and to execute the LoRA model training workflow using the multimodal dataset to obtain a textile and footwear upper generation model. The shoe upper image generation module is used to generate a creative shoe upper image by calling the textile shoe upper generation model based on the text description input by the user; An image scoring module is used to score and evaluate the creative image of the shoe upper. The image scoring module includes: an image preprocessing submodule, a traditional image quality feature extraction submodule, an AI defect feature extraction submodule, a multimodal large model scoring submodule, and a hierarchical weighted fusion and fatal defect correction submodule, and outputs a final quality score. Specifically, the image preprocessing submodule scales the image to a fixed size to eliminate resolution differences, converts the image from RGB space to YCbCr space and Lab space, and performs block processing, while retaining the RGB space image for subsequent feature extraction and local defect detection. The hierarchical weighted fusion and fatal defect correction submodule divides the features into four layers, performs weighted summation to obtain a fusion score, and reduces the fusion score by a correction factor when a fatal defect is triggered to obtain the final quality score. The iterative control module is used to compare the final quality score with a preset threshold, and prompt to regenerate the shoe upper creative image or automatically regenerate the shoe upper creative image when the final quality score is less than the preset threshold, and output a qualified shoe upper creative image when the final quality score is not less than the preset threshold.

Images