An upper map intelligent generation method and device

Abstract

This application discloses an intelligent method and apparatus for generating shoe upper textures. By acquiring shoe outline data and performing standardized preprocessing, and then inputting it into an integrated shoe upper texture generation model, multi-dimensional joint optimization processing is performed, significantly improving texture production efficiency and shortening the production cycle. The model's built-in distortion correction function can automatically compensate for stretching and misalignment caused by the curved surface of the shoe last, ensuring that the pattern is not distorted in areas with severe curvature such as the toe and heel. Combined with preset weaving clarity indicators, the texture clarity is optimized in real time, making the design drawing highly matched with the actual fly-knitted or warp-knitted sample, avoiding printing misalignment and blurring. The final generated texture file includes weaving control parameters, achieving seamless integration from design to production, facilitating batch adaptation of different shoe models, and greatly improving production yield and process synergy.

Description

Technical Field

[0001] This application relates to the field of weaving technology, and in particular to a method and apparatus for intelligent generation of shoe upper textures. Background Technology

[0002] In the footwear manufacturing industry, especially in the production of flyknit and warp-knitted uppers, the visual design of upper patterns and textures is a key element in enhancing product added value. Currently, the industry mainly relies on traditional manual processing methods for generating and adapting upper textures. Designers typically use general image processing software (such as Adobe Photoshop) to manually apply textures, using a two-dimensional unfolded upper as a base. They then manually scale, rotate, and transform the textures to place and adjust them one by one onto the corresponding areas of the base. Based on experience, they simulate the visual effect of the textures on a three-dimensional shoe last, repeatedly correcting pixel stretching and misalignment in the two-dimensional image.

[0003] However, this existing technology has significant drawbacks and shortcomings in practical applications. In terms of production efficiency, it relies entirely on manual Photoshop operations. The textures for each shoe model must be created from scratch. For patterns containing complex gradients, geometric grids, or intricate designs, designers need to manually match and adjust them pixel by pixel and area by area. The texture creation for a single shoe upper typically takes several hours to several days. When it involves a series or multiple color schemes, the amount of repetitive work is enormous, which seriously slows down the overall R&D cycle from design to prototyping.

[0004] Regarding texture quality, the final shoe upper needs to be fitted to the surface of a shoe last with complex three-dimensional curvature. Traditional two-dimensional texture mapping methods lack mathematical mapping of three-dimensional surface geometry. When a two-dimensional planar texture is mapped onto the shoe last model, significant stretching, compression, misalignment, or texture distortion will inevitably occur in areas with drastic curvature changes (such as the toe, heel, and sides of the arch). Designers cannot preview or effectively correct this three-dimensional distortion in real time in the Photoshop environment, resulting in severe pattern deformation during prototype production, requiring repeated modifications, and even causing material waste.

[0005] Regarding shoe model adaptation, existing methods require manual adjustments to the unfolded pattern in Photoshop for each shoe model or size when creating textures for different shoe models (such as sneakers, casual shoes, and high heels) or different sizes of the same model. Different shoe models have varying outlines, structural dividing lines, and curved surface unfolding forms, making it impossible to achieve universal reuse of patterns through simple scaling. This necessitates companies maintaining a large, model-specific texture library, resulting in extremely low resource reuse rates.

[0006] In matching with flyknit or warp-knitting processes, traditional Photoshop texture mapping only focuses on two-dimensional visual representation and does not consider the physical constraints of the knitting process, such as the continuity of the yarn, the influence of the knitting direction on pattern stretching, and the dimensional variation coefficients of different needle bed areas. Therefore, serious printing misalignments often occur between the design drawing and the actual woven shoe upper in terms of pattern position, proportion, and seams; that is, design elements (such as logos and seam lines) cannot be aligned with the actual structural seams of the shoe upper. This disconnect between design and process makes it difficult to guarantee the yield rate of high-precision patterns.

[0007] In summary, existing technologies lack a method for automatically generating textures and performing real-time distortion correction based on the geometric features of the shoe last's 3D data or the unfolded pattern of the shoe upper. The entire process relies on human experience and lacks an intelligent, parametric mapping mechanism from "design drawing" to "3D surface adaptation" and then to "process production matching". Summary of the Invention

[0008] To address the problems of the prior art, this application provides a method and apparatus for intelligent generation of shoe upper textures, thereby resolving one or more technical problems existing in the prior art.

[0009] In a first aspect, this application provides a method for intelligently generating shoe upper textures, the method comprising: Obtain the shoe outline data of the shoe model to be processed, and perform standardized preprocessing on the shoe outline data; The preprocessed shoe outline data is input into the pre-trained integrated shoe upper texture generation model for multi-dimensional joint optimization processing, and the initial shoe upper texture is output. The texture clarity of the initial shoe upper texture is verified and optimized in real time using the integrated shoe upper texture generation model and the preset weaving clarity index to generate the target shoe upper texture. A shoe upper texture file is generated based on the target shoe upper texture. The shoe upper texture file includes the target shoe upper texture in a preset image format and the corresponding weaving control parameters.

[0010] In some embodiments, the multi-dimensional joint optimization process includes at least one of texture generation, color matching optimization, splicing processing, and distortion correction.

[0011] In some embodiments, the integrated shoe upper texture generation model is an end-to-end fusion large model, and integrates a texture encoder, a color generator, a splicing learner, and a deformation predictor.

[0012] In some embodiments, the training process of the integrated shoe upper texture generation model includes: Construct a multi-dimensional training dataset, which includes at least shoe shape contour samples, shoe upper texture samples, color scheme samples, curved surface deformation samples, and weaving process constraint samples. The multi-dimensional training dataset is uniformly labeled and its features are extracted to obtain feature information of various types of samples; The extracted feature information of each type of sample is uniformly encoded and fused to obtain a feature vector of uniform dimension; A multi-task joint learning approach was adopted, with the optimization objectives of texture generation accuracy, color matching naturalness, splicing rationality, and distortion correction error as the optimization targets. An initial integrated shoe upper texture generation model was obtained through end-to-end overall training. The model parameters of the texture encoder, color generator, splicing learner, and deformation predictor are synchronously and iteratively updated using the backpropagation algorithm to obtain the final integrated shoe upper texture generation model.

[0013] In some embodiments, the splicing process includes: Based on the shoe upper seams, sole bonding boundaries, and weaving start and end areas, the system automatically identifies the shoe upper structure division lines. At the dividing line, texture gradient, smooth transition and seamless splicing operations are performed.

[0014] In some embodiments, the distortion correction includes: Based on the curvature of the last surface in the shoe profile data, the stretching area is calculated and the stretching factor is determined. According to the stretching factor, UV remapping is performed on the shoe last surface to compensate for the texture ratio; By combining mesh deformation optimization and optical flow field adjustment, local distortion elimination and pixel-level smoothing are performed on the remapped texture.

[0015] In some embodiments, calculating the stretching region and determining the stretching factor based on the curvature of the shoe last surface in the shoe profile data includes: The three-dimensional shoe last surface is triangularly meshed, and the principal curvature of each mesh node is calculated. Calculate the Gaussian curvature and the mean curvature based on the principal curvature; The stretching factor is calculated based on the Gaussian curvature. The formula for calculating the stretching factor is as follows: , where S is the stretching factor, K is the Gaussian curvature, and ε is a minimum positive number.

[0016] In some embodiments, the multi-dimensional joint optimization process further includes a fusion process, which is used to consistently fuse the shoe outline data processed by different optimization paths to form a continuous initial shoe upper texture.

[0017] In some embodiments, the real-time verification and optimization of the texture clarity of the initial shoe upper texture using the integrated shoe upper texture generation model and a preset weave clarity index includes: The preset weaving clarity index of the initial shoe upper texture is calculated in real time. The preset weaving clarity index is used to comprehensively evaluate the texture edge sharpness, detail retention and weaving restoreability. When the preset weaving clarity index is lower than the preset index threshold, the internal parameters of the integrated shoe upper mapping generation model are automatically adjusted for optimization until the calculated preset weaving clarity index is not lower than the preset index threshold.

[0018] In some embodiments, the integrated shoe upper texture generation model further integrates a weaving constraint layer; the method further includes: Through the weaving constraint layer, based on preset weaving equipment parameters, the generated target shoe upper texture is mapped and aligned pixel-level with the weaving path, needle bed distribution, and yarn feed trajectory of the weaving equipment.

[0019] In some embodiments, the weaving control parameters include at least a pixel-level color map, a yarn color correspondence table, jacquard control instructions, weaving start point coordinates, yarn feeder movement path, and alignment marks for preventing pattern offset.

[0020] In a second aspect, this application provides a smart shoe upper texture generation device, which is used to implement the smart shoe upper texture generation method as described in any of the first aspects.

[0021] In a third aspect, embodiments of this application disclose a computer device including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and when the computer program is executed by the processor, it implements the intelligent generation method for shoe upper textures as described in any of the first aspects.

[0022] In a fourth aspect, embodiments of this application disclose a computer-readable storage medium storing a computer program, which, when executed, implements the intelligent generation method for shoe upper textures as described in any of the first aspects.

[0023] According to the specific embodiments provided in this application, the following technical effects are disclosed: In this embodiment, by acquiring shoe outline data and performing standardized preprocessing, and then inputting it into an integrated shoe upper texture generation model, multi-dimensional joint optimization processing such as texture generation, color matching, splicing, and distortion correction is completed simultaneously, significantly improving texture production efficiency and shortening the production cycle. The model's built-in distortion correction function can automatically compensate for stretching and misalignment caused by the curved surface of the shoe last, ensuring that the pattern is not distorted in areas with severe curvature such as the toe and heel. Combined with preset weaving clarity indicators, texture clarity is optimized in real time, making the design drawing highly matched with the actual fly-knitted or warp-knitted sample, avoiding printing misalignment and blurring. The final generated texture file includes weaving control parameters, achieving seamless integration from design to production, facilitating batch adaptation of different shoe models, and greatly improving production yield and process synergy.

[0024] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0025] The disclosure of this invention will become more readily understood with reference to the accompanying drawings. It will be readily understood by those skilled in the art that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Furthermore, similar numbers in the drawings are used to denote similar components, wherein: Figure 1 This is an architecture diagram of the integrated shoe upper texture generation model provided in some embodiments of this application; Figure 2 This is a flowchart of a method for intelligently generating shoe upper textures provided in some embodiments of this application; Figure 3 This is a training framework diagram of the integrated shoe upper texture generation model provided in some embodiments of this application; Figure 4 This is a schematic diagram of the structure of a computer device provided in some embodiments of this application. Detailed Implementation

[0026] Some embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0027] As described in the background section, current shoe upper texture creation relies entirely on manual Photoshop operations, resulting in low efficiency and lengthy cycles, with texture creation for a single shoe model taking hours to days. Due to the lack of 3D mapping capabilities on the curved surfaces of the shoe last, textures are prone to stretching, misalignment, and distortion in areas with significant curvature, such as the toe and heel. When changing shoe models or sizes, manual adjustments are required from scratch, making batch adaptation difficult. Furthermore, the poor compatibility between the designed textures and the flyknit and warp-knitting processes frequently leads to printing misalignments, making it difficult to guarantee the yield rate of high-precision patterns. None of the above methods possess automated texture generation and distortion correction capabilities, severely hindering the intelligent upgrade of shoe upper design and manufacturing processes to meet the demands for high efficiency, high consistency, and rapid model changes.

[0028] To address one or more of the aforementioned problems, this application proposes an intelligent method and apparatus for generating shoe upper textures. The method first acquires the shoe outline data of the shoe model to be processed and performs standardized preprocessing to eliminate differences in the original data format. Then, the preprocessed data is input into a pre-trained integrated shoe upper texture generation model, undergoing multi-dimensional joint optimization processing including texture generation, color matching optimization, splicing processing, and distortion correction. This results in a one-time output of the initial shoe upper texture, completely breaking the limitations of fragmented processes and repeated manual intervention in traditional workflows. Next, the model, combined with preset weaving clarity indicators, verifies and optimizes the texture clarity of the initial texture in real time, generating the target shoe upper texture. Finally, a shoe upper texture file containing the target shoe upper texture in a preset image format and corresponding weaving control parameters is generated. This fundamentally solves problems such as surface distortion, poor process matching, and difficulty in adapting to different styles, significantly improving design efficiency, texture quality, and consistency in mass production. It achieves a leap from manual experience-driven to data-driven shoe upper texture generation, providing an efficient collaborative solution from design to production for flyknit or warp-knitted shoe manufacturing.

[0029] The solution of this application will now be described in detail with reference to the accompanying drawings and various embodiments.

[0030] Example 1 The intelligent shoe upper texture generation method provided in this application embodiment is based on a pre-trained integrated shoe upper texture generation model, referring to... Figure 1As shown, this integrated shoe upper texture generation model is an end-to-end fusion model, internally integrating a texture encoder, a color generator, a stitching learner, and a deformation predictor. The texture encoder employs a convolutional neural network (CNN) for feature extraction, performing multi-layer convolution and pooling operations on the input shoe outline data and reference pattern materials to progressively extract texture features from local to global levels, ultimately outputting a high-dimensional texture feature vector to provide structured representation for subsequent modules. The color generator is based on a generative adversarial network (GAN) architecture. Through adversarial training between the generator and discriminator, the generator can automatically learn the reasonable color distribution and gradient rules of various areas of the shoe upper. The discriminator is responsible for evaluating the realism and style consistency of the color matching results. After optimization through this game between the two, a visually natural color scheme that meets design requirements is output. The stitching learner can incorporate a Graph Neural Network (GNN) to abstract the spatial adjacency relationships between different sections of the shoe upper into a graph structure. Using segmented areas of the shoe upper as nodes and stitching boundaries between areas as edges, it learns the fusion weights and transition characteristics between adjacent textures through graph convolution operations. This effectively eliminates visual breaks and color abrupt changes at stitching seams, achieving smooth and natural overall texture blending. The deformation predictor can adopt a Transformer temporal model architecture, treating the standardized preprocessed shoe outline data as a spatial position sequence. It utilizes the Transformer's self-attention mechanism to capture the long-distance nonlinear dependency between 2D design coordinates and 3D shoe last surfaces, predicting pixel-by-pixel the geometric offset and stretch compensation of the texture in areas with varying curvature, such as the toe, heel, and arch, thus simultaneously completing distortion correction during texture generation. These four components are jointly trained and collaboratively inferred within an end-to-end framework, requiring no manual step-by-step intervention, and outputting an initial shoe upper texture that has undergone texture synthesis, color optimization, seamless stitching, and surface distortion correction in a single step.

[0031] It should be noted that "end-to-end" here refers to a model architecture where the model completes the entire task directly from the original input to the final output, without requiring manual splitting of multiple independent sub-modules or reliance on manually designed intermediate features. In this application, the input of the integrated shoe upper texture generation model is standardized preprocessed shoe outline data, and the output is a complete initial shoe upper texture (including texture, color matching, splicing, distortion correction, etc.). There is no need for manual execution of texture drawing, color adjustment, seam trimming, distortion compensation, etc. The model completes all tasks in one go, hence the term "end-to-end."

[0032] "Fusion" emphasizes that the model integrates multiple sub-tasks (texture generation, color optimization, stitching, distortion correction, and fusion processing) that originally required multiple independent algorithms or manual operations into a unified model framework for simultaneous learning and joint inference. Different sub-tasks share feature representations, which can promote each other, avoid error accumulation, and achieve the best overall performance, rather than simply stringing multiple modules together.

[0033] "Large-scale models" typically refer to deep learning models with a large number of parameters (millions, hundreds of millions, or even more) trained on massive datasets, possessing stronger feature extraction capabilities, nonlinear mapping capabilities, and generalization performance. In this application, the large-scale model integrates multiple functional components such as a texture encoder, color generator, stitching learner, and deformation predictor. These components are jointly trained within the same end-to-end framework of the large-scale model, enabling it to learn the complex mapping relationship from two-dimensional shoe contours to three-dimensional surface adaptation, thereby automatically completing distortion correction and weave clarity optimization.

[0034] Figure 2 This is a flowchart of the intelligent generation method for shoe upper textures provided in the embodiments of this application, refer to... Figure 2 As shown, the method includes the following steps: S100: Obtain the shoe outline data of the shoe model to be processed, and perform standardized preprocessing on the shoe outline data.

[0035] In step S100, the shoe outline data of the shoe model to be processed is first acquired. Shoe outline data refers to a set of key information that uniquely characterizes the three-dimensional geometric shape and two-dimensional unfolded pattern of the shoe model. Specifically, it may include, but is not limited to: three-dimensional scan data or a three-dimensional design model of the shoe last, two-dimensional unfolded outlines of each piece of the upper (such as the toe, body, heel counter, and tongue), and the splicing relationships and seam positions between the pieces. This data can be acquired by directly reading standard file formats output by the design department (such as OBJ, STL, DXF, or formats from specialized shoe pattern design software), or by reverse engineering existing shoe patterns using image acquisition or scanning equipment. It should be noted that the original input data output by different design software or scanning equipment varies significantly in terms of coordinate system, point cloud integrity, noise level, and outline scale. For example, the three-dimensional shoe last model may be in millimeters and the origin position may not be fixed; the point cloud data may contain missing holes or outliers; and the two-dimensional unfolded diagram may lack a unified dimensional reference. To this end, step S100 further standardizes and preprocesses the acquired shoe outline data, specifically including coordinate alignment, missing point completion, point cloud denoising, and outline normalization. The above operations are automatically completed by the system preprocessing module without manual intervention.

[0036] Coordinate alignment refers to uniformly transforming all input data to a preset world coordinate system, eliminating inconsistencies in origin, axis, and units between data from different sources, and ensuring a unified spatial reference for subsequent processing. Missing point completion is used to repair holes and breaks in 3D point clouds or 2D contours caused by scanning occlusion or data export omissions. It typically uses interpolation or local curvature-based completion algorithms to automatically generate reasonable replacement points, restoring the complete geometric shape. Point cloud denoising utilizes statistical filtering or radius filtering methods to automatically remove outliers, isolated patches, and obvious measurement errors from the original data, improving the signal-to-noise ratio. Contour normalization refers to scaling the contour data to the standard size range required by the model input based on the structural characteristics of different shoe types (such as shoe length, width, girth, and other key dimensions), while maintaining the original geometric proportions.

[0037] After the above preprocessing operations are completed, the standardized data can be encoded into a multidimensional tensor format acceptable to the model. For example, a two-dimensional unfolded contour map can be converted into a fixed-resolution multi-channel image tensor, or a three-dimensional shoe last point cloud can be sampled into an ordered point sequence matrix. Through the fully automated execution of the preprocessing module, the output shoe contour data has a unified coordinate system, a complete point cloud structure, clean geometric information, and standardized contour scale, which can be directly input into the subsequent integrated shoe upper texture generation model for end-to-end intelligent processing. This step effectively avoids the problem of model inference failure or texture quality degradation caused by the heterogeneity of the original data, laying a data foundation for the realization of fully automated, batch shoe upper texture generation.

[0038] S200: Input the pre-processed shoe outline data into the pre-trained integrated shoe upper texture generation model, perform multi-dimensional joint optimization processing, and output the initial shoe upper texture.

[0039] In some embodiments, the multi-dimensional joint optimization process includes at least one of texture generation, color matching optimization, splicing processing, and distortion correction.

[0040] The shoe outline data, which has been standardized and preprocessed in step S100, is input into the pre-trained integrated shoe upper texture generation model. The model simultaneously performs multi-dimensional joint optimization processing such as texture generation, color matching optimization, splicing processing, and distortion correction, and outputs the initial shoe upper texture in one go.

[0041] As mentioned above, the integrated shoe upper texture generation model in this embodiment is an end-to-end fusion model that integrates a texture encoder, color generator, stitching learner, and deformation predictor. Unlike traditional methods that break down tasks such as texture drawing, color adjustment, stitching trimming, and distortion compensation into multiple independent steps and rely on manual operation one by one, the integrated model in this step receives standardized shoe outline data and, through the collaborative reasoning of the above four components, simultaneously performs multiple generation and optimization tasks without the need for manual step-by-step intervention.

[0042] Specifically, the texture encoder first extracts deep features from the preprocessed shoe outline data and optional reference pattern materials, encoding the geometric structure and desired texture style of each area of ​​the shoe upper into high-dimensional feature vectors. Based on this, the color generator automatically generates color distribution schemes adapted to different areas of the shoe upper according to the extracted feature vectors, achieving color optimization from local to global, ensuring the texture meets design style requirements in terms of color. The stitching learner automatically learns the fusion weights between adjacent areas based on the spatial adjacency relationships between the pieces or panels implicitly contained in the shoe outline, predicting the texture transition characteristics at the stitching boundary and eliminating visual breaks and color abrupt changes that may occur due to piece separation. The deformation predictor, based on the three-dimensional curved surface geometric features reflected by the shoe outline data, implicitly learns the nonlinear mapping relationship between the two-dimensional texture coordinates and the three-dimensional shoe last surface, predicting the geometric offset and stretch compensation of the texture at locations with drastic curvature changes (such as the toe, heel, and sides of the arch) pixel by pixel or region by region, thus completing distortion correction simultaneously with texture generation.

[0043] The four tasks mentioned above are not executed sequentially, but are performed synchronously and coupled within the end-to-end framework of the model. Texture features provide the underlying basis for color matching and stitching, the color matching result affects the visual continuity of the stitching boundary, the distortion correction offset also participates in the calculation of texture sampling coordinates, and the stitching weight is also adaptively adjusted according to the distorted mesh. Through joint training, the model learns to balance the optimization objectives of each task in a unified feature space, so that the final output texture map achieves overall optimality in terms of texture detail, color harmony, stitching smoothness, and geometric conformity. After the above synchronous processing, the model directly outputs the initial shoe upper texture map. This initial shoe upper texture map already contains complete texture information, an optimized color scheme, a seamless stitching effect, and a geometric mapping that conforms to the 3D shoe last surface after distortion correction, without the need for additional manual retouching or secondary correction. Compared with the traditional method of processing each step separately and then manually synthesizing, this step significantly shortens the texture generation time, while fundamentally avoiding the problems of cumulative errors and local distortion caused by independent processing of each step.

[0044] S300: The texture clarity of the initial shoe upper texture is verified and optimized in real time through the integrated shoe upper texture generation model and the preset weaving clarity index to generate the target shoe upper texture.

[0045] In step S300, the initial shoe upper texture output in step S200 is verified and optimized in real time using the integrated shoe upper texture generation model and the preset weaving clarity index, with a focus on improving its texture clarity, and finally generating the target shoe upper texture.

[0046] It should be noted that although the initial shoe upper texture generated in step S200 has a high degree of integrity and visual quality in terms of texture synthesis, color matching, splicing, and distortion correction, the fineness of its texture details has not yet been adapted to the actual process capabilities of subsequent flyknitting or warp knitting equipment. Different weaving equipment, yarn specifications, and weaving densities have clear upper limits on the recognizable precision of the texture in the texture image: if the texture image contains overly dense or fine texture details (such as extremely fine lines, micro-granular patterns, or high-frequency grayscale changes), it may result in jagged lines, blurring, broken lines, or loss of details during the actual weaving process due to limitations in yarn diameter, stitch length, or weaving action; conversely, if the texture details are too sparse or coarse, the performance of high-precision weaving equipment cannot be fully utilized, resulting in a product visual effect lower than expected. To solve this problem, this step introduces a preset weaving clarity index as a quantitative benchmark for measuring whether the texture details of the texture image are adapted to the target weaving process.

[0047] Compared to traditional methods where designers rely on experience to estimate texture clarity and require multiple sampling trials to determine the appropriate detail density, this step integrates process constraints directly into the texture generation closed loop through a pre-set weaving clarity index and a real-time verification and optimization mechanism based on the integrated model. This ensures that the output texture meets the resolution and process window requirements of the weaving equipment from the source, significantly reducing trial production waste and cycle delays caused by clarity mismatch. Furthermore, because this index can be flexibly adjusted according to different equipment or yarn types, the same design texture can be quickly adapted to multiple production processes, further enhancing the flexibility and process compatibility of shoe upper texture generation.

[0048] S400: Generate an upper texture file based on the target upper texture, the upper texture file including the target upper texture in a preset image format and the corresponding weaving control parameters.

[0049] In step S400, a complete upper texture file is further generated based on the target upper texture generated in step S300. This upper texture file is not a single image file, but contains two core parts: one is the target upper texture in a preset image format, and the other is the weaving control parameters corresponding to the texture.

[0050] It's important to note that in traditional shoe upper pattern creation processes, designers typically output only one or more images in a common format (such as JPG, PNG, BMP, etc.) after completing the pattern. These images are then manually interpreted by process engineers, who then manually input or adjust corresponding process parameters in the weaving equipment control system, such as yarn needle selection information, jacquard position offset, density adjustment coefficient, and weaving actions for each area. This process not only relies on the experience and judgment of the process engineers but is also prone to errors due to the semantic gap between the pattern and equipment parameters, further exacerbating the misalignment between design and finished product. To address these issues, this step achieves joint output of the pattern and process parameters at the design stage, creating a directly executable production file.

[0051] Specifically, after the target upper texture is generated, the integrated upper texture generation model or subsequent file generation module automatically derives the corresponding weaving control parameters based on the texture distribution, color scheme, splicing boundaries, and spatial mapping relationships generated during distortion correction in the texture. Weaving control parameters refer to a set of process instructions that can be directly parsed and executed by the control system of flyknitting or warp knitting equipment. Meanwhile, the target upper texture itself is encapsulated in a preset image format (such as high-resolution lossless compressed PNG or a bitmap format specifically for jacquard equipment), which can be used for visualization and archiving, as well as serving as a reference benchmark for parameter verification.

[0052] In some embodiments, the weaving control parameters include at least a pixel-level color map, a yarn color correspondence table, jacquard control instructions, weaving start-up point coordinates, a yarn feeder movement path, and alignment marks to prevent pattern shift. The pixel-level color map maps each pixel of the target image to the corresponding needle position and knitting row number of the yarn color number, achieving precise control of a single needle. The yarn color correspondence table records the correspondence between color numbers and physical yarns (material, color number, yarn crease position). The jacquard control instructions include action parameters such as needle selection sequence, stitch lifting, and loop shifting, which, together with the color map, determine the pattern structure. The weaving start-up point coordinates specify the starting position of the knitting, ensuring consistent pattern alignment across multiple batches. The yarn feeder movement path defines the yarn feeding trajectory, and after alignment by the weaving constraint layer, ensures that the texture direction matches the equipment movement. The alignment marks embed feature points or QR codes that can be recognized by the equipment for online detection and compensation of accumulated errors, preventing pattern shift.

[0053] The shoe upper texture file can be packaged into a single file (such as a compressed file or a custom format file), or it can generate multiple files linked together according to a preset directory structure. After exporting, the file can be directly transferred to the equipment control system in the weaving workshop. The control system automatically parses the target texture and weaving control parameters within it, eliminating the need for manual secondary input or adjustment. For different sizes or color schemes of the same shoe model, the system only needs to scale the outline data or perform color mapping to quickly generate corresponding variant files, enabling rapid style changes.

[0054] Reference Figure 3 As shown, in some embodiments, the training process of the integrated shoe upper texture generation model includes: Construct a multi-dimensional training dataset, which includes at least shoe shape contour samples, shoe upper texture samples, color scheme samples, curved surface deformation samples, and weaving process constraint samples. The multi-dimensional training dataset is uniformly labeled and its features are extracted to obtain feature information of various types of samples; The extracted feature information of each type of sample is uniformly encoded and fused to obtain a feature vector of uniform dimension; A multi-task joint learning approach was adopted, with the optimization objectives of texture generation accuracy, color matching naturalness, splicing rationality, and distortion correction error as the optimization targets. An initial integrated shoe upper texture generation model was obtained through end-to-end overall training. The model parameters of the texture encoder, color generator, splicing learner, and deformation predictor are synchronously and iteratively updated using the backpropagation algorithm to obtain the final integrated shoe upper texture generation model.

[0055] In this embodiment, the training process of the integrated shoe upper texture generation model is an end-to-end multi-task joint learning process, aiming to enable the model to simultaneously master texture generation, color optimization, splicing processing, distortion correction, and process adaptation capabilities. Specifically, the training process includes the following core steps: First, a multi-dimensional training dataset is constructed. This dataset contains at least five types of samples: shoe outline samples, covering 3D shoe last data and corresponding 2D unfolded patterns for various shoe types (sneakers, casual shoes, high heels, etc.); shoe upper texture samples, including high-quality shoe upper texture patterns of different styles (geometric, floral, abstract, gradient, etc.); color scheme samples, including color matching data of various successfully launched shoe models and corresponding color psychology and trend labels; surface deformation samples, recording the actual distortion field data (i.e., displacement and stretching of each pixel) of 2D planar textures after mapping to 3D shoe lasts; and weaving process constraint samples, including process parameter sets under different weaving equipment, yarn specifications, and weaving densities, as well as corresponding texture clarity thresholds and weaveability labels. The above samples can be obtained through various means such as historical production data, public datasets, simulation generation, and manual annotation to ensure that the dataset has broad coverage and balanced distribution.

[0056] Secondly, the multi-dimensional training dataset was uniformly labeled and its features extracted to obtain feature information for various types of samples. The labeling work was completed under the guidance of professional designers and process engineers, specifically including texture labeling (texture type, density, direction, period), color labeling (primary color, secondary color, gradient range, color number, color gamut distribution), splicing labeling (upper parting, seam, boundary line, transition area), distortion labeling (surface stretching area, distortion degree, UV offset ground truth), and process labeling (woven area, out-of-bounds area, minimum aperture, maximum density). All labeled information, after quality verification, together with the original sample data, constituted a structured training sample library. Subsequently, feature extraction was performed on the labeled samples to form three different modalities of feature data: image features (texture images, color distribution maps, surface deformation maps, etc.), text / tag features (process parameters, structural descriptions, color numbers, and other tag information), and geometric features (curvature, UV coordinates, grid vertex positions, etc.).

[0057] Then, a multimodal feature fusion encoding method is adopted to map the above three types of heterogeneous features to the same representation space. For example, for image features, a pre-trained ResNet50 can be used as the backbone network to extract deep features, and then after multiple convolutional downsampling operations, it is finally compressed into a set of 256-dimensional feature vectors. This vector retains the key visual information of texture primitives, color distribution, and surface deformation field. For text or label features (such as weaveable area labels, texture type labels, Pantone color codes, etc. in process annotations), discrete labels can be mapped into dense vectors through the embedding layer, and then further transformed by the fully connected layer to finally output a 128-dimensional feature vector to capture process constraints and structural semantics. For geometric features (such as Gaussian curvature of shoe last surface, UV mapping coordinates, and grid vertex positions), all geometric coordinates are first normalized to eliminate scale and displacement differences, and then position encoding is introduced to inject spatial position information into the feature sequence, finally outputting a 128-dimensional feature vector to represent the local deformation characteristics and splicing relationship of the surface. After encoding the three types of features, they are simply concatenated to form a 512-dimensional original fusion vector. To eliminate distribution differences between different modal features (for example, the numerical ranges of image features and text features may differ significantly), LayerNorm (layer normalization) is used to normalize the 512-dimensional vector after concatenation, making the distribution of each feature dimension more consistent, thereby stabilizing the subsequent model training process and accelerating convergence. The final 512-dimensional unified feature vector serves as the input representation of the integrated shoe upper texture generation model, carrying comprehensive information on shoe geometry, texture appearance, color style, concatenation structure, distortion rules, and manufacturing constraints.

[0058] Next, a multi-task joint learning approach is adopted, with the optimization objectives of texture generation accuracy, color matching naturalness, stitching rationality, and distortion correction error. An initial integrated shoe upper texture generation model is obtained through end-to-end training. Specifically, it internally shares a core feature extraction network (i.e., the aforementioned multimodal fusion encoder), and four sub-networks branch off from this core: texture encoder (CNN), color generator (GAN), stitching learner (GNN), and deformation predictor (Transformer). During training, the four sub-tasks are optimized simultaneously, with each branch outputting prediction results independently, but sharing the underlying feature representation of the core network. The specific optimization objectives for the four tasks are as follows: Task 1 is texture generation, and the optimization objective uses a combination of regression loss and perceptual loss. The regression loss calculates the pixel-level absolute error between the generated texture and the real texture to ensure the accuracy of the basic structure; the perceptual loss uses a pre-trained VGG network to extract high-level semantic features to ensure that the generated texture is visually similar to the real texture.

[0059] Task 2 is color matching optimization, which aims to combine color difference loss and color gamut loss. Color difference loss can be calculated based on the CIEDE2000 color difference formula to determine the color difference between the predicted color and the labeled true value; color gamut loss constrains the distribution range of the generated colors to ensure that they fall within the preset reasonable color gamut boundaries, avoiding oversaturation or abnormal color shifts.

[0060] Task 3 focuses on the rationality of the stitching process, and the optimization objective employs a combination of boundary loss and continuity loss. Boundary loss measures the pixel differences at the seams between adjacent upper sections, penalizing the inconsistency between the predicted stitching boundary and the labeled boundary; continuity loss evaluates the smoothness of the texture on both sides of the stitching in the gradient domain, ensuring a natural texture transition without visual breaks.

[0061] Task 4 is distortion correction, and the optimization objective adopts a combination of UV error loss and curvature consistency loss. The UV error loss calculates the mean square error between the UV offset output by the deformation predictor and the true value of the distortion annotation, directly driving the model to learn an accurate 2D to 3D mapping; the curvature consistency loss constrains the deformed texture to maintain a reasonable degree of stretching in areas of drastic curvature changes in the shoe last, avoiding excessive distortion or wrinkles.

[0062] The total loss is calculated by weighted summation of the loss functions for the four tasks mentioned above. L 总 = λ1L 纹理 + λ2L 配色 + λ3L 拼接 + λ4L 畸变 Among them, L 纹理 For the total texture loss, L 配色 For total color matching loss, L 拼接 For the total splicing loss, L 畸变 λ1, λ2, λ3, and λ4 represent the total distortion loss and are weights.

[0063] The weights for each task can be set as learnable parameters that are adaptively adjusted during training, or a fixed equal weight allocation can be used, such as λ1=λ2=λ3=λ4=0.25, etc. Through the above multi-task joint learning, the model simultaneously takes into account texture details, color harmony, splicing smoothness, and surface adaptation accuracy in the shared feature space, avoiding suboptimal solutions or inter-task conflicts that may occur during single-task independent training.

[0064] Finally, the backpropagation algorithm is used to synchronously and iteratively update all model parameters of the backbone network and the four sub-networks. The specific training process is as follows: each iteration inputs a batch of standardized training data. After the backbone network extracts multimodal fusion features, the four sub-networks output their respective prediction results; then the total loss L is calculated. 总Next, starting from the loss layer, the gradient of each trainable parameter is calculated layer by layer upwards, and the AdamW optimizer is used to update all weights of the backbone network and the four sub-networks. During training, the initial learning rate is set to 1e. -4 The batch size is set according to the GPU memory capacity, and the overall training rounds are 50 to 100 epochs. After each training round, the texture generation accuracy (such as structural similarity (SSIM), peak signal-to-noise ratio (PSNR), color naturalness (Freche start distance (FID), distortion correction error (mean UV offset error)) and other indicators are monitored on the validation set. If the validation set loss no longer decreases for several consecutive epochs, an early stopping mechanism is triggered to stop training to prevent overfitting, and the best-performing model parameters on the validation set are saved. The core of the above process is that each forward propagation completes the parallel prediction of four tasks, and one backpropagation completes the synchronous update of all components (backbone + four sub-networks), realizing true end-to-end joint training. The final integrated shoe upper texture generation model obtained after training convergence only needs to input standardized preprocessed shoe outline data in actual use, and can output the initial shoe upper texture after texture generation, color optimization, stitching processing, distortion correction, and sharpness verification in one go, without any manual intervention.

[0065] In some embodiments, the splicing process includes: Based on the shoe upper seams, sole bonding boundaries, and weaving start and end areas, the system automatically identifies the shoe upper structure division lines. At the dividing line, texture gradient, smooth transition and seamless splicing operations are performed.

[0066] Specifically, the stitching learner in the integrated upper texture generation model first receives feature maps from the texture encoder and structural information implicit in the shoe outline data. The fusion criteria include three key boundaries: upper seam lines, i.e., the stitching paths between different pieces (such as the toe and body, or the body and heel counter); sole bonding boundaries, i.e., the edge contours of the area where the upper and sole meet, where the texture usually needs to fade or terminate to avoid visual conflict with the adhesive layer; and weaving start and end areas, i.e., the positions where yarn begins and ends during flyknitting or warp knitting, where density changes or yarn end treatments may exist, requiring transition zones to be reserved in the texture. Based on these three types of boundary information, the model uses a graph neural network (GNN) to partition and analyze the upper structure, automatically identifying all structural dividing lines with stitching or transition properties, and encoding these dividing lines into a graph structure as node edge weights.

[0067] After identifying the dividing lines, the model performs targeted fusion operations on the texture regions on both sides of each dividing line. The texture gradient operation generates a gradient weight mask along a direction perpendicular to the dividing line, gradually attenuating the pixel intensity of the texture on one side and gradually enhancing the texture on the other, thus eliminating visual discontinuities caused by sudden jumps. The transition smoothing operation uses filtering or interpolation methods to locally equalize the color and brightness in the neighborhood of the dividing line, ensuring that there are no obvious bright or dark edges at the splicing point. The seamless splicing operation further blends the features of the textures on both sides in a deep semantic space, maintaining not only pixel-level continuity but also ensuring that texture primitives (such as the shape of the mesh and the direction of the stripes) maintain a consistent direction and period when crossing the dividing line. These three operations can be executed in parallel or superimposed sequentially according to a preset order, ultimately outputting a visually continuous and integrated shoe upper texture at the splicing boundary. From a design perspective, the shoe upper, originally composed of multiple independent pieces, appears on the texture as if it were woven from a single piece.

[0068] In some embodiments, the distortion correction includes: Based on the curvature of the last surface in the shoe profile data, the stretching factor is calculated and the stretching area is determined. According to the stretching factor, UV remapping is performed on the shoe last surface to compensate for the texture ratio; By combining mesh deformation optimization and optical flow field adjustment, local distortion elimination and pixel-level smoothing are performed on the remapped texture.

[0069] Specifically, the deformation predictor in the integrated shoe upper texture generation model first receives standardized preprocessed shoe outline data and extracts the geometric information of the three-dimensional curved surface of the shoe last. The curvature includes principal curvature, Gaussian curvature, and mean curvature, used to quantify the degree of bending of the surface at each point. Regions with greater curvature (such as the toe sphere, heel arc, and arch sidewall) experience more severe stretching or compression deformation during traditional two-dimensional texture mapping. The deformation predictor automatically calculates the stretching region based on the curvature value corresponding to each grid vertex or pixel position. That is, continuous regions with absolute curvature values ​​exceeding a preset threshold are marked as stretching regions, and the stretching factor of each region in the U and V directions is further determined. The stretching factor is a two-dimensional vector, representing the scaling ratio of the texture coordinates in the longitudinal and latitudinal directions, respectively. For example, a stretching factor greater than 1 indicates that the texture needs to be enlarged in that direction to offset compression, and less than 1 indicates that the texture needs to be reduced to offset stretching.

[0070] Based on the obtained stretch factor, the deformation predictor performs UV remapping on the shoe last surface. Traditional UV mapping typically distributes 2D texture coordinates evenly to 3D mesh vertices, leading to texture scaling in areas of severe curvature. In this step, the deformation predictor recalculates the UV coordinates of each vertex or pixel using the stretch factor as a local scaling factor: for example, in areas with a stretch factor greater than 1, the UV coordinate step size is increased accordingly, distributing more texture pixels per unit surface area to compensate for the visual stretching caused by surface curvature; in areas with a stretch factor less than 1, the UV coordinate step size is decreased accordingly to reduce texture pixel accumulation and avoid excessive texture compression. The remapped UV coordinate set constitutes a distortion-corrected texture mapping relationship, ensuring that the projection of the originally distorted pattern onto the 3D surface presents a proportion close to the design intent.

[0071] After UV remapping, the remapped texture may still exhibit localized non-uniform distortions due to the stretching factor being a continuously changing quantity calculated based on curvature. For example, the texture transition between adjacent meshes may not be smooth enough, or there may be minor jagged edges. To further eliminate these residual distortions, the deformation predictor introduces mesh deformation optimization and optical flow field adjustment. Mesh deformation optimization uses the remapped UV mesh as the initial state and iteratively adjusts the positions of mesh vertices, constrained by the local consistency of feature points within the texture (such as stripe edges and mesh centers), to make the deformation gradient between adjacent mesh units as smooth as possible, eliminating sudden misalignments or wrinkles. Optical flow field adjustment, starting from pixel-level accuracy, calculates the pixel motion field between the remapped texture and the desired distortion-free texture. It obtains the offset vector for each pixel through bidirectional optical flow estimation, and then performs low-pass filtering and interpolation on this offset vector field to generate the final pixel-level smooth displacement. Applying the optical flow field to the remapped texture eliminates sub-pixel-level local distortions and minor misalignments, ensuring that the texture maintains straight lines and continuous patterns even in areas of curvature variation.

[0072] Through the multi-level distortion correction process described above, which includes "curvature-driven region detection → stretching factor calculation → UV remapping → mesh deformation optimization → optical flow field fine-tuning", the model can automatically complete comprehensive correction from macroscopic scale compensation to microscopic pixel adjustment during the texture generation stage. This ensures that the final output initial shoe upper texture, after being mapped onto the real shoe last, has accurate texture proportions, no local distortion, and overall visual smoothness, fundamentally solving the technical problem that curved surface deformation cannot be effectively corrected in traditional manual PS texture mapping.

[0073] In some embodiments, calculating the stretching region and determining the stretching factor based on the curvature of the shoe last surface in the shoe profile data includes: The three-dimensional shoe last surface is triangularly meshed, and the principal curvature of each mesh node is calculated. Calculate the Gaussian curvature and the mean curvature based on the principal curvature; The stretching factor is calculated based on the Gaussian curvature, and the stretching region is determined based on the average curvature. The calculation formula is as follows: , where S is the stretching factor, K is the Gaussian curvature, and ε is a minimum positive number.

[0074] Specifically, firstly, the 3D shoe last surface corresponding to the shoe model to be processed is triangulated, discretizing the continuous surface into a mesh model composed of a large number of triangular facets, with each facet sharing vertices. This meshing operation can be automatically completed using standard meshing algorithms (such as Delaunay triangulation or curvature-based adaptive meshing), ensuring that the mesh has higher resolution in areas with drastic curvature changes (such as the toe, instep, and heel) to guarantee computational accuracy.

[0075] Next, the principal curvatures are calculated for each grid node (i.e., the vertex of the triangle) in the triangular mesh. The principal curvatures are the maximum and minimum curvatures among all normal sections passing through a point on the surface, denoted as [insert values ​​here]. and ( The principal curvature can be calculated by fitting a local quadratic surface or by using a discrete curvature estimation algorithm based on the finite element method. The principal curvature of each vertex is then obtained. and Then, the Gaussian curvature was further calculated. and mean curvature .

[0076] in, , Gaussian curvature This reflects the overall degree and type of curvature of the surface at that point: >0 indicates an elliptical point (such as a convex area on a sphere). <0 indicates a hyperbolic point (such as a saddle-shaped region). ≈0 represents a parabolic point or a planar region. For the curved surface of a shoe last, this includes raised areas such as the toe and instep. >0, areas such as the arch of the foot and the lateral waist may be affected. <0, while the flatter side ≈0.

[0077] Next, based on Gaussian curvature Calculate the stretch factor S at each mesh vertex. Use the following formula: Where S is the stretching factor, K is the Gaussian curvature, and ε is a very small positive number used to prevent calculation overflow due to a zero denominator. The physical meaning of this formula is that the larger the absolute value of the Gaussian curvature, the more severe the surface curvature, and the more severe the stretching or compression deformation caused by traditional 2D textures in this area. Therefore, a larger stretching factor is needed to compensate for the texture scale. Specifically, S>1 indicates that the texture needs to be enlarged (i.e., the number of texture pixels per unit surface area) to offset the visual stretching caused by the surface bulge; S<1 indicates that the texture needs to be reduced; and when S=1, no adjustment is needed. Furthermore, this embodiment is not solely based on the stretching factor. Instead of using a single threshold to divide the absolute value, the Gaussian curvature is used. Determined stretching factor With mean curvature The positive and negative signs are jointly determined to determine the same... The system accurately distinguishes distortion attributes at the numerical level, providing a differentiated compensation strategy for subsequent UV remapping. The specific logic is as follows: Gaussian curvature Determines the stretch factor The numerical value ( ), used to quantify the severity of texture scaling misalignment; Mean curvature Used to distinguish the local concavity and convexity of curved surfaces, and to determine the specific type of texture distortion: when When the surface is identified as a convex curved surface (corresponding to the raised areas of the instep and toe on the shoe last), the texture in such areas is mainly due to stretching distortion, and correction should focus on texture amplification compensation; when When the surface is identified as a concave curved surface (corresponding to the concave instep and heel groove), the texture in this area is mainly characterized by compression distortion. Correction requires additional compression texture correction on top of the magnification compensation. When the plane is approximately developable (corresponding to the flat area of ​​the shoe body), it is determined to have almost no stretching or compression distortion, and only simple scaling or no correction is made.

[0078] Based on the aforementioned joint determination rules, this embodiment performs comprehensive classification of the mesh nodes on the curved surface of the shoe last. For example, a stretching factor threshold is set. As the dividing line between strong and weak distortion, and combined with The positive and negative signs are used to obtain the following classification results: when and At that time, it was classified as a convex surface high tensile compensation zone; when and At that time, it was classified as a concave strong compression correction region; when and At that time, it was classified as a convex surface weak tensile compensation zone; when and At that time, it was classified as a concave weak compression correction region; when or At this time, it is classified as a distortion-free flat region, and the UV coordinates remain unchanged. To illustrate with a specific example: suppose at a certain vertex of the instep region of the shoe last, the calculated Gaussian curvature K=0.3, and the average curvature... ( ), take ε=10 8 Then the stretching factor ≈1.83. Since 1.83 > 1.2 and This vertex is identified as a convex surface with strong stretching compensation, meaning that in subsequent UV remapping, the texture coordinate step size of this region needs to be increased by approximately 1.83 times to compensate for the texture stretching effect caused by the curved surface. For comparison, assume that a vertex in the concave region of the shoe opening has the same Gaussian curvature. But mean curvature ( ), at this time the calculated result It is also approximately 1.83, but due to The vertex is identified as a concave strong compression correction region. Its subsequent correction logic is not simply to increase the texture step size, but to introduce an additional compression texture correction term on the basis of amplification compensation in order to eliminate texture accumulation and compression distortion caused by concave curvature.

[0079] By traversing all mesh vertices and performing the above calculations and judgments, the precise distribution of the stretching region on the entire shoe last surface and the stretching factor corresponding to each region can be obtained, providing a quantitative basis for subsequent UV remapping.

[0080] In some embodiments, the multi-dimensional joint optimization process further includes a fusion process, which is used to consistently fuse the shoe outline data processed by different optimization paths to form a continuous initial shoe upper texture.

[0081] Specifically, the "fusion processing" in the multi-dimensional joint optimization process is a key step in ensuring that the final texture maintains a high degree of visual and structural consistency. This fusion processing is not a simple image overlay, but rather an intelligent integration at the data level of the shoe outline data processed by different optimization paths (e.g., texture generation, color optimization, splicing processing, or distortion correction) in the integrated shoe upper texture generation model. This eliminates inconsistencies that may be introduced by each branch of processing, thereby generating a seamless and coherent initial shoe upper texture.

[0082] In some embodiments, the fusion process specifically includes: first, acquiring at least two texture regions to be fused. These texture regions originate from different optimization processing paths. For example, the first region may have primarily undergone texture detail enhancement, while the adjacent second region may have focused on color saturation optimization or geometric distortion correction. Next, identifying and determining the feature differences of each texture region to be fused at a predetermined fusion boundary. Exemplarily, color feature differences (e.g., color difference gradients in RGB or Lab color spaces) and texture feature differences (e.g., differences in texture frequency, orientation, and other features extracted based on the gray-level co-occurrence matrix or Gabor filter) of these regions at predetermined overlap boundaries or splicing adjacencies can be calculated. The quantification results of these differences form the basis for subsequent fusion weight allocation. Then, the fusion weight of each texture region to be fused can be dynamically determined based on the calculated color and / or texture feature differences. Exemplarily, if the feature difference of a certain boundary region is small, it indicates that the texture at that location is... Figure 1 If the similarity is high, adjacent regions are assigned a more balanced fusion weight; conversely, if the differences are large, the weights are dynamically adjusted according to the gradient direction of the differences, so that regions with more continuous features dominate the fusion process, ensuring a visually smooth transition in the fusion result. Finally, at least two texture regions to be fused can be weighted and fused according to the dynamically determined fusion weights. During the fusion process, the fusion weights are distributed continuously and smoothly in space, rather than abruptly. In this way, texture data generated by different optimization paths can be smoothly transitioned and synthesized in overlapping or connecting areas, ultimately generating a continuous initial shoe upper texture that is consistent and unified in color, texture, and geometry.

[0083] In some embodiments, the real-time verification and optimization of the texture clarity of the initial shoe upper texture using the integrated shoe upper texture generation model and a preset weave clarity index includes: The preset weaving clarity index of the initial shoe upper texture is calculated in real time. The preset weaving clarity index is used to comprehensively evaluate the texture edge sharpness, detail retention and weaving restoreability. When the preset weaving clarity index is lower than the preset index threshold, the internal parameters of the integrated shoe upper mapping generation model are automatically adjusted for optimization until the calculated preset weaving clarity index is not lower than the preset index threshold.

[0084] Specifically, the process includes calculating the preset weave clarity index of the initial shoe upper texture in real time, and automatically adjusting the internal parameters of the model for iterative optimization when the threshold is not met.

[0085] It's important to note that the preset weaving clarity index is a comprehensive evaluation metric used to measure the performance of the initial upper texture in three dimensions. The first is texture edge sharpness, which refers to the clarity of pattern boundaries and line outlines in the texture. This is typically quantified by calculating the distribution of edge gradient amplitude or average gradient energy. Higher sharpness indicates more distinct boundaries formed by yarn color changes or jacquard movements during weaving, reducing the likelihood of blurring or jagged edges. The second is detail retention, referring to the integrity and recognizability of subtle texture features (such as mesh holes, fine stripes, and granular jacquard) in the texture. This can be evaluated using the signal-to-noise ratio or local variance of high-frequency components. Insufficient detail retention can lead to the loss of fine patterns in the woven upper, affecting the product's quality. The third is weaving reproducibility, reflecting whether the texture details in the texture can be actually reproduced under the constraints of the target weaving equipment's resolution, yarn fineness, and weaving density. This dimension involves comparing the spatial frequency of the texture with the Nyquist frequency of the equipment to determine if there are excessively dense textures beyond the equipment's capabilities, while also evaluating the degree of matching between the texture direction and the weaving direction. The calculation results of the above three dimensions are weighted and fused according to preset weights to obtain a comprehensive score between 0 and 1 or a percentage system, which is the preset weaving clarity index.

[0086] During real-time verification, the integrated shoe upper texture generation model first performs multi-scale analysis on the initial shoe upper texture, calculating the three sub-indicators mentioned above pixel by pixel or region by region to obtain a sharpness index distribution map of the entire texture. Then, this distribution map is compared with a preset index threshold. The preset index threshold can be an empirical value or a lower limit value calibrated through process experiments; for example, it can be set to 0.75 (out of 1.0), indicating that the overall texture sharpness score must not be lower than this value; otherwise, it is considered unqualified.

[0087] When the model detects that the pre-set weave sharpness index of a certain area or the whole in the initial shoe upper texture is lower than the pre-set threshold, the model automatically activates its internal optimization mechanism without manual intervention. This optimization mechanism is achieved by adjusting the internal parameters of the integrated shoe upper texture generation model, specifically including but not limited to the following: adjusting the feature extraction scale in the texture encoder to focus on higher resolution texture details; modifying the color smoothing parameters in the color generator to avoid edge blurring caused by overly smooth color transitions; adjusting the width of the transition zone in the stitching learner to make the texture gradient at the stitching boundary steeper to retain sharp edges; or adjusting the optical flow field smoothing intensity in the deformation predictor to reduce detail loss due to over-smoothing. These parameter adjustments are not random attempts, but rather based on pre-defined rules or through a supplementary optimization network (such as reinforcement learning or a gradient-based meta-optimizer) for rapid optimization. The model forms a closed-loop iteration between validation and parameter adjustment: after each adjustment, the sharpness index is recalculated; if it still does not meet the standard, adjustments continue until the sharpness index of all areas is not lower than the pre-set threshold. At this point, the texture is output as the target shoe upper texture.

[0088] Through this real-time verification and optimization mechanism, the present invention ensures that the final generated target shoe upper texture meets the process window requirements of the weaving equipment at the source, significantly reducing the probability of rework due to texture blurring or loss of details found in subsequent trial production, and realizing automated closed-loop control of texture clarity.

[0089] In some embodiments, the integrated shoe upper texture generation model further integrates a weaving constraint layer to embed physical constraints of the weaving process during texture generation, ensuring that the final output target shoe upper texture can seamlessly interface with the actual production equipment. Based on this, the method further includes: using the weaving constraint layer, according to preset weaving equipment parameters, performing pixel-level mapping and path alignment between the generated target shoe upper texture and the weaving path, needle bed distribution, and yarn feed trajectory of the weaving equipment.

[0090] Specifically, after generating the target shoe upper texture in step S300 or during its generation, the weaving constraint layer is activated. This layer pre-stores or receives the parameter configuration of the weaving equipment in real time. These parameters include, but are not limited to: the needle bed pitch (the number of needles per unit length, such as needles per inch), yarn diameter and material, the number and degrees of freedom of the yarn feeders, the yarn feeder movement trajectory output by the weaving path planning algorithm (represented as a two-dimensional or three-dimensional coordinate sequence in the machine tool coordinate system), and the needle selection sequence of each needle position during the weaving process. The above parameters vary depending on the equipment model, batch, and process requirements. The weaving constraint layer allows for the dynamic loading of corresponding parameter sets according to different equipment.

[0091] During the mapping and alignment process, the weaving constraint layer first establishes the transformation relationship between the pixel coordinate system of the target shoe upper texture and the machine coordinate system of the weaving equipment. Since textures typically exist as two-dimensional raster images, while weaving equipment executes needle position sequences and yarn feeder trajectories, these belong to different representation spaces. The weaving constraint layer achieves pixel-level mapping by mapping each pixel position in the texture to the specific needle number and knitting row number on the machine's needle bed. Specific operations include: calculating the physical length (mm / pixel) corresponding to each pixel based on the texture resolution and the actual physical dimensions of the shoe upper (obtainable from shoe outline data); then, combining this with the needle pitch, converting this physical length into a multiple of the needle spacing, thereby determining which needle and knitting row should perform the needle selection or jacquard action for a given pixel.

[0092] Furthermore, the weaving constraint layer performs path alignment, which involves spatially registering the texture direction in the target upper texture with the actual movement trajectory of the yarn feeder. The yarn feeder's movement trajectory determines the direction and order of yarn feeding. If the stripes, lines, or gradient directions in the texture are inconsistent with the yarn feeder's trajectory, the woven pattern will experience systematic shifts or distortions. To address this, the weaving constraint layer reads preset weaving paths (e.g., a zigzag path from the toe to the heel, or a local compensation path for a specific jacquard area) and calculates the position vector field of the yarn feeder at each moment. This vector field is then compared with the spatial gradient field of the texture, calculating the angular deviation and offset between them, and automatically adjusting the local coordinates of each pixel in the texture to ensure that the main direction of the texture (e.g., the direction of the stripes, the baseline of the jacquard pattern) remains parallel to or aligned with the yarn feeder's movement trajectory at the designed angle. For areas of uneven speed caused by the turning and acceleration / deceleration of the yarn feeder, the weaving constraint layer can also dynamically resample the texture, appropriately stretch the texture spacing in the deceleration section of the yarn feeder, and compress the texture spacing in the acceleration section to compensate for the difference between the actual speed of the yarn feeder and the ideal uniform speed assumption, so as to ensure that the woven pattern has uniform density and no local accumulation or thinning.

[0093] Through the aforementioned pixel-level mapping and path alignment, the weaving constraint layer transforms the abstract target shoe upper texture into a "process instruction diagram" that can be directly executed by the weaving equipment. This eliminates defects in traditional processes such as pattern misalignment, orientation distortion, and uneven density caused by the mismatch between the texture coordinate system and the equipment coordinate system. This layer is integrated within the unified model, enabling the conversion process from design intent to executable code to machine code without manual intervention, achieving a true digital closed loop.

[0094] Example 2 Corresponding to Embodiment 1 above, this application also provides a smart shoe upper texture generation device, which is used to implement the smart shoe upper texture generation method as described in any one of Embodiment 1. In this embodiment, content that is the same as or similar to that in Embodiment 1 above can be referred to the above description and will not be repeated hereafter.

[0095] Example 3 Corresponding to Embodiment 1 above, this application also provides a computer device, including: a processor and a memory, wherein the memory stores a computer program that can run on the processor, and when the computer program is executed by the processor, it executes the intelligent generation method for shoe upper texture provided in any of the above embodiments.

[0096] in, Figure 4 An exemplary computer device 1500 is shown, which may specifically include a processor 1510, a video display adapter 1511, a disk drive 1512, an input / output interface 1513, a network interface 1514, and a memory 1520. The processor 1510, video display adapter 1511, disk drive 1512, input / output interface 1513, network interface 1514, and memory 1520 can communicate with each other via a communication bus 1530.

[0097] The processor 1510 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solution provided by the present invention.

[0098] The memory 1520 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 1520 can store the operating system 1521 for controlling the operation of the electronic device, and the basic input / output system (BIOS) for controlling the low-level operations of the electronic device. Additionally, it can store a web browser 1523, a data storage management system 1524, and a device identification information processing system 1525, etc. The aforementioned device identification information processing system 1525 can be the application program that specifically implements the aforementioned steps in this embodiment of the invention. In summary, when implementing the technical solution provided by this invention through software or firmware, the relevant program code is stored in the memory 1520 and is called and executed by the processor 1510.

[0099] Input / output interface 1513 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.

[0100] Network interface 1514 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0101] The bus includes a pathway for transmitting information between various components of the device, such as processor 1510, video display adapter 1511, disk drive 1512, input / output interface 1513, network interface 1514, and memory 1520.

[0102] In addition, the electronic device can also obtain information on specific claim conditions from the virtual resource object claim condition information database for condition judgment, and so on.

[0103] It should be noted that although the above-described device only shows the processor 1510, video display adapter 1511, disk drive 1512, input / output interface 1513, network interface 1514, memory 1520, bus, etc., in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the present invention, and not necessarily all the components shown in the figures.

[0104] Example 4 Corresponding to Embodiment 1 above, this application also provides a computer-readable storage medium. In this embodiment, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description and will not be repeated hereafter.

[0105] The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the intelligent generation method for shoe upper textures as described above.

[0106] In some implementations of this application, when the computer program is executed by the processor, it can also implement the steps corresponding to the method described in Embodiment 1. For details, please refer to the detailed description in Embodiment 1, which will not be repeated here.

[0107] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0108] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0109] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0110] It should be noted that, unless otherwise specified, the relevant parts of the different embodiments described above can be referred to and learned from each other. This application will not repeat the same parts.

[0111] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0112] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0113] Although embodiments of this application have been shown and described above, it is to be understood that the various embodiments described herein are presented only to aid in understanding and teaching the claimed features. These embodiments are provided only as representative examples of embodiments and are not exhaustive and / or exclusive. It should be understood that the advantages, embodiments, examples, functions, features, structures and / or other aspects described herein should not be considered as limitations on the scope of the invention as defined by the claims or on the equivalents of the claims, and other embodiments may be utilized and modifications may be made without departing from the scope of the claimed invention. Various embodiments of the invention may suitably include, consist of, or substantially consist of suitable combinations of the disclosed elements, components, features, parts, steps, means, etc., rather than those specifically described herein. Furthermore, this disclosure may include other inventions not currently claimed but which may be claimed in the future.

Claims

1. A method for intelligently generating shoe upper textures, characterized in that, The method includes: Obtain the shoe outline data of the shoe model to be processed, and perform standardized preprocessing on the shoe outline data; The preprocessed shoe outline data is input into the pre-trained integrated shoe upper texture generation model for multi-dimensional joint optimization processing, and the initial shoe upper texture is output. The texture clarity of the initial shoe upper texture is verified and optimized in real time using the integrated shoe upper texture generation model and the preset weaving clarity index to generate the target shoe upper texture. A shoe upper texture file is generated based on the target shoe upper texture. The shoe upper texture file includes the target shoe upper texture in a preset image format and the corresponding weaving control parameters.

2. The intelligent generation method for shoe upper textures according to claim 1, characterized in that, The multi-dimensional joint optimization process includes at least one of texture generation, color matching optimization, splicing processing, and distortion correction.

3. The intelligent generation method for shoe upper textures according to claim 2, characterized in that, The integrated shoe upper texture generation model is an end-to-end fusion large model, and integrates a texture encoder, color generator, splicing learner and deformation predictor.

4. The intelligent generation method for shoe upper texture according to claim 3, characterized in that, The training process of the integrated shoe upper texture generation model includes: Construct a multi-dimensional training dataset, which includes at least shoe shape contour samples, shoe upper texture samples, color scheme samples, curved surface deformation samples, and weaving process constraint samples. The multi-dimensional training dataset is uniformly labeled and its features are extracted to obtain feature information of various types of samples; The extracted feature information of each type of sample is uniformly encoded and fused to obtain a feature vector of uniform dimension; A multi-task joint learning approach was adopted, with the optimization objectives of texture generation accuracy, color matching naturalness, splicing rationality, and distortion correction error as the optimization targets. An initial integrated shoe upper texture generation model was obtained through end-to-end overall training. The model parameters of the texture encoder, color generator, splicing learner, and deformation predictor are synchronously and iteratively updated using the backpropagation algorithm to obtain the final integrated shoe upper texture generation model.

5. The intelligent generation method for shoe upper texture according to any one of claims 2 to 4, characterized in that, The splicing process includes: Based on the shoe upper seams, sole bonding boundaries, and weaving start and end areas, the system automatically identifies the shoe upper structure division lines. At the dividing line, texture gradient, smooth transition and seamless splicing operations are performed.

6. The intelligent generation method for shoe upper texture according to any one of claims 2 to 4, characterized in that, The distortion correction includes: Based on the curvature of the last surface in the shoe profile data, the stretching factor is calculated and the stretching area is determined. According to the stretching factor, UV remapping is performed on the shoe last surface to compensate for the texture ratio; By combining mesh deformation optimization and optical flow field adjustment, local distortion elimination and pixel-level smoothing are performed on the remapped texture.

7. The intelligent generation method for shoe upper texture according to claim 6, characterized in that, The step of calculating the stretching factor and determining the stretching region based on the curvature of the shoe last surface in the shoe shape contour data includes: The three-dimensional shoe last surface is triangularly meshed, and the principal curvature of each mesh node is calculated. Calculate the Gaussian curvature and the mean curvature based on the principal curvature; The stretching factor is calculated based on the Gaussian curvature, and the stretching region is determined based on the average curvature. The formula for calculating the stretching factor is as follows: , where S is the stretching factor, K is the Gaussian curvature, and ε is a minimum positive number.

8. The intelligent generation method for shoe upper texture according to any one of claims 1 to 4, characterized in that, The multi-dimensional joint optimization process also includes a fusion process, which is used to consistently fuse the shoe shape contour data after processing through different optimization paths to form a continuous initial shoe upper texture.

9. The intelligent generation method for shoe upper texture according to any one of claims 1 to 4, characterized in that, The process of real-time verification and optimization of the texture clarity of the initial shoe upper texture using the integrated shoe upper texture generation model and preset weave clarity index includes: The preset weaving clarity index of the initial shoe upper texture is calculated in real time. The preset weaving clarity index is used to comprehensively evaluate the texture edge sharpness, detail retention and weaving restoreability. When the preset weaving clarity index is lower than the preset index threshold, the internal parameters of the integrated shoe upper texture generation model are automatically adjusted for optimization until the calculated preset weaving clarity index is not lower than the preset index threshold.

10. The intelligent generation method for shoe upper texture according to any one of claims 1 to 4, characterized in that, The integrated shoe upper texture generation model also integrates a weaving constraint layer; the method further includes: Through the weaving constraint layer, based on preset weaving equipment parameters, the generated target shoe upper texture is mapped and aligned pixel-level with the weaving path, needle bed distribution, and yarn feed trajectory of the weaving equipment.

11. The intelligent generation method for shoe upper texture according to claim 10, characterized in that, The weaving control parameters include at least a pixel-level color mapping map, a yarn color correspondence table, jacquard control instructions, weaving start point coordinates, yarn feeder movement path, and alignment marks to prevent pattern offset.

12. A smart device for generating shoe upper textures, characterized in that, The device is used to implement the intelligent generation method for shoe upper textures as described in any one of claims 1-11.

Images