Migration method of textile pattern generation style

Abstract

The invention relates to the technical field of textile, in particular to a textile pattern generation style migration method. Comprising the following steps: adopting a textile pattern structure and style separation coding mechanism; introducing a pattern rule structure field PSF, and introducing a style representation mapping domain SFD to perform high-dimensional deconstruction on color distribution density and texture frequency response extraction; a joint orthogonal deembedding network is utilized to maximize the independence of the structure and the style in the submerged space; adopting style component reconfigurable mapping; further refining the style characteristics into five types of core components; an inductive weight bias function is introduced, so that the system understands aesthetic tendency to make migration strategy adjustment; constructing a structure constraint graph matching mechanism; keeping the similarity of the pattern structure field PSF in the submerged space; introducing a topology order-preserving loss function, and keeping the original topology continuity for the repetitive unit arrangement; deformation of the structure caused by style replacement is limited through a structure and style cooperative adjustment curve; and generating a network by adopting dynamic adaptation of a style migration pattern.

Description

Technical Field

[0001] The present invention relates to the technical field of textiles, and particularly to a method for migrating textile pattern generation styles. Background Art

[0002] Although certain progress has been made in the current application of textile pattern generation driven by the development of image processing, computer vision, and deep learning, there are still obvious deficiencies and technical bottlenecks at multiple levels such as industrial actual requirements, textile process adaptability, and style expression control, restricting its promotion and application in scenarios such as high precision, customization, and practical implementation.

[0003] Firstly, most existing methods are built on the general architectures of traditional image style transfer networks such as Gram matrix transfer, neural style networks, or generative adversarial networks. These methods focus on the overall style distribution convergence with image-level texture and color statistics as the core, ignoring the "structural logic", "repetition rules", and "local functionality" unique to textile patterns. For example, traditional fabric patterns often have strict pixel arrangements, symmetry requirements, or texture periods, which are difficult to model in existing style transfer methods, easily leading to errors inapplicable to weaving such as structural misalignment, incoherent pattern repeat units, and boundary pixel overflow, thus affecting the reduction effect of subsequent printing or jacquard. Secondly, current methods generally adopt the "overall style to overall image" transfer method and cannot achieve semantic deconstruction and independent control of style components. In the actual textile pattern design process, users often hope to only transfer certain style elements, such as color tone, line decoration, background texture, while keeping other elements unchanged. This "directional style transfer" requirement is almost unsolvable in traditional methods because their core expression unit is the overall image rather than semantic components, resulting in overly strong or completely out-of-control style transfer, with the generated results deviating from the design intent and seriously affecting the user experience and design efficiency. In addition, in the perceptual modeling of style transfer, existing methods rely too much on "pixel similarity" and "texture statistical consistency" indicators in computer vision to measure the quality of style expression, ignoring more complex subjective preference factors in human aesthetics, such as cultural semantic tendencies, emotional tones, and style personalities. This "technology-driven detachment from the aesthetic context" transfer method often results in the embarrassing situation of "visually realistic but aesthetically distorted" style of the generated patterns, which is particularly prominent in commercial brand design or cultural product pattern creation.

[0004] Furthermore, most traditional image style transfer methods are static one-stage generation methods, lacking a process-adjustable structure, unable to achieve progressive control and mid-course repair of the style generation process. Once the generation fails, it can only be re-entered, resulting in a waste of time and resources, and showing extremely high uncontrollability in the user interaction experience. Especially in industrial scenarios, if the style transfer algorithm cannot provide a process perception and feedback adjustment mechanism, designers cannot intervene and optimize at a certain stage during the generation process, leading to low actual deployment efficiency and poor interpretability. To sum up, there are still many drawbacks in the existing textile pattern style transfer methods at multiple levels. Summary of the Invention

[0005] The object of the present invention is to provide a method for transferring the generation style of textile patterns, so as to solve some of the drawbacks and deficiencies pointed out in the background art.

[0006] The method for transferring the generation style of textile patterns adopted by the present invention to solve its above technical problems includes: S1. Adopt a textile pattern structure and style separation and encoding mechanism:

[0007] S1.1. Introduce a pattern rule structure field PSF, encode the repeating units, arrangement patterns, and geometric layouts in the pattern as a structure tensor; at the same time, introduce a style representation mapping domain SFD to perform high-dimensional deconstruction by extracting the color distribution density and texture frequency response.

[0008] S1.2. Use a joint orthogonal de-embedding network to maximize the independence of the structure and style in the latent space, forming two independently controllable encoding streams.

[0009] S2. Adopt a style component reconstructable mapping:

[0010] S2.1. Further refine the style features into five categories of core components: main color mapping, texture microstructure, decorative linear semantics, edge style performance, and contrast domain; use a learnable weight matrix to construct a style factor scheduler, which can perform combined reconstruction on the five categories of components.

[0011] S2.2. Introduce a perceptual weight bias function to enable the system to understand the aesthetic tendency and make adjustments to the transfer strategy.

[0012] S3. Construct a style adaptation mapping:

[0013] S3.1. Construct a structure constraint graph matching mechanism; maintain the similarity of the pattern structure field PSF in the latent space; introduce a topological order-preserving loss function to keep the original topological continuity of the repeating unit arrangement.

[0014] S3.2. Restrict the deformation caused by style replacement to the structure through a structure and style collaborative adjustment curve; adapt the style flow density according to the fabric properties including plain weave and twill to avoid interfering with subsequent weaving.

[0015] S4. Use a dynamic adaptation generation network with style transfer patterns:

[0016] S4.1. Design a multi-stage progressive generator PPG, which is divided into four stages: sketch generation, initial stylization, intermediate adjustment, and convergence optimization; in each stage, introduce an error feedback adjustment factor RCF to repair the dimensions that deviate from the target structure or style.

[0017] S4.2. At the terminal output stage, add a textile perception discriminator to evaluate whether the pattern has fabric surface naturalness according to the real texture samples of the fabric, and strengthen the learning of realism.

[0018] Furthermore, the style component reconstructable mapping method includes:

[0019] Perform five-category semantic decomposition on the style features in the style image: main color mapping, texture microstructure, decorative linear semantics, edge style performance, and contrast domain. After the style features are extracted by different channel extractors respectively, they are converted into independently controllable embedding tensors as the basic input for subsequent fusion; perform quantitative modeling on each type of style component to construct a component response function:

[0020]

[0021] Where:

[0022] is the comprehensive response intensity of the i-th type of style component in the image; Ω i is the perceptual area of the i-th component in the style image; φ i (x) is the basic visual response of this component at position x, including color concentration and texture change value; θ i (x) is the change rate of this = component, used to reflect its dynamics and activity; α i , β i are the weighted parameters that respectively control the static significance and change influence of this component.

[0023] Furthermore, the style component reconstructable mapping method includes:

[0024] Adopt a style factor scheduler to dynamically regulate the activation intensity of each type of style component and the influence relationship between them; according to the pattern content semantics and structural layout, learn and adjust the coupling relationship between style components to enhance the adjustability of style transfer; and express the linkage between style components, define a non-linear response function:

[0025] Ψ ij (x,y) = γ·exp(-|ρ i (x)ρ j (y)| δ)·ζ ij

[0026] in:

[0027] Ψ ij (x,y) represents the degree of linkage response between the i-th and j-th style components at positions x and y; ρ i (x) is the gradient response of style component i at position x; δ is the nonlinear deformation coefficient of the coupling curve, which controls the influence of the difference between different components on the response function; γ is the global coupling modulation coefficient, which controls the overall strength of the coupling response; ζ ij is the coupling weight parameter of style factors i and j.

[0028] Furthermore, the style component reconfigurable mapping method includes:

[0029] Introducing the perceptual bias function, based on user historical interactions, scene cultural labels, and semantic style labels, a nonlinear bias mapping between aesthetic emotions and style factors is established, and a dynamic adjustment style fusion strategy is constructed; Perceptual bias function:

[0030] B i (u,c)=κ i ·(1+tanh(μ u ·σ i (c)λ i ))

[0031] in:

[0032] B i (u,c) is the bias gain coefficient of the i-th style factor under user u and cultural label c; μ u is the strength of the user's aesthetic preference for style; i (c) is the recommendation degree of the current culture / context c for the i-th style component; i is the neutral threshold of style component i; κ i is the coefficient that controls the overall bias amplification or suppression; tanh(·) is the hyperbolic tangent function used to model the nonlinear boundary of aesthetic decision-making, which is suitable for expressing the gradual changes in emotions and cultural choices.

[0033] Furthermore, the method for constructing a dynamic adaptive generation network of the style transfer pattern includes:

[0034] A multi-stage progressive pattern generation framework is adopted, and the pattern generation process is divided into four progressive stages: sketch generation, initial stylization, intermediate adjustment and convergence optimization. Each stage is responsible for solving the dimension problem; the structure is controlled in the early stage, the style is embedded in the middle stage, and the detail compensation and style fusion are performed in the later stage. The generation state of each stage is uniformly modeled and converged, and the stage progressive fusion function is defined:

[0035]

[0036] Among them:

[0037] P s (x) is the pattern style fusion value at pixel position x in the final output image, representing the integration result of all-stage style information; τ ∈ [0, 1] is the normalized stage progress representing the pattern generation process; ξ(x, τ) is the style mapping response generated at pixel position x in the τ-th stage; Δ τ (x) is the style or structure difference between this point in the τ-th stage and the output image of the previous stage; η τ is the learning weight for controlling the influence intensity of the style mapping in the τ-th stage; ω τ is the correction amplitude for controlling the error term to the output result of this stage, expressing the deviation convergence trend.

[0038] Furthermore, the method for constructing a dynamic adaptive generation network for the style transfer pattern includes:

[0039] Introduce an error feedback adjustment factor mechanism RCF. After each stage ends, according to the deviation between the actual output and the target style / structure, adjust the generation behavior of the next stage; use the historical output style deviation trend to perform weighted adjustment on the style channels of the next stage to form a problem-driven style correction mechanism; model through the following function:

[0040]

[0041] Among them:

[0042] R i (t) is the adjustment output intensity of the i-th style dimension at time step t; χ i (t) is the actual response intensity of the system to the i-th style dimension in the current stage, reflecting the current activation degree of this dimension; is the response change rate of the style dimension at time s; α i is the coefficient reflecting the static importance of this dimension in the adjustment function; β i is the adjustment weight of the historical fluctuation term, used to control the correction of style drift or instability; t represents the stage time step in the current generation process, linearly advancing according to the generation process; s is the integration time variable, used to capture the cumulative effect of style fluctuations in past stages.

[0043] Furthermore, the method for constructing a dynamic adaptive generation network for the style transfer pattern includes:

[0044] Build a textile perception discriminator module to detect whether the finally generated pattern has the naturalness of the fabric surface. By comparing the differences between the generated pattern and the real fabric texture samples, a perception model is established in the dimensions of texture continuity, edge weavability, and color landing to score and feedback the final result; for the quantitative perception of the fabric surface naturalness, design a similarity scoring function:

[0045]

[0046] Where:

[0047] Q(x,y) represents the naturalness score between position x in the pattern and position y in the real fabric sample; is the local texture feature vector extracted from the pattern image at position x; is the texture feature vector of the real fabric sample at position y; is the vector difference between the local part of the pattern and the real fabric sample in the texture layer, representing the matching degree; δ is the non-linear contrast enhancement coefficient, which determines whether the micro-differences are amplified; γ is the difference penalty amplitude, which controls the sensitivity of the score; ζ is the normalization ratio coefficient of the overall score, which controls the final scoring range.

[0048] Advantages of the present invention:

[0049] By decomposing the complex style into five core components (main color mapping, texture microstructure, decorative linear semantics, edge style performance, contrast domain), the limitation of "overall style black box processing" in traditional style transfer is broken, enabling the system to have the ability to independently model, schedule, and fuse each type of style attribute, and improving the interpretability and controllability of style transfer. The stage-by-stage generation framework combined with the error feedback adjustment factor mechanism enables the pattern to achieve an orderly evolution of structure dominance, style penetration, error repair, and continuous optimization throughout the process from sketch to final style fusion, significantly solving common industry problems such as pattern structure misalignment, style drift, and detail abruptness.

[0050] By introducing a perceptual bias function, a non-linear mapping relationship is established between user aesthetic preferences, cultural backgrounds, semantic style labels, and style components, realizing the transformation of style generation towards "personality-driven" and "semantic consistency", enabling the system to adapt to different users and different market demand style design directions. Through the textile perception discriminator module to evaluate and optimize the pattern in dimensions such as texture continuity, color reducibility, and structure weavability, the pattern generation has achieved a leap from "good-looking" at the image level to "truly usable" at the fabric level, greatly improving the application success rate of the generated pattern in process links such as digital printing and jacquard weaving. Description of the Drawings

[0051] Figure 1Flowchart of the method for migrating the textile pattern generation style of the present invention.

[0052] Figure 2 Flowchart of the method for reconstructable mapping of style components of the present invention.

[0053] Figure 3 Flowchart of the method for constructing a dynamic adaptation generation network for style-migrated patterns of the present invention. Detailed implementation manners

[0054] The following makes a detailed description of the specific implementation manners of the present invention with reference to the accompanying drawings.

[0055] Refer to the attached Figure 1, the method for transferring the style of textile pattern generation of the present invention, step S1 adopts the textile pattern structure and style separation coding mechanism, aiming to effectively extract the structural information and style information with independent control ability from the source pattern, so as to provide a controllable, stable and accurate feature basis for the subsequent style transfer. S1.1 first introduces the pattern rule structure field (PatternStructureField, PSF), which is a high-dimensional structure tensor expression method used to structure the spatial layout information in the pattern. It contains the basic configuration features such as the repeated unit form, arrangement method, geometric coordinate logic, etc. of the pattern. For example, some elements in the fabric pattern appear in a repeated grid distribution, and their arrangement period and unit size are explicitly recorded in the PSF in a way of embedding mathematical rules, so as to ensure that the structural information will not be lost due to style changes during the migration process. At the same time, the style representation mapping domain (Style Feature Domain, SFD) is introduced. This domain extracts style features that can be used for high-dimensional description by analyzing the visual attributes of multiple dimensions such as color distribution density, main color gradient law, texture particle frequency, contrast intensity, etc. in the pattern. SFD is not a single set of style labels, but constructs a continuous style vector space, which enables the system to understand the composition ratio, hierarchical distribution and spatial scalability of style elements, and then realizes the disassembly, fusion and migration of style; then, the joint orthogonal disembedding network is introduced in S1.2. This network is a deep learning architecture with structural decoupling function, which can project the input pattern encoding vector into two mutually orthogonal potential spaces respectively, which are used to carry the structure tensor (by PS F encoding), and the other is used to carry the perceptual visual flow formed by the style distribution vector (constructed by SFD). By imposing orthogonal constraints and distribution preservation constraints in the latent space during the training process, the network can maximize the independence of structural information and style information in the encoding dimension, avoid the structure being disturbed by the style and the style being distorted by the structure, and finally form two decoupled and independently operable encoding streams. The core value of this mechanism is that it establishes a dynamic mapping foundation with constant structure and variable style for subsequent style migration, and also provides encoding-level support and controllable interface for complex operations such as multi-style superposition and local style control, ensuring that the generated pattern has creative freedom at the visual level without destroying the original pattern composition rules and textile adaptability.

[0056] Step S2 adopts style component reconfigurable mapping, aiming to solve the problems of rough overall style transfer, insufficient control force, and single style expression in traditional style transfer methods. Through in-depth analysis and modular expression of style information, the system is equipped with a more delicate and adjustable transfer ability. S2.1 first refines the style features in the semantic dimension and decomposes them into five categories of core style components, namely main color mapping, texture microstructure, decorative linear semantics, edge style performance, and contrast domain. Among them, the main color mapping is responsible for defining the color tone, color temperature tendency, and color matching strategy of the pattern, which is the basis of the first impression of the overall style perception. The texture microstructure contains information at the microscopic level such as fine texture particles, texture, and dot-like repeating units in the pattern, affecting the simulation of the visual touch on the fabric. The decorative linear semantics is used to express the guiding contour lines, border decorations, or geometric lines in the pattern, which is an important component for maintaining the sense of structure and aesthetic rhythm. The edge style performance involves the processing methods of the pattern boundary, such as blurring, edge pressing, and embossing, which play a decisive role in the naturalness and visual fusion degree of the pattern boundary. The contrast domain measures the degree of visual conflict between colors, light and dark, or graphics inside the pattern, determining the sense of tension and hierarchy of the pattern. The above five components are all represented as independent style embedding vectors and form a style factor scheduler by constructing a learnable weight matrix. The scheduler can dynamically activate, inhibit, or recombine the five categories of style components according to information such as the input structure diagram, style target diagram, and even user preferences, so as to achieve the fusion transfer or local adjustment between multiple styles. The core of its principle lies in that the style transfer is no longer a holistic imitation, but becomes a "generation and reconstruction based on the combination of style particles", endowing the system with higher transfer controllability and expression freedom. Immediately afterwards, in S2.2, a perceptual weight bias function is introduced, aiming to incorporate subjective intentions such as user aesthetic preferences, cultural preferences, and visual habits into the system decision-making logic. The function assigns different bias weights to each style component in the style factor scheduler by analyzing high-level information such as user interaction data (such as preference labels, style scores, selection behaviors), cultural style labels (such as Chinese style, minimalist, retro, childlike, etc.), and emotional semantic dimensions (such as soft, rational, impact). The bias function presents a non-linear adjustment relationship, retaining the default generation ability of the system and being able to flexibly adjust the style synthesis strategy according to the user profile. For example, for the "Oriental minimalist" aesthetic preference, the system will reduce the response intensity of decorative linearity and contrast to create a sense of blank space, while for the "retro European style", it will enhance the main color saturation and edge decoration to enhance the sense of heaviness and decoration. The core significance of this perceptual bias mechanism is to enable the system to change from passively transferring styles to actively understanding "how people perceive styles", so as to achieve "aesthetic-driven style generation" in the true sense, improving the personalization, diversity, and cultural fit of the pattern output, and making it more applicable and market-adaptable in actual textile design, product customization, brand visual language and other scenarios.

[0057] Step S3 is to construct a style adaptation mapping. Its core goal is to solve the problem of style information interfering with the structural stability during the pattern style transfer process, ensuring that the generated result not only has style expressiveness but also retains the geometric composition rules of the original pattern, so as to meet the rigid requirements of fabric design for the rigor and repetitive continuity of the pattern structure. First, in S3.1, a structure constraint graph matching mechanism is constructed. Based on the Pattern Structure Field (PSF), this mechanism abstracts the information such as the position, arrangement sequence, periodic function, and direction relationship of each repeating unit in the pattern into a high-dimensional structure vector graph and performs real-time comparison and matching on it in the latent space of style transfer, that is, it requires a high similarity distribution between the structure tensors before and after style transfer, so that the overall skeleton of the pattern is not damaged macroscopically in the system-generated result. Further, a topological order-preserving loss function is introduced, whose role is to perform topological continuity constraints on the connection relationships between the repeating units in the pattern, preventing structural disorders such as unit misalignment, rotational distortion, and spacing breakage during the style enhancement process. This loss function does not directly act on the pixel level of the image but constrains its adjacency matrix, connectivity, and boundary cycle relationships in the graph embedding space of the structure field, so as to retain the arrangement order and graphic element geometric rhythm of the pattern, thus achieving the core requirement of "changing style while keeping structure constant". Then, in S3.2, to further alleviate the problems of detail distortion and local expansion caused by style enhancement to the pattern structure, a structure-style co-regulation curve mechanism is introduced. This mechanism models the degree of interference of each type of style component on the structure when acting on the pattern area as an adjustable response curve. For example, the enhancement of a certain type of high-density texture style in the decoration area will cause deformation of the graphic element edges. This mechanism can adjust the density, direction, and gradient diffusion range of the style particle flow in real time based on the structure intensity map, making it avoid the structure boundary or reduce the weight in the structure-sensitive area, thus achieving the "differentiated friendly adaptation" of style to structure. In addition, under different fabric properties (such as plain weave, twill, satin, etc.), due to the different ways of pattern bearing by the fabric structure, the transferable density of the style also needs to be dynamically adjusted. Therefore, the system introduces a fabric property perception model to adjust the style flow density and spatial scalability according to the input fabric type. For example, more detailed decoration styles are allowed to be concentratedly expressed on plain weave fabrics, while high-frequency graphic elements need to be reduced in twill fabrics to prevent misaligned interlacing. This module ensures that the style transfer not only conforms to the aesthetic expression at the visual level but also is compatible with the actual process bearing capacity of different fabrics, ensuring that the pattern has structural reducibility, cycle stability, and process adaptability in terminal processes such as digital printing and jacquard weaving, thus truly pushing the style transfer algorithm from the image level to the usable level of textile engineering.

[0058] Step S4 adopts a dynamic adaptation generation network for style transfer patterns, aiming to solve problems in traditional pattern style transfer methods such as one-time generation, uncontrollable process, and lack of dynamic coordination between style and structure. Through a phased progressive optimization and adaptive feedback adjustment mechanism, it realizes the full-process intelligent evolution of the pattern from structure construction to style fusion and then to enhanced realism. In S4.1, the system designs a multi-stage progressive generator PPG (Progressive Pattern Generator), which divides the entire pattern generation process into four stages: sketch generation, initial stylization, intermediate adjustment, and convergence optimization. In the sketch generation stage, the system mainly constructs the basic contour and pixel distribution framework of the pattern according to the input structure field information, providing a geometric skeleton guarantee for style transfer; in the initial stylization stage, the extracted main style factors (such as main color, main texture) are transferred to the sketch to form a preliminary framework for style expression; in the intermediate adjustment stage, local style details and structure boundaries are refined, enhanced, or repaired to establish a closer fusion relationship between the style and the structure of the pattern; in the convergence optimization stage, operations such as color transition, edge sharpness enhancement, and decorative line fine-tuning are concentrated to improve the overall visual integrity and industrial implementation friendliness; to ensure that each stage can self-correct style deviation or structural perturbation problems, the system introduces an error feedback adjustment factor RCF (Residual Correction Factor) in each stage. This mechanism analyzes the deviation between the generated result of the current stage and the target style map or structure field, identifies abnormal responses in dimensions such as color, texture, contrast, and pixel density, and dynamically adjusts the activation weights and parameter distributions of the corresponding generation modules in the next stage through the feedback mechanism. For example, if the system detects that the linear contour of the pattern in the intermediate stage is deformed or broken, it enhances the generation ability of the decorative linear semantic channel in the convergence stage to complete pixel closure or path smoothing; the introduction of RCF enables the system to have the ability of phased error correction and self-adjustment, not only improving the generation quality, but also enabling the pattern style transfer process to have the ability of continuous optimization and semantic consistency maintenance; in S4.2, to further ensure that the final output pattern is not only visually beautiful in style, but also has reducibility and process matching in actual fabric presentation, the system introduces a textile perception discriminator in the terminal output stage. This module uses real fabric texture samples as the learning reference, evaluates the naturalness of the output pattern on the fabric surface from dimensions such as texture continuity, color level transition, and pattern repeat cycle consistency, and feeds the discrimination result back to the generation network for end-point fine-tuning to ensure that there will be no problems such as edge tearing, particle chaos, or color block jumping that are not suitable for weaving or printing due to excessive stylization when the output pattern is presented on the fabric surface. The introduction of this module extends the traditional image-level style transfer to a complete industrial pattern design path with "perception - adaptation - feedback", making the generated pattern not only "look like fabric", but also "be truly applicable to fabric", significantly improving the industrial usability and commercial implementation value of the system generation result.

[0059] Example 1:

[0060] Combined with the attached Figure 2 In this embodiment, in the pattern intelligent design department of a textile design company, a designer received a customer customization task. The customer is an enterprise mainly engaged in high-end women's clothing brands and hopes to design a fabric pattern for a new Chinese-style spring and summer dress. It is required that the pattern reflects light dark green in the main color tone, retains the symmetrical flower pattern in the traditional Chinese composition in terms of structure, and the overall style should be fresh and natural while having a certain sense of modern aesthetic lines. The customer provided an "inspiration picture", which is an ink painting of orchids in the style of Chinese painting, with highly extensible lines, naturally transitioning edges, and low contrast. The designer decided to apply the style component reconstructable mapping technology in the textile pattern generation style transfer method, using the inspiration picture provided by the customer as the style source picture, and combining with the existing standardized structure templates to generate a new textile pattern.

[0061] During the execution process, the system first decomposes the style image into five types of style components, and the specific decomposition results are as follows:

[0062] 1) The main color tone mapping identifies the main colors as light dark green (H: 120°, S: 0.25, V: 0.8) and a transitional background in the ash cyan color system;

[0063] 2) The texture microstructure detects scattered dot-like ink patterns presented by fine brushstrokes, with a high-frequency but low-density texture arrangement;

[0064] 3) In the decorative linear semantics, the main line is extracted as the trend of medium-thick arc-shaped orchid leaves, with strong directionality and obvious linear tension;

[0065] 4) The edge style performance mainly shows the characteristics of feathered boundaries and gradient water marks, presenting a strong transitional blur effect;

[0066] 5) The contrast domain analysis result is a low-contrast feature, with a gentle change in light and dark in the overall picture and no strong boundary mutations.

[0067] The system converts the above five types of components into five groups of embedding tensors respectively, and numerically models each type of style component in the perception region Ω i and measures it using the following style response function:

[0068]

[0069] Set the parameter range as follows:

[0070] α i (Static significance weight) is set in the range of 0.5 - 2.0, reflecting the basic expression strength of the style component in the overall;

[0071] β i(Dynamic fluctuation sensitivity) The setting range is 0.1–1.0, which is used to control the "activity" of the components, that is, whether there are rapid changes or the complexity of details;

[0072] φ(x): The system extracts the local style value from the image. For example, the color concentration of the area with the main color of grayish green is mapped to 0.75, and the gray value of the decorative line is 0.82;

[0073] θ i (x): Calculated based on the local gradient of texture changes. For example, the change rate in the texture particle area is about 0.12, and the change rate in the linear stroke area is 0.25.

[0074] Taking the decorative linear semantic dimension (i.e., i = 3) as an example, the system analyzes the response of this dimension in the region Ω3 (covering the main line area of the orchid). The length of this region is set to 300 pixels, and a total of 150 local sampling points are set. Set α3 = 1.5, β3 = 0.3. Among the sampling points, the average φ3(x) = 0.82 and θ3(x) = 0.25. Then the response of this style component can be approximately estimated as:

[0075]

[0076] This value, as the style response intensity of the decorative linear semantics, is passed into the scheduler and fused and modeled together with other style factors. Subsequently, according to the customer's aesthetic labels of "neo - Chinese style", "natural simplicity and elegance", and "low contrast", the system retrieves a set of perceptual bias function coefficients from the user preference library, corresponding to:

[0077] B 主色调 = 1.2,

[0078] B 纹理 = 0.9,

[0079] B 线性语义 = 1.1,

[0080] B 边缘 = 1.0,

[0081] B 对比度 = 0.7,

[0082] Finally, the system performs weighted reconstruction on the response values of the five types of style factors in the scheduler, enhances the expression of the main color and line contour during the pattern generation process, moderately suppresses the sharpness of the pattern edge and high - contrast details, and at the same time compresses the texture orderly, and outputs an ink - wash style pattern with a complete structure, unified main color, soft boundary, and clear linear direction.

[0083] After the designer completed the decomposition and quantitative modeling of the five types of style components of the ink-wash style pattern, they entered the crucial stage of style component fusion. They decided to use the style factor scheduler mechanism to further optimize the style combination expression, so that the final pattern not only has the visual style expected by the customer, but also is more natural and coordinated in terms of structural logic and aesthetic performance. At this stage, the system based on the five types of style factor response tensors generated in the previous stage dynamically regulates the activation intensity of each component in the local area of the pattern, and considers its linkage coupling relationship with other style factors to ensure that the style fusion does not cause conflicts or structural misalignments, and improves the naturalness of the transition at the detail level. The system first analyzes the pattern structure semantic map and finds that the structure of this new Chinese orchid pattern has obvious "line-plane combination" characteristics, that is, the linear contour (orchid leaves) occupies the dominant space composition, and the color and texture details are mainly concentrated in the background area. Therefore, in the style fusion process, the regulatory ability of decorative linear semantics on the main color and edge style should be strengthened, and the interference of texture factors on the element edges should be suppressed, so as to achieve a clear primary and secondary style coordination logic. The system models the coupling relationship between style factors through the following non-linear response function:

[0084] Ψ ij (x,y) = γ·exp(-|ρ i (x)·ρ j (y)| δ )·ζ ij

[0085] where Ψ ij (x,y) represents the linkage response intensity of style components i and j at pattern positions x and y; ρ i (x), ρ j (y) are the gradient response values of components i and j at this position, representing the "active change degree" of the style characteristics at this point; δ is the non-linear deformation control coefficient, with a setting range of 1.2–2.5, used to control the deformation degree of the response curve. A lower value enhances the local smooth reaction, and a higher value enhances the ability to suppress dissimilar coupling; γ is the global modulation coefficient, with a setting range of 0.8–1.5, used to adjust the overall coupling response intensity; ζ ij is the coupling coefficient between components, obtained from statistical learning in the training set, with a range of 0.3–1.2. Here, a typical coupling term is selected for substitution calculation to verify the feasibility of the scheme. For example, it is set that the i = 3 term is "decorative linear semantics" and the j = 1 term is "main color mapping", and the coupling situation in the main orchid line area in the middle of the pattern (coordinate position x = 120, y = 130) is analyzed. After system analysis, the change rate of linear semantics at this point is ρ3(120) = 0.65, and the color gradient response of the main color at this point is ρ1(130) = 0.45. Considering the new Chinese design requirement that the main color unfolds around the structural line, the style coupling should be moderately enhanced. The system selects δ = 1.6, γ = 1.2, ζ3,1 =0.9, substitute into the formula:

[0086] Ψ 3,1 (120,130)=1.2·exp(-|0.65·0.45| 1.6 )··0.9=1.2·exp(-0.292 1.6 )·0.9

[0087] Ψ 3,1 ≈1.2·exp(-0.155)·0.9=1.2·0.856·0.9≈0.924

[0088] The coupling response result is close to 1, indicating that the linear contour of this area is highly coordinated with the main color tone. The system will give priority to enhancing the expression of the main color tone in this area in the scheduler, thus forming a new Chinese style layout of "lines wrapped around the main color". At the same time, the texture and contrast components are selected in the pattern background area for coupling evaluation to control unnecessary interference. For example, the background area position is x = 40, t = 55, and ρ2(40) = 0.38 and ρ5(55) = 0.51 are set. Since the customer does not want obvious graininess and strong contrast, the system sets a higher δ = 2.2 to suppress the coupling effect, and sets the coupling coefficient to ζ 2,5 =0.5, substitute it into the calculation and we get:

[0089] Ψ 2,5 (40,55)=1.2·exp(-|0.38·0.51| 2.2 )·0.5=1.2·exp(-0.194 2.2 )·0.5

[0090] Ψ 2,5 ≈1.2·exp(-0.048)·0.5=1.2·0.953·0.5≈0.571

[0091] This value is significantly lower than 0.9, indicating that the system will weaken the linkage effect of texture and contrast in this area, and the background transition part of the generated pattern will maintain a soft, grainless, low-contrast style tone, thereby avoiding interference with the main expression of the orchid.

[0092] After the designer completed the extraction of style components, response modeling, and coupling scheduling among factors, they entered the stage of perceptual aesthetic adaptation. As a high-end women's clothing brand that emphasizes cultural tone, the core audience of the client is urban women aged 25 - 40, who prefer aesthetic styles of "neo-Chinese", "natural elegance", and "implicit decoration". Therefore, the system activated the perceptual bias function mechanism, aiming to make the fusion strategy of the final pattern style not only reflect the rationality of style ratio in terms of technology but also the aesthetic expectations of the target user group. This mechanism constructs the following bias function by analyzing the matching relationship between the user portrait u, cultural context label c, and style components:

[0093] B i (u, c) = κ i ·(1 + tanh(μ u ·σ i (c) - λ i ))

[0094] The system extracts the aesthetic preference coefficient μ from the client's past pattern selection behaviors. The clustering analysis results given by the internal user preference database show that this client attaches a relatively high importance to "main color consistency" (μ u = 0.88), has a medium to high acceptance of "decorative linear semantics" (μ u = 0.72), and is more sensitive to "contrast", preferring low contrast (μ u = 0.35). Corresponding to the current project's cultural context of "neo-Chinese spring women's clothing", the system sets the recommended degree σ u (c) of the style factors labeled "neo-Chinese" in the aesthetic corpus as follows: the main color component σ1(c) = 0.85, the linear semantics σ3(c) = 0.72, the edge transition σ4(c) = 0.65, and the contrast σ5(c) = 0.25. The neutral threshold λ i is set in the range of 0.3 - 0.7, reflecting the sensitivity of different style components to change from the "neutral" to the "preferred" state; the bias amplification coefficient κ i ranges from 1.0 - 1.8, used to control the expression weight of the overall bias effect in the final fusion strategy. i Taking the "main color mapping" style component (i = 1) as an example, the current aesthetic preference intensity of the client is μ

[0095] = 0.88, the cultural recommendation weight is σ1(c) = 0.85, the neutral threshold λ1 = 0.5, and the bias amplification coefficient κ1 = 1.6. Substituting these values into the calculation gives: u = 0.88, the cultural recommendation weight is σ1(c) = 0.85, the neutral threshold λ1 = 0.5, and the bias amplification coefficient κ1 = 1.6. Substituting these values into the calculation gives:

[0096]

[0097] The results show that the main color components will be significantly enhanced during the style fusion process. The system adjusts the weight of the main color channel in the scheduler to nearly double, ensuring that the main color of the pattern remains within the range of customer preferences and avoiding color deviation or overly cluttered background colors. Taking the "contrast and contrast domain" component (i = 5) as an example, μ u = 0.35, σ5(c) = 0.25, λ5 = 0.45, κ5 = 1.2. Substituting into the formula:

[0098]

[0099] The result is less than 1. The system will lower the weight of the contrast and contrast component in the final pattern to about 78% of the original value, and avoid using black lines or color block mutations with too high contrast in the generated pattern to conform to the customer's brand tone of "low impact, elegant and natural". After the system applies the perceptual bias function to all five types of style components, it outputs a preliminary pattern draft with the characteristics of "strong liquidity of the main color in the ink painting sense, blurred and soft boundaries, moderately clear lines, and overall low contrast". After secondary fine-tuning by the designer, the output version meets the customer's satisfaction standard.

[0100] Example 2:

[0101] Combined with the attached Figure 3 , in this embodiment, in the project background of a textile design company customizing a "new Chinese style ink painting orchid pattern" for a high-end women's clothing brand, after the designer team completes the pre-steps such as style extraction, style component response modeling, and perceptual bias adjustment, it enters the final pattern generation stage. The dynamic adaptive generation network construction method of this style transfer pattern is applied to solve the problems of style jump, detail fragmentation, and structural misalignment caused by the one-time mapping mode in the traditional pattern style transfer system. The generation network adopts a multi-stage progressive pattern generation framework, dividing the entire pattern construction process into four progressive stages: sketch generation, initial stylization, intermediate adjustment, and convergence optimization. Each stage focuses on a core goal: the sketch generation stage emphasizes the accuracy and symmetry of the main structure, such as identifying the main direction of orchid leaves, the axis of symmetry, the repetition period of flower shapes, etc.; the initial stylization stage quickly injects the main color and texture background atmosphere to ensure that the overall atmosphere is close to the target style; the intermediate adjustment stage is responsible for fine-tuning the graphic elements and micro-fusing the style factors, strengthening the transparency of the lines and spreading the local ink painting texture; the convergence optimization stage performs global balance, boundary feathering, detail compensation, and enhancement of texture continuity, making the final pattern overall coordinated, highly weavable, and aesthetically natural.

[0102] To uniformly control the output connection, style response coherence, and error repair effect between the four stages, the system uses the following stage progressive fusion function for modeling:

[0103]

[0104] Among them:

[0105] P s P(x) is the fused style expression value of the pattern at pixel position x finally.

[0106] τ ∈ [0, 1] is the progress of the normalization stage, corresponding to sketch generation (0.0–0.25), initial stylization (0.25–0.5), intermediate adjustment (0.5–0.75), and convergence optimization (0.75–1.0) respectively.

[0107] ξ(x, τ) is the style response value at this pixel position in the τ-th stage (the range is set to 0–1, indicating the proportion of the style influence in this stage).

[0108] Δ τ Δ(x) is the style / structure difference between the τ-th stage and the previous stage at this position. The larger the value, the more obvious the style jump or the greater the generation error at this point.

[0109] η τ is the style influence weight, and the value range is 0.3–1.5. The higher the value, the stronger the dominant style generation effect in this stage.

[0110] ω τ is the error correction weight, and the value range is 0.1–1.0. The higher the value, the more the system tends to use this stage to repair the residuals of the previous stage.

[0111] In actual generation, taking the pixels x = (210, 75) in the area of the central main leaf of the orchid pattern as an example, the main lines are drawn in the sketch stage in this area (ξ(x, 0.15) = 0.65), the ink wash background and light ink tone are applied in the initial stylization stage (ξ(x, 0.4) = 0.75, Δ 0.4 (x) = 0.12), the line clarity and texture overlay are optimized in the intermediate adjustment stage (ξ(x, 0.65) = 0.82, Δ 0.65 (x) = 0.08), and the structural symmetry and color balance are strengthened in the convergence stage (ξ(x, 0.9) = 0.78, Δ 0.9 (x) = 0.04). The corresponding weight parameters set by the system are: η 0.4 = 1.2, η 0.65 = 1.4, ω 0.4 = 0.6, ω 0.65 = 0.4, ω 0.9 = 0.2. The pattern fusion expression value is estimated by numerical integration as follows:

[0112]

[0113] P s(x) ≈ (1.2·0.75 + 0.6·0.0144)·0.25 + (1.4·0.82 + 0.4·0.0064)·0.25 + (1.1·0.78 + 0.2·0.0016)·0.25

[0114] P s (x) ≈ (0.9 + 0.0086)·0.25 + (1.148 + 0.0026)·0.25 + (0.858 + 0.00032)·0.25P s (x) ≈ 0.2271 + 0.2876 + 0.2146 = 0.7293

[0115] This result indicates that the fusion value of the pattern in this area reaches a style expression density of 72.9%, which is within the high-quality generation range, and the error converges well between stages and the style transition is natural. Compared with the standard generation when the system does not use this function mechanism, the response value fluctuates violently (between approximately 0.61 - 0.83), and there is a breakpoint phenomenon in texture fusion.

[0116] The design team has successfully completed sketch construction, style injection, and convergence optimization using a multi-stage progressive generator. However, the designer noticed that during the adjustment process of the intermediate stage, some key style components - especially the decorative linear semantics and texture microstructure - showed style deviation near the structural contour, manifested as: the linear trend of the orchid leaves was overly softened in the intermediate stage, resulting in the inability to restore the due backbone tension in the convergence stage. At the same time, the particle density of the texture microstructure changed significantly between the front and back stages, affecting the unity of the pattern layout. To solve such problems, the design team initiated the error feedback adjustment factor mechanism (RCF), which weights and adjusts the style activation intensity in the upcoming next-stage generation process by capturing the response trend of the style dimension in the historical stage, specifically modeled by the following function:

[0117]

[0118] where R i (t) is the adjustment output intensity of the i-th type of style dimension at the stage time step t, χ i (t) is the current response of the system to this style dimension, represents the change rate of this dimension at the historical time point s, α i is the static importance coefficient of this dimension, β i is the historical fluctuation adjustment weight. In actual values, α i is set in the range of 0.8 - 1.5 (used to represent the proportion intensity of the basic expression), and β i is set to 0.1 - 0.8 (to control whether to amplify the impact of style fluctuations on the system adjustment behavior).

[0119] In the current pattern, taking the decorative linear semantic dimension (i = 3) as an example, the team observed that its response value χ3(0.5) = 0.62 at stage t = 0.5 (mid-term adjustment stage), showing a decrease compared to 0.79 in the previous stage. At the same time, the integral of the squared response derivative within the historical time period was calculated as follows:

[0120]

[0121] This indicates that the stability of this style component during the generation process is poor and there is a rapid downward trend. To prevent this trend from continuing to expand, the system sets the static weight of this dimension to α3 = 1.2 and the fluctuation adjustment coefficient to β3 = 0.5, and substitutes them into the formula:

[0122] R3(0.5) = 1.2·0.62 + 0.5·0.046 = 0.744 + 0.023 = 0.767

[0123] This result is slightly lower than the expected value (the target activation intensity is 0.85). The system immediately feeds this value back to the generation engine in the convergence stage, increases the convolution kernel response intensity of the tensor channel in the decorative linear semantic module by 1.2 times, and suppresses the interference of texture noise in the same area to prevent the lines from weakening further.

[0124] Taking the texture microstructure dimension (i = 2) as another example, its style response at stage t = 0.65 is χ2(0.65) = 0.51, and the historical integral of the squared derivative is 0.034. Setting α2 = 1.0 and β2 = 0.3, then:

[0125] R2(0.65) = 1.0·0.51 + 0.3·0.034 = 0.51 + 0.0102 = 0.5202

[0126] This value is lower than the expected threshold of 0.6 for style fusion scheduling. The system increases the texture coverage rate of this dimension by 10% in the final stage, but limits the coverage range to non-edge areas to avoid damaging the main line performance. This "value-driven self-adjustment mechanism" enables the system to not only identify abnormal style generation but also judge whether it is a structural deviation based on the style trend, so as to intervene in advance and correct it precisely.

[0127] After the "New Chinese Ink Orchid Pattern" project customized by the design company for the client entered the final stage, the design team faced a crucial decision: Although the pattern had passed multiple controls in terms of structural logic, style coordination, and user aesthetic bias visually, whether it could be truly implemented on the fabric and maintain the naturalness of the texture, the clarity of the edges, and the accuracy of color transition in the actual fabric material was still the ultimate basis for determining whether the solution could be put into production. To ensure that the generated pattern had industrial-grade fabric naturalness expressiveness, the system introduced a textile perception discriminator module. This module no longer simply evaluated the pattern effect from the visual similarity, but starting from "whether it could be restored by the fabric", established a comprehensive perception scoring mechanism for three dimensions: texture continuity, edge weavability, and color implementation.

[0128] In the application, the system selected a set of real fabric sample data as a reference. The samples were from the fabric pattern data of historical best-selling models after high-definition scanning, including orchid shapes, light ink gray-green color matching, and the fine-grained texture composition of the real fabric under high-frequency distribution. When comparing the consistency between the generated pattern and the real samples, the system called the following fabric naturalness scoring function for local feature comparison modeling:

[0129]

[0130] In the function, Q(x,y) is the scoring value, ranging from 0 to ζ, represents the local texture feature vector extracted from the generated pattern at position x, represents the texture feature extracted from the real fabric sample at position y, and the difference vector represents the texture expression deviation between the two. The parameter settings are as follows: δ∈[1.2,2.4], which is used to control the amplification degree of small differences, γ∈[2.0,5.0] controls the penalty intensity of the score for texture inconsistency, ζ∈[0.8,1.2] adjusts the final scoring range, and when set to 1, it represents the original output, and being larger or smaller can be used for weight redistribution.

[0131] Taking the pixel block x=(180,90) of the main leaf part in the central area of the pattern corresponding to the fabric sample area y=(178,92) as an example, the local texture feature vector of the generated pattern extracted in this area is The real fabric texture is Then the difference is:

[0132]

[0133] Selecting the parameters δ = 1.8, γ = 3.2, ζ = 1.0 and substituting them into the scoring function, we get:

[0134]

[0135] Since the scoring function is "perceived difference", the smaller this value is, the smaller the difference and the higher the naturalness. Therefore, the system classifies it into the "highly restored" area. The system repeats the above calculations to traverse all key positions of the pattern (such as the edges, color mutations, and the centers of flower patterns), and calculates the overall naturalness score of the fabric, forming the following statistics:

[0136] Mean value of texture continuity index: 0.023±0.007 (lower value is better)

[0137] Mean value of edge transition score: 0.031±0.011

[0138] Mean value of color landing score: 0.027±0.009

[0139] Comprehensive naturalness score Q total : 0.0287 (the difference from the best sample is less than 0.03)

[0140] The system sets the industrial production threshold as Q total <0.045, and the current pattern fully meets the standard. Based on this, the designer judges that this version of the pattern not only meets the new Chinese aesthetic in terms of vision, but also has a high degree of restoration in fabric presentation.

[0141] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for transferring the style of textile pattern generation, characterized in that Including the following steps: S1. Adopt a textile pattern structure and style separation coding mechanism: S1.

1. Introduce a pattern rule structure field PSF, and encode the repeating units, arrangement patterns, and geometric layouts in the pattern as structure tensors; at the same time, introduce a style characterization mapping domain SFD to perform high-dimensional deconstruction by extracting color distribution density and texture frequency response; S1.

2. Use a joint orthogonal de-embedding network to maximize the independence of the structure and style in the latent space, forming two independent and controllable coding streams; S2. Adopt a style component reconstructable mapping: S2.

1. Further refine the style features into five core components: main color mapping, texture microstructure, decorative linear semantics, edge style performance, and contrast domain; use a learnable weight matrix to construct a style factor scheduler, which can perform combined reconstruction on the five components; S2.

2. Introduce a perceptual weight bias function to enable the system to understand the aesthetic tendency and make migration strategy adjustments; S3. Construct a style adaptation mapping: S3.

1. Construct a structure constraint graph matching mechanism; maintain the similarity of the pattern structure field PSF in the latent space; introduce a topological order-preserving loss function to keep the original topological continuity of the repeating unit arrangement; S3.

2. Restrict the deformation caused by style replacement to the structure through a structure and style co-regulation curve; adapt the style flow density according to the fabric properties including plain weave and twill to avoid interfering with subsequent weaving; S4. Adopt a dynamic adaptation generation network for style-transferred patterns: S4.

1. Design a multi-stage progressive generator PPG, which is divided into four stages: sketch generation, initial stylization, intermediate adjustment, and convergence optimization; introduce an error feedback adjustment factor RCF in each stage to repair the dimensions that deviate from the target structure or style; S4.

2. At the terminal output stage, add a textile perception discriminator to evaluate whether the pattern has fabric naturalness according to the real texture samples of the fabric, and strengthen the realness learning.

2. The textile pattern generation style transfer method according to claim 1, characterized in that The style component reconstructable mapping method includes: Perform five-category semantic decomposition on the style features in the style image: main color mapping, texture microstructure, decorative linear semantics, edge style performance, and contrast domain. After the style features are extracted by different channel extractors respectively, they are transformed into independent and controllable embedding tensors, which are used as the basic input for subsequent fusion; perform quantitative modeling on each category of style components to construct a component response function: Where: is the comprehensive response intensity of the i-th style component in the image; Ω i is the perceptual region of the i-th component in the style image; φ i (x) is the basic visual response of this component at position x, including color concentration and texture change value; θ i (x) is the change rate of this component, used to reflect its dynamics and activity; α i and β i are the weighted parameters that respectively control the static saliency and change influence of this component.

3. The textile pattern generation style transfer method according to claim 2, characterized in that The style component reconstructable mapping method includes: Adopt a style factor scheduler to dynamically regulate the activation intensity of each category of style components and the influence relationship between them; learn and adjust the coupling relationship between style components according to the pattern content semantics and structural layout to enhance the adjustability of style transfer.

4. The textile pattern generation style transfer method according to claim 3, wherein The style component reconstructable mapping method includes: Introduce a perceptual bias function to establish a non-linear bias mapping between aesthetic emotions and style factors according to user historical interactions, scene culture tags, and semantic style tags, and construct a dynamic adjustment style fusion strategy; perceptual bias function: B i (u, c) = κ i ·(1 + tanh(μ u ·σ i (c)λ i )) Where: B i (u, c) is the bias gain coefficient of the i-th style factor under user u and cultural tag c; μ u is the aesthetic preference intensity of the user for the style; σ i (c) is the recommendation degree of the current culture / situation c for the i-th style component; λ u is the neutral threshold of style component i; κ i is the coefficient for controlling the overall bias amplification or suppression; tanh(·) models the non-linear boundary of aesthetic decision-making using the hyperbolic tangent function, which is suitable for expressing the gradual change of emotions and cultural choices.

5. The textile pattern generation style transfer method according to claim 1, characterized in that The method for constructing a dynamic adaptation generation network for style-transferred patterns includes: Adopt a multi-stage progressive pattern generation framework, divide the pattern generation process into four progressive stages: sketch generation, initial stylization, intermediate adjustment, and convergence optimization. Each stage is responsible for solving dimensional problems; and control the structure in the early stage, embed the style in the middle stage, and perform detail compensation and style fusion in the later stage, and uniformly model and converge the generation state of each stage.

6. The textile pattern generation style transfer method according to claim 5, characterized in that The method for constructing a dynamically adaptive generation network for style transfer patterns includes: Introduce an error feedback adjustment factor mechanism RCF. After each stage, according to the deviation between the actual output and the target style / structure, adjust the generation behavior of the next stage; utilize the style deviation trend of historical outputs to weight-adjust the style channels of the next stage to form a problem-driven style correction mechanism; model through the following function: Where: R i The adjusted output intensity of the i-th style dimension at time step t; χ i The actual response intensity of the system to the i-th style dimension at the current stage, reflecting the current activation degree of this dimension; is the response change rate of the style dimension at time s; α i is the coefficient reflecting the static importance of this dimension in the adjustment function; β i is the adjustment weight of the historical fluctuation term, used to control the correction of style drift or instability; t represents the stage time step in the current generation process, linearly advancing according to the generation process; s is the integration time variable, used to capture the cumulative effect of style fluctuations in the past stage.

7. The textile pattern generation style transfer method according to claim 6, characterized in that The method for constructing a dynamically adaptive generation network for style transfer patterns includes: Construct a textile perception discriminator module to detect whether the finally generated pattern has fabric naturality. By comparing the differences between the generated pattern and real fabric texture samples, establish a perception model in the dimensions of texture continuity, edge weavability, and color landing to score and feedback the final result; for the quantitative perception of fabric naturality, design a similarity scoring function: Where: Q(x, y) represents the naturalness score between the position x in the pattern and the position y in the real fabric sample; is the local texture feature vector extracted from the pattern image at position x; is the texture feature vector of the real fabric sample at position y; is the vector difference between the local part of the pattern and the real fabric sample in the texture layer, representing the matching degree; δ is the non - linear contrast enhancement coefficient, determining whether the micro - differences are amplified; γ is the difference penalty amplitude, controlling the sensitivity of the score; ζ is the normalization ratio coefficient of the overall score, controlling the final scoring range.

Images