Clothing pattern design method and system based on AI agent

Abstract

The application discloses a clothing pattern design method and system based on an AI agent, relates to the technical field of clothing pattern design, and comprises the following steps: calling a traditional pattern element deconstruction and reorganization algorithm by the AI agent to perform morphological atomic decomposition and reorganization on a source pattern, thereby generating a reorganized pattern primitive set; calling a clothing pattern multi-modal adaptive generation model to fuse text and visual features and output a multi-modal pattern layout candidate set; calling a clothing pattern color space correction algorithm to combine fabric base color and reflection characteristics to implement color gamut mapping and color drift compensation on the candidate set; generating a quantitative score sequence by using a pattern design effect quantitative evaluation platform; and performing joint iterative optimization on generation strategy parameters according to the score sequence by the AI agent, and finally outputting a clothing pattern. The application realizes collaborative optimization and high-fidelity color correction of clothing pattern design, and improves the overall coordination and design quality of the final pattern.

Description

Technical Field

[0001] This invention relates to the field of clothing pattern design technology, and in particular to a method and system for clothing pattern design based on AI intelligent agents. Background Technology

[0002] Clothing pattern design is a core creative link in the upstream of the textile and apparel industry chain, and its design quality directly affects the aesthetic value and market competitiveness of the end products. With the continuous growth of consumer demand for personalized and diversified clothing products, the traditional model of pattern creation relying on manual hand-drawing or computer-aided drawing software is gradually revealing its limitations in terms of efficiency and innovation. In recent years, deep learning technology has made significant progress in the field of image generation, providing a new technical path for the automated design of clothing patterns. The industry has begun to explore introducing algorithms such as generative adversarial networks, variational autoencoders, and diffusion models into the clothing pattern creation process, aiming to shorten the design cycle and expand the diversity boundaries of pattern forms. Existing automatic garment pattern generation solutions generally adopt an end-to-end deep generative model architecture. The typical implementation is as follows: First, a large number of historical garment pattern images and their corresponding text labels are collected to form a training dataset. Then, a conditional generative adversarial network or diffusion model is constructed, and the text description or style reference image is used as a conditional input to drive the generator to output pattern that meets the conditional constraints. Some solutions introduce color histogram matching or tone mapping functions after generation to make the color distribution of the generated pattern approach the color gamut range of the target garment fabric. Finally, the generated results are filtered and sorted by calculating peak signal-to-noise ratio, structural similarity index or aesthetic scoring model, and several candidate pattern designs are output for designers to manually select or further modify.

[0003] Existing technologies have the following drawbacks: Firstly, existing solutions treat pattern element deconstruction, pattern layout generation, color correction, and effect evaluation as independent sequential processing steps. There is a lack of a collaborative optimization mechanism based on quantitative feedback between these steps, causing errors in any step to accumulate progressively and become impossible to compensate for in subsequent steps, resulting in poor overall coordination of the final output pattern. Secondly, existing color correction methods only perform coarse-grained stretching or compression of the global color gamut, failing to consider the unique bidirectional reflective distribution characteristics and spatially non-uniform color differences of the fabric substrate. This leads to a significant deviation between the color rendering of the corrected pattern on the actual fabric and the screen preview, reducing the engineering usability of the design solution. Summary of the Invention

[0004] In order to overcome the shortcomings and deficiencies of existing technologies, this invention provides a method and system for designing clothing patterns based on AI intelligent agents.

[0005] The technical solution adopted in this invention is a clothing pattern design method based on an AI agent, comprising the following steps: S1, responding to design requirement instructions, the AI ​​agent calls a pattern element deconstruction and recombination algorithm to perform morphological atom decomposition on the source pattern, generating a recombined pattern primitive set with texture gradient distribution characteristics; S2, based on the recombined pattern primitive set, a multimodal adaptation generation model for clothing patterns is called to map text description features and visual style features to a common embedding space, and a multimodal pattern layout candidate set is fused and output; S3, for the multimodal pattern layout candidate set, a clothing pattern color space correction algorithm is called, based on the target clothing fabric... Based on the base color parameters and light reflection characteristics, color gamut mapping and non-rigid color drift compensation are performed on each pattern in the candidate set, and the color-corrected pattern is output. S4, the color-corrected pattern is input into the pattern design effect quantitative evaluation platform, and the quantitative scoring sequence of each pattern is generated by calculating the pattern complexity, color contrast and structural symmetry index. S5, the AI ​​agent performs joint optimization iteration on the recombined pattern primitive set, multimodal fusion weight and color correction intensity according to the quantitative scoring sequence, and updates the pattern generation strategy parameters. S6, S2 to S5 are repeated until the iteration termination condition is met, and the final clothing pattern is output.

[0006] Furthermore, the traditional pattern element deconstruction and recombination algorithm includes the following expression: ,in, This represents the set of recombined pattern primitives. Indicates the first A recombined pattern primitive, Indicates the domain of the source pattern. Represents the morphological atom decomposition operator. Indicates the spatial position of the source pattern The gray intensity function at that location, Indicated by Center position parameter and with The first of the scale parameters A Gaussian receiver domain function Domain Lebesgue measure differential on, This indicates the total number of recombinant primitives; subsequently, for each primitive... Apply deconstruction and reconstruction mapping: ,in, This represents the optimized result after deconstruction and recombination. A recombination element, Represents the gradient operator. Indicates the first Texture transformation functions, Indicates the first The curvature control parameters corresponding to the transformation, This represents the total number of transformation functions. express Norm, This indicates the product operation.

[0007] Furthermore, the multimodal adaptation and generation model for clothing patterns includes the following expression: ,in, This represents the fusion latent feature matrix in the common embedding space. This represents the hyperbolic tangent activation function. This represents the text modality projection weight matrix. This represents the text description feature vector. Represents the visual modality projection weight matrix. Represents visual style feature vectors. Represents the bias vector; based on the fused latent feature matrix Further, a multimodal pattern layout candidate set is generated: ,in, This represents the candidate set of multimodal pattern layouts. Indicates the first A matrix of candidate layout patterns, Indicates the generation of the first The number of fusion atoms used when selecting candidate patterns. Indicates the first The candidate pattern is the first one Attention weight coefficients of each fused atom This represents the natural exponential function. Indicates the first The candidate pattern is the first one The cumulative affinity matrix of the fused atoms, Indicating continuous variables The affinity density function at that point, Denotes the upper limit constant of integration. , .

[0008] Furthermore, the color space correction algorithm for clothing patterns includes the following expression: ,in, This represents the color matrix of the pattern output after color correction. , Indicates the total number of cascaded calibration stages. Indicates the first Stage gamut mapping diagonal matrix, This indicates the input pattern color matrix. This represents the Hadamard product operation. Indicates the first The spectral reflectance compensation matrix is ​​constructed based on the base colorimetric parameters of the target apparel fabric. Indicates the first Stage-based non-rigid color drift compensation bias matrix.

[0009] Furthermore, the AI ​​agent includes a graphics processor cluster, a tensor processing unit array, and a high-bandwidth storage unit. The source pattern originates from a pattern database in vector graphics format. The design requirement instructions include pattern theme tags, clothing category tags, and style preference tags. The multimodal fusion weights are dynamically adjusted by the AI ​​agent based on a confidence allocation mechanism. The adjustment step size decreases exponentially with the number of iterations, and the decay factor ranges from 0.92 to 0.98. The color correction intensity is controlled by a color gamut boundary constraint coefficient, which is 0.75 in the early stage of iteration and gradually converges to 0.25 in the later stage of iteration. The iteration termination condition is that the sum of the absolute values ​​of the changes in two adjacent iterations of the quantized scoring sequence is lower than a preset threshold, which ranges from 0.015 to 0.025. The upper limit of the number of iterations is set to 80 to 120. After each iteration, the recombined pattern primitive set retains the top 60% to 75% of the primitives in terms of score from high to low to participate in clothing pattern generation, and discards the primitives in the bottom 25% to 40%.

[0010] Further, S2 includes the following sub-steps: S21, extracting the geometric contours and texture density distributions of each primitive in the recombined pattern primitive set into a structural descriptor sequence; S22, performing cross-modal attention alignment between the structural descriptor sequence and the text description features to generate an attention weight distribution map; S23, performing weighted aggregation of visual style features based on the attention weight distribution map to obtain a stylized structure fusion tensor; S24, inputting the stylized structure fusion tensor into a multilayer perceptron decoder to output a multimodal pattern layout candidate set.

[0011] Further, step S3 includes the following sub-steps: S31, extracting the principal components of the base chromaticity and the characteristic parameters of the bidirectional reflectance distribution function of the target garment fabric under a standard light source; S32, constructing a color gamut boundary constraint surface based on the principal components of the base chromaticity, and projecting the chromaticity values ​​of each pixel in the pattern candidate set into the constraint surface; S33, calculating the color difference compensation vector field under different viewing angles based on the characteristic parameters of the bidirectional reflectance distribution function, and applying directional drift correction to the projected chromaticity values; S34, performing local neighborhood smoothing interpolation on the drift-corrected pattern color matrix, and outputting the color-corrected pattern.

[0012] Further, S4 includes the following sub-steps: S41, performing grayscale co-occurrence matrix statistics on the color-corrected pattern to extract texture contrast and energy feature values; S42, converting the pattern to a uniform color space and calculating the variance and covariance of each color channel to quantify the harmony index; S43, performing Fourier shape descriptor extraction on the boundary contour of the pattern and calculating rotational symmetry score and mirror symmetry score; S44, performing a weighted linear combination of texture contrast, energy feature values, color harmony index, rotational symmetry score, and mirror symmetry score to generate a quantitative score sequence for each pattern.

[0013] Further, S5 includes the following sub-steps: S51, parsing the pattern corresponding to each score in the quantitative scoring sequence and extracting the multimodal fusion weight vector corresponding to the pattern with the highest score; S52, performing gradient-guided coarse adjustment on the current color correction intensity based on the color correction intensity parameter of the pattern with the highest score; S53, removing primitives with scores below the elimination threshold from the recombined pattern primitive set and supplementing new primitives by affine mutation based on the morphological characteristics of the primitives with the highest scores; S54, encapsulating the updated fusion weight vector, color correction intensity parameter, and recombined pattern primitive set into a new generation generation strategy parameter for the AI ​​agent to call and execute the next iteration.

[0014] A clothing pattern design system based on AI agents, applied to a clothing pattern design method based on AI agents, includes: a pattern primitive deconstruction and reconstruction engine, which receives source pattern data, internally runs a morphological atom decomposition protocol and a Gaussian receptive domain convolution extraction process, and outputs a reconstructed pattern primitive set; a multimodal fusion layout generator, whose input is coupled to the output of the pattern primitive deconstruction and reconstruction engine, receives the reconstructed pattern primitive set, and internally includes a text encoding channel and a visual encoding channel. The outputs of the two channels perform tensor splicing and nonlinear activation operations in a common embedding space, and output a multimodal pattern layout candidate set; and a color gamut drift correction processor, whose input is coupled to the output of the multimodal fusion layout generator, internally embedding a base chromaticity constraint module and a directional color difference compensation module, and performing color gamut mapping on the multimodal pattern layout candidate set. The system outputs a color-corrected pattern after radiation and drift correction; a pattern quality evaluator, with its input coupled to the output of the color gamut drift correction processor, includes a texture statistics calculation unit, a color distribution analysis unit, and a shape symmetry measurement unit. The weighted summation of the outputs of each unit generates a quantitative scoring sequence; an intelligent agent strategy iteration controller, with its input coupled to the output of the pattern quality evaluator, and its output coupled to the control ports of the pattern primitive deconstruction and reconstruction engine, the multimodal fusion layout generator, and the color gamut drift correction processor, dynamically adjusts the internal operating parameters of each engine, generator, and processor according to the quantitative scoring sequence; and a pattern output cache interface, with its input coupled to the termination determination port of the intelligent agent strategy iteration controller, is used to receive and store the final garment pattern when the iteration termination condition is met.

[0015] Beneficial Effects: This invention proposes a method and system for designing clothing patterns based on an AI agent. By organically coupling a traditional pattern element deconstruction and recombination algorithm, a multimodal adaptation generation model for clothing patterns, a color space correction algorithm for clothing patterns, a quantitative evaluation platform for pattern design effects, and an AI agent strategy iteration controller, a closed-loop collaborative optimization architecture is constructed, encompassing the entire chain from pattern primitive extraction to pattern generation, color correction, and quantitative evaluation. This solves the shortcomings of existing technologies where each stage is independent, leading to the gradual accumulation of errors that cannot be compensated. Based on the scoring sequence output by the quantitative evaluation platform, the AI ​​agent simultaneously performs joint optimization iterations on the composition of the recombined pattern primitive set, the allocation of the multimodal fusion weight matrix, and the parameter values ​​for color correction intensity. This allows the structural features generated in the pattern deconstruction stage to adaptively influence the layout decisions in the subsequent generation stage, while the color gamut constraint results in the color correction stage... Furthermore, it can reverse the style fusion tendency of the generation stage, forming a dynamic collaborative adjustment mechanism based on numerical feedback between each stage, avoiding the irreversible error amplification problem caused by unidirectional serial processing; at the same time, when performing color correction, the system fully incorporates the spectral reflectance compensation and non-rigid color drift compensation operations unique to the target clothing fabric substrate, and performs directional drift correction on the chromaticity value of each pixel through multi-stage cascaded integration operations, improving the color fidelity and viewing consistency of the corrected pattern on the real fabric, and effectively suppressing the color deviation of the design scheme from virtual preview to physical implementation; in addition, the quantitative evaluation platform conducts a comprehensive quantitative analysis of the pattern from multiple dimensions such as texture contrast, color harmony and symmetry scores, providing the intelligent agent with accurate iterative guidance signals, thereby quickly converging to the optimal pattern design result that combines aesthetic quality and engineering adaptability within a limited number of iterations. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 This is a flowchart of method step S2 of the present invention; Figure 3 This is a flowchart of method step S3 of the present invention; Figure 4 This is a flowchart of method step S4 of the present invention; Figure 5 This is a flowchart of step S5 of the method of the present invention; Figure 6 This is a diagram showing the system unit composition of the present invention. Detailed Implementation

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1 As shown, the clothing pattern design method based on AI intelligent agents includes the following steps: S1, responding to design requirement instructions, the AI ​​intelligent agent calls a pattern element deconstruction and recombination algorithm to perform morphological atomic decomposition on the source pattern, generating a recombined pattern primitive set with texture gradient distribution characteristics; S2, based on the recombined pattern primitive set, the clothing pattern multimodal adaptation generation model is called to map text description features and visual style features to a common embedding space, and the multimodal pattern layout candidate set is fused and output; S3, for the multimodal pattern layout candidate set, the clothing pattern color space correction algorithm is called, based on the base color of the target clothing fabric. Based on parameters and light reflection characteristics, color gamut mapping and non-rigid color drift compensation are performed on each pattern in the candidate set, outputting a color-corrected pattern; S4, the color-corrected pattern is input into the pattern design effect quantitative evaluation platform, and a quantitative scoring sequence for each pattern is generated by calculating pattern complexity, color contrast and structural symmetry indicators; S5, the AI ​​agent performs joint optimization iteration on the recombined pattern primitive set, multimodal fusion weight and color correction intensity according to the quantitative scoring sequence, and updates the pattern generation strategy parameters; S6, S2 to S5 are repeated until the iteration termination condition is met, and the final clothing pattern is output.

[0019] In step S1, responding to the design requirement instruction, which specifically includes pattern theme tags, clothing category tags, and style preference tags, the AI ​​agent, composed of a graphics processor cluster, tensor processing unit array, and high-bandwidth storage unit, retrieves matching source patterns from a vector graphics format pattern database based on the pattern theme tags in the design requirement instruction. It then calls a traditional pattern element deconstruction and reconstruction algorithm to perform a morphological atom decomposition operation on the source pattern. This morphological atom decomposition operation divides the source pattern into multiple independent morphological atoms according to the geometric contour curvature variation characteristics. Each morphological atom corresponds to a local structural unit in the source pattern. During the morphological atom decomposition operation, a Gaussian receptive field function with a scale of 0.3 to 0.7 is used to perform multi-scale filtering on the source pattern, extracting the texture gradient extreme points at each scale. As boundary anchors for morphological atoms, the regions between these anchors constitute the morphological atoms. After morphological atom decomposition, the algorithm clusters and recombines morphological atoms with similar gradient directions based on the texture gradient distribution direction of each morphological atom. During the clustering and recombination process, a similarity threshold of 0.65 to 0.85 is set, and morphological atoms with similarity higher than this threshold are grouped into the same recombination primitive cluster. Within each recombination primitive cluster, the morphological atoms are sorted from high to low according to their texture gradient magnitude. The top 60% to 80% of the morphological atoms are selected to form recombination pattern primitives, and the texture gradient distribution within the recombination pattern primitives is non-uniformly resampled. During the non-uniform resampling process, the density control factor is set to 1.2 to 1.8, ultimately generating a set of recombination pattern primitives with texture gradient distribution characteristics. This set includes 30 to 60 recombination pattern primitives.

[0020] In step S2, based on the recombined pattern primitive set, the AI ​​agent calls the multimodal adaptation generation model for clothing patterns. This model includes a text encoding channel and a visual encoding channel. The text encoding channel converts the pattern theme tags and style tendency tags in the design requirement instructions into 768-dimensional text description feature vectors. The visual encoding channel converts the geometric contour features and texture density distribution features of each primitive in the recombined pattern primitive set into 512-dimensional visual style feature vectors. The outputs of the two channels are subjected to tensor concatenation in a common embedding space. After tensor concatenation, a 1280-dimensional fusion feature tensor is obtained. A nonlinear activation function is applied to the fusion feature tensor. The nonlinear activation function uses the hyperbolic tangent function to compress the values ​​of each element in the fusion feature tensor to the range of negative 1 to positive 1, obtaining the fusion latent feature matrix. The dimension of the fusion latent feature matrix is ​​set to 32 x 64. Based on the fusion of latent feature matrices, the generative model further decodes and outputs a multimodal pattern layout candidate set. The decoding process employs an attention mechanism with eight attention heads. Each attention head independently calculates the association weights between different positions in the input sequence. After normalization, the association weights are used as weighted aggregation coefficients to perform weighted permutations and combinations on the recombined pattern primitive set. During the permutation and combination process, each candidate layout pattern uses 8 to 15 primitives, generating a total of 5 to 8 candidate layout patterns. Each candidate layout pattern is stored in the form of a pixel matrix with a resolution of 512 x 512. The candidate layout patterns differ in the spatial arrangement, rotation angle, and scaling ratio of the primitives. Among the aforementioned differential parameters, the rotation angle ranges from 0 degrees to 360 degrees, and the scaling ratio ranges from 0.5 to 2.0, ultimately forming a multimodal pattern layout candidate set.

[0021] In step S3, for the multimodal pattern layout candidate set, the AI ​​agent calls the clothing pattern color space correction algorithm. First, it extracts the principal components of the base chromaticity of the target clothing fabric under a standard light source. The principal components of the base chromaticity include the mean value of the L channel luminance, the mean value of the red-green axis chromaticity of the a channel, and the mean value of the yellow-blue axis chromaticity of the b channel. The value ranges of the three channel means are 0 to 100, -128 to +127, and -128 to +127, respectively. At the same time, it extracts the bidirectional reflectance distribution function feature parameters of the target clothing fabric. These feature parameters include the specular reflection component intensity value and the diffuse reflection component attenuation coefficient value. The specular reflection component intensity value is 0.1 to 0.4, and the diffuse reflection attenuation coefficient value is 0.6 to 0.9. Based on the principal components of the base chromaticity, a color gamut boundary constraint surface is constructed. The color gamut boundary constraint surface is represented by a convex hull in three-dimensional space. The L, a, and b channel chromaticity values ​​of each pixel in each pattern pattern in the candidate set are projected. To project the image into the convex hull, a nearest neighbor projection strategy is used. If a pixel's chroma value is outside the convex hull, it moves along the gradient descent direction to the nearest point on the convex hull surface. If the pixel's chroma value is inside the convex hull, it remains unchanged. After completing the color gamut projection, the color difference compensation vector under different viewing angles is calculated based on the bidirectional reflectance distribution function characteristic parameters. The viewing angles are 0 degrees, 30 degrees, and 60 degrees. The color difference compensation amount under each angle is calculated. A directional drift correction is applied to the chroma values ​​of each pixel after projection. The drift step size of the directional drift correction is set to 0.02 to 0.05, and the drift direction is determined by the direction of the color difference compensation vector. After completing the drift correction, local neighborhood smoothing interpolation is performed on the pattern color matrix. The local neighborhood window size is set to 3 x 3 pixels, and the smoothing interpolation weight uses a Gaussian kernel function with a spatial standard deviation of 0.8. Finally, the color-corrected pattern is output.

[0022] In step S4, the color-corrected pattern is input into the pattern design effect quantitative evaluation platform. The evaluation platform first performs gray-level co-occurrence matrix statistics on the color-corrected pattern. The pixel spacing of the gray-level co-occurrence matrix is ​​set to 1 pixel, and the directions are 0 degrees, 45 degrees, 90 degrees, and 135 degrees. The texture contrast value and energy feature value are calculated in each direction. The texture contrast value is calculated by weighting the squares of the gray-level differences between adjacent pixels in the gray-level co-occurrence matrix. The energy feature value is calculated by weighting the squares of the gray-level differences between adjacent pixels in the gray-level co-occurrence matrix. The average texture contrast values ​​in the four directions are used as the final texture contrast index, and the average energy characteristic values ​​in the four directions are used as the final energy characteristic index. Then, the pattern is converted from the L, a, b color space to a uniform color space. In the uniform color space, the variance and covariance values ​​of each color channel are calculated. The variance reflects the dispersion of chromaticity within each channel, and the covariance reflects the correlation of chromaticity between different channels. The variance and covariance values ​​of each channel are weighted and combined, with weighting coefficients set to 0.3, 0.3, and 0.4 respectively. The combined result is used as a color harmony index. Then, Fourier shape descriptor extraction is performed on the boundary contours of the pattern. During extraction, the first 10 Fourier coefficients are selected as shape features. Based on these shape features, rotational symmetry and mirror symmetry scores are calculated. The rotational symmetry score is calculated by rotating the pattern 360 degrees around its center and dividing it into 12 equal angular intervals. The percentage of overlapping area within each interval is counted. The ratio of the number of intervals with an overlapping area exceeding 0.8 to the total number of intervals is the rotational symmetry score. The mirror symmetry score... The calculation method involves mirroring the pattern along the horizontal and vertical axes, calculating the percentage of overlapping area of ​​the contours before and after the mirroring, and the average of the percentages in the two directions is the mirror symmetry score. Finally, the texture contrast index, energy characteristic index, color harmony index, rotational symmetry score, and mirror symmetry score are weighted and linearly combined. The weighting coefficients of the five indicators are set to 0.2, 0.15, 0.3, 0.2, and 0.15, respectively. After weighted summation, a quantitative score sequence for each pattern is generated, with the quantitative score ranging from 0 to 100.

[0023] In step S5, the AI ​​agent performs joint optimization iteration on the recombined pattern primitive set, multimodal fusion weights, and color correction intensity based on the quantized scoring sequence. The AI ​​agent first analyzes the pattern corresponding to each score in the quantized scoring sequence and extracts the multimodal fusion weight vector corresponding to the pattern with the highest score. The multimodal fusion weight vector includes text modality weight values ​​and visual modality weight values, with the sum of the two weight values ​​being 1. The text modality weight ranges from 0.2 to 0.8, and the visual modality weight ranges from 0.2 to 0. 8. Based on the multimodal fusion weight vector of the pattern with the highest score, a gradient-guided strategy is used to update the current multimodal fusion weights, with an update step size set to 0.05. Simultaneously, a coarse adjustment operation is performed on the current color correction intensity based on the color correction intensity parameters corresponding to the pattern with the highest score. The color correction intensity parameters include the gamut mapping intensity value and the drift compensation amplitude value. The gamut mapping intensity value ranges from 0.5 to 1.0, and the drift compensation amplitude value ranges from 0.1 to 0.5. The adjustment step size for the coarse adjustment operation is set to... 0.03, then the set of recombined pattern primitives is updated, and primitives with scores below the elimination threshold are removed. The elimination threshold is set to 0.6 times the median of the quantized score sequence. Primitives with scores below this threshold are marked as primitives to be eliminated. The number of eliminated primitives accounts for 25% to 40% of the total set. Affine mutation operations are performed on the morphological characteristics of the top 20% of primitives. Affine mutation operations include translation mutation, rotation mutation, and scaling mutation. The translation mutation range is from -10 pixels to +10 pixels. The rotation variation ranges from -15 degrees to +15 degrees, and the scaling variation ranges from 0.8 to 1.2. New primitives are added to the set through affine variation, restoring the total number of primitives in the set to the original number. The updated multimodal fusion weight vector, color correction intensity parameters, and recombined pattern primitive set are encapsulated as a new generation generation strategy parameter for the AI ​​agent to call and execute the next iteration. During the iteration process, the adjustment step size of the multimodal fusion weight decreases exponentially with the number of iterations, and the decay factor ranges from 0.92 to 0.98.

[0024] In step S6, S2 to S5 are repeated until the iteration termination condition is met. The iteration termination condition is that the sum of the absolute values ​​of the changes in two consecutive iterations of the quantized scoring sequence is lower than a preset threshold and the number of iterations reaches a lower limit. The preset threshold is between 0.015 and 0.025, the lower limit of the number of iterations is set to 10, and the upper limit of the number of iterations is set to 80 to 120. During each iteration, the AI ​​agent records the quantized scoring sequence of the current iteration round and the corresponding final pattern. Termination is not performed before the number of iterations reaches the lower limit. After the number of iterations reaches the lower limit, the absolute value of the difference between the current quantized scoring sequence and the quantized scoring sequence of the previous iteration is calculated after each iteration. The total change is obtained by adding the absolute values ​​of all differences. If the total change is lower than the preset threshold, the iteration termination condition is met, the iteration stops, and the pattern of the current round is output as the final garment pattern. If the total change is not lower than the preset threshold and the number of iterations has not reached the upper limit, the next round of iteration continues. If the number of iterations reaches the upper limit and the total change is still not lower than the preset threshold, the iteration is forcibly terminated, and the pattern with the highest quantitative score is selected from all recorded patterns as the final garment pattern. The final garment pattern is output in the form of a pixel matrix with a resolution of 1024 x 1024 and a color space of L, a, b. The corresponding quantitative score value is also output to record the design quality.

[0025] Preferably, the traditional pattern element deconstruction and recombination algorithm includes the following expression: ,in, This represents the set of recombined pattern primitives. Indicates the first A recombined pattern primitive, Indicates the domain of the source pattern. Represents the morphological atom decomposition operator. Indicates the spatial position of the source pattern The gray intensity function at that location, Indicated by Center position parameter and with The first of the scale parameters A Gaussian receiver domain function Domain Lebesgue measure differential on, This indicates the total number of recombinant primitives; subsequently, for each primitive... Apply deconstruction and reconstruction mapping: ,in, This represents the optimized result after deconstruction and recombination. A recombination element, Represents the gradient operator. Indicates the first Texture transformation functions, Indicates the first The curvature control parameters corresponding to the transformation, This represents the total number of transformation functions. express Norm, This indicates the product operation.

[0026] Specifically, in the traditional pattern element deconstruction and recombination algorithm, the construction logic of the first formula stems from the need for hierarchical analysis of the spatial structure of the source pattern. The morphological atom decomposition operator treats the gray-level intensity function of the source pattern as a continuous distribution field on the domain. This distribution field is then extracted through multi-scale filtering using a Gaussian receiver domain function. The center position parameter of the Gaussian receiver domain function determines the spatial positioning accuracy of the morphological atoms, while the scale parameter controls the spatial range of the receiver domain. The synergistic effect of these two parameters effectively captures abrupt texture changes at different spatial locations. Lebesgue metric differentiation ensures the mathematical rigor of integral operations on irregular domains. The total number of recombined primitives is dynamically determined based on the complexity of the source pattern, typically between 30 and 60. The second formula is built upon the optimization theory framework, aiming to reconstruct the texture structure of the primitives obtained from the initial decomposition. The secondary refinement involves gradient operators used to quantify the intensity of texture changes within primitives. Various texture transformation functions correspond to different curvature control parameters, ranging from 0.5 to 2.0, controlling different transformation behaviors such as edge sharpening, directional smoothing, and local contrast stretching. The product operation superimposes multiple transformation effects, with the p-value in the Lp norm set to 2, corresponding to the Euclidean distance metric, used to measure the difference between the transformed primitives and the target texture features. The derivation relationship between the two formulas is as follows: first, the spatial positioning and preliminary separation of the pattern primitives are completed using the first formula; then, the multi-target texture transformation optimization is applied to each separated primitive using the second formula, so that the recombined primitives retain the original structural information of the source pattern while obtaining richer texture expressiveness. The entire derivation process follows the logic of decomposition followed by optimization.

[0027] Preferably, the multimodal adaptation generation model for clothing patterns includes the following expression: ,in, This represents the fusion latent feature matrix in the common embedding space. This represents the hyperbolic tangent activation function. This represents the text modality projection weight matrix. This represents the text description feature vector. Represents the visual modality projection weight matrix. Represents visual style feature vectors. Represents the bias vector; based on the fused latent feature matrix Further, a multimodal pattern layout candidate set is generated: ,in, This represents the candidate set of multimodal pattern layouts. Indicates the first A matrix of candidate layout patterns, Indicates the generation of the first The number of fusion atoms used when selecting candidate patterns. Indicates the first The candidate pattern is the first one Attention weight coefficients of each fused atom This represents the natural exponential function. Indicates the first The candidate pattern is the first one The cumulative affinity matrix of the fused atoms, Indicating continuous variables The affinity density function at that point, Denotes the upper limit constant of integration. , .

[0028] Specifically, in the multimodal adaptation and generation model for clothing patterns, the first formula is based on the premise that text description and visual style are two different modalities of information carriers, requiring projection onto the same mathematical space to achieve semantic alignment. Both the text modality projection weight matrix and the visual modality projection weight matrix are obtained through pre-training, with dimensions of 768x512 and 512x512 respectively. The bias vector is set to 512, adjusting the baseline offset of the projected features. The hyperbolic tangent activation function constrains the fusion result within a finite interval to prevent numerical overflow. The fusion latent feature matrix has a dimension of 32x64, serving as a bridge connecting the two modalities. The second formula, built upon the aforementioned fusion features, uses an attention mechanism to strengthen the association between each primitive and the fusion feature. The metric is quantified into attention weight coefficients. The natural exponential function makes the association strength non-negative and amplifies the significant differences. The cumulative affinity matrix is ​​introduced to describe the long-term dependency between primitives and fused features in continuous space. The value of its integral upper limit constant L is determined according to the sequence length and is usually set to a fixed value between 64 and 128. The summation operation aggregates the intermediate results generated by different fused atoms. The number of fused atoms used for each candidate pattern is set to 8 to 15. The derivation relationship between the two formulas is that the first formula is used to establish the fused representation of multimodal features, and the second formula is used to perform attention-driven layout decoding on the representation, so that the output candidate layout pattern is semantically consistent with the text conditions and visually consistent with the source pattern features.

[0029] Preferably, the color space correction algorithm for clothing patterns includes the following expression: ,in, This represents the color matrix of the pattern output after color correction. , Indicates the total number of cascaded calibration stages. Indicates the first Stage color gamut mapping diagonal matrix, This indicates the input pattern color matrix. This represents the Hadamard product operation. Indicates the first The spectral reflectance compensation matrix is ​​constructed based on the base colorimetric parameters of the target apparel fabric. Indicates the first Stage-based non-rigid color drift compensation bias matrix.

[0030] Specifically, the color space correction algorithm for clothing patterns decomposes color correction into multiple cascaded stages of successive fine-tuning operations. The total number of cascaded correction stages, Q, is set to 3 to 5 stages. Each stage independently performs color gamut mapping and color drift compensation. The diagonal elements of the color gamut mapping diagonal matrix correspond to the mapping gain coefficients of the L, a, and b color channels, respectively. The value ranges of the three channel gain coefficients are 0.8 to 1.2, 0.7 to 1.3, and 0.7 to 1.3, respectively. Their function is to perform channel-by-channel scaling on the input pattern color matrix to achieve the purpose of compressing or stretching the color gamut range. The number of rows in the input pattern color matrix corresponds to the total number of pixels, and the number of columns is fixed at 3, storing the chromaticity values ​​of the L, a, and b channels, respectively. The spectral reflectance compensation matrix has the same dimension as the input matrix, and the values ​​of each element are calculated based on the base chromaticity parameters of the target clothing fabric. The base chromaticity parameters include the base chromaticity parameters of the L channel. The three reference values—the standard value, the a-channel reference value, and the b-channel reference value—are taken from statistical data of fabric colorimetric measurement experiments. The specific values ​​fluctuate within the standard colorimetric range depending on the type of fabric. The Hadamard product operation realizes element-wise multiplication between the input matrix and the compensation matrix, so that the color compensation effect is precisely applied to each channel of each pixel. The non-rigid color drift compensation bias matrix introduces an additional directional offset through addition operations. This offset is calculated based on the bidirectional reflection distribution characteristics of the fabric. The offset varies at different pixel positions to achieve spatially non-uniform color drift compensation. The product operation accumulates the correction effects of all cascaded stages in sequence, with the output of the previous stage serving as the input of the next stage. The mapping gain coefficient and the compensation bias matrix of each stage are dynamically adjusted according to iterative feedback. The adjustment step size of the gain coefficient is 0.02, and the adjustment step size of the bias matrix is ​​0.1 to 0.5.

[0031] Preferably, the AI ​​agent includes a graphics processor cluster, a tensor processing unit array, and a high-bandwidth storage unit. The source pattern originates from a pattern database in vector graphics format. The design requirement instructions include pattern theme tags, clothing category tags, and style preference tags. The multimodal fusion weights are dynamically adjusted by the AI ​​agent based on a confidence allocation mechanism. The adjustment step size decreases exponentially with the number of iterations, and the decay factor is between 0.92 and 0.98. The color correction intensity is controlled by a color gamut boundary constraint coefficient, which is 0.75 in the early stage of iteration and gradually converges to 0.25 in the later stage of iteration. The iteration termination condition is that the sum of the absolute values ​​of the changes in two adjacent iterations of the quantized scoring sequence is lower than a preset threshold, which is between 0.015 and 0.025. The upper limit of the number of iterations is set to 80 to 120. After each iteration, the recombined pattern primitive set retains the top 60% to 75% of the primitives in terms of score from high to low to participate in clothing pattern generation, and discards the primitives in the bottom 25% to 40%.

[0032] Preferred, such as Figure 2 As shown, S2 includes the following sub-steps: S21, extracting the geometric contours and texture density distributions of each primitive in the recombined pattern primitive set into a structural descriptor sequence; S22, performing cross-modal attention alignment between the structural descriptor sequence and the text description features to generate an attention weight distribution map; S23, performing weighted aggregation of visual style features based on the attention weight distribution map to obtain a stylized structure fusion tensor; S24, inputting the stylized structure fusion tensor into a multilayer perceptron decoder to output a multimodal pattern layout candidate set.

[0033] Specifically, in step S2, S21 extracts the geometric contours and texture density distributions of each primitive in the recombined pattern primitive set into a sequence of structural descriptors. When extracting the geometric contours, the Canny edge detection operator is used, with a low threshold of 30 and a high threshold of 90. When extracting the texture density distribution, a local binary mode algorithm is used, with a neighborhood radius of 2 pixels and 8 sampling points. Each primitive corresponds to one structural descriptor, and all descriptors are arranged sequentially according to the original arrangement of the primitives in the set to form a sequence of structural descriptors. The length of this sequence is a specific value between 30 and 60 of the total number of primitives, and each descriptor has a dimension of 128. In step S22, the sequence of structural descriptors is aligned with the text description features through cross-modal attention. The text description features are output as a 768-dimensional vector from the text encoding channel. During the alignment process, a query matrix, a key matrix, and a value matrix are introduced, with dimensions set to 128 x 64 and 768, respectively. Multiply by 64 and 128 by 64, calculate the product of the query matrix and the transpose of the key matrix, divide by the scaling factor 8, and obtain the attention weight map after normalization. The dimension of the attention weight map is the sequence length multiplied by the sequence length. S23 performs weighted aggregation of visual style features based on the attention weight map. The visual style features are output as a 512-dimensional vector from the visual encoding channel. Multiply the attention weight map with the visual style feature matrix to obtain the weighted aggregated stylized structure fusion tensor. The dimension of this tensor is set to the sequence length multiplied by 64. S24 inputs the stylized structure fusion tensor into the multilayer perceptron decoder. The multilayer perceptron decoder includes 3 hidden layers with 256, 128 and 64 neurons in each hidden layer, respectively. The activation function is a modified linear unit. The output layer has 3 neurons, corresponding to the horizontal coordinate offset, vertical coordinate offset and scaling factor of the candidate layout pattern. Finally, a multimodal pattern layout candidate set is output, which includes 5 to 8 layout patterns.

[0034] Preferred, such as Figure 3 As shown, step S3 includes the following sub-steps: S31, extracting the principal components of the base chromaticity and the characteristic parameters of the bidirectional reflectance distribution function of the target clothing fabric under a standard light source; S32, constructing a color gamut boundary constraint surface based on the principal components of the base chromaticity, and projecting the chromaticity values ​​of each pixel in the pattern candidate set into the constraint surface; S33, calculating the color difference compensation vector field under different viewing angles based on the characteristic parameters of the bidirectional reflectance distribution function, and applying directional drift correction to the projected chromaticity values; S34, performing local neighborhood smoothing interpolation on the drift-corrected pattern color matrix, and outputting the color-corrected pattern.

[0035] Specifically, in step S3, S31 extracts the principal components of the base chromaticity and the characteristic parameters of the bidirectional reflectance distribution function of the target garment fabric under a standard light source. The standard light source is a D65 light source. The principal components of the base chromaticity are obtained by performing principal component analysis on the chromaticity measurements of multiple points on the fabric surface. The eigenvectors corresponding to the first three principal components are taken as the principal component basis. The characteristic parameters of the bidirectional reflectance distribution function include the specular reflection peak angle and the diffuse reflection attenuation index. The specular reflection peak angle ranges from 15 degrees to 45 degrees, and the diffuse reflection attenuation index ranges from 0.5 to 1.5. In step S32, a color gamut boundary constraint surface is constructed based on the principal components of the base chromaticity. The color gamut boundary constraint surface is constructed using a triangular mesh partitioning method. The vertices of the triangular mesh are determined by the extreme value combination of each dimension of the principal components of the base chromaticity, with a total number of vertices of 8 to 12. The chromaticity values ​​of each pixel in the pattern candidate set are projected into the constraint surface. Iteration is used during the projection process. The nearest point algorithm has an upper limit of 50 iterations and a convergence threshold of 0.001. S33 calculates the color difference compensation vector field under different viewing angles based on the characteristic parameters of the bidirectional reflectance distribution function. The viewing angles are 0 degrees, 30 degrees, 45 degrees, and 60 degrees. After calculating the color difference compensation vector under each angle, a continuous vector field is generated by radial basis function interpolation. The radial basis function adopts a Gaussian kernel function with a kernel width parameter set to 1.5. A directional drift correction is applied to the projected chromaticity values ​​with a drift correction step size factor set to 0.03. S34 performs local neighborhood smoothing interpolation on the drift-corrected pattern color matrix. The local neighborhood window shape is a cross-shaped window with a window radius of 2 pixels. The interpolation weight is calculated according to the reciprocal of the pixel spatial distance. Pixels with a distance less than or equal to the radius participate in the interpolation, while pixels with a distance greater than the radius do not participate in the interpolation. The color-corrected pattern is then output.

[0036] Preferred, such as Figure 4 As shown, step S4 includes the following sub-steps: S41, performing grayscale co-occurrence matrix statistics on the color-corrected pattern to extract texture contrast and energy feature values; S42, converting the pattern to a uniform color space and calculating the variance and covariance of each color channel to quantify the harmony index; S43, performing Fourier shape descriptor extraction on the boundary contour of the pattern and calculating rotational symmetry score and mirror symmetry score; S44, performing a weighted linear combination of texture contrast, energy feature values, color harmony index, rotational symmetry score, and mirror symmetry score to generate a quantitative score sequence for each pattern.

[0037] Specifically, in step S4, S41 performs gray-level co-occurrence matrix (GLCM) statistics on the color-corrected pattern, extracting texture contrast and energy feature values. The pixel spacing of the GLCM is set to 1 pixel, and the directions are 0 degrees, 45 degrees, 90 degrees, and 135 degrees. The GLCM values ​​for each direction are statistically analyzed separately, and the average value is taken as the final GLCM. The texture contrast value is obtained by calculating the weighted sum of squares of the off-diagonal elements in the GLCM, with the weights being the squares of the absolute values ​​of the pixel gray-level differences. The energy characteristic values ​​are obtained by calculating the sum of squares of each element in the gray-level co-occurrence matrix. S42 converts the pattern to a uniform color space, using CIEL*a*b* space. During the conversion, the reference white point is D65. The variance and covariance of each color channel are calculated. The variance values ​​of channel L, channel a, and channel b reflect the degree of chromatic dispersion of each channel, while the covariance values ​​of L and a, L and b, and a and b reflect the degree of correlation between channels. The variance values ​​are then compared with the covariance values... After combining the variance values, the harmony index is calculated using the chromaticity harmony function. The variance weighting coefficients for each channel in the harmony function are set to 0.25, 0.35, and 0.4, respectively, and the covariance weighting coefficients are all set to 0.1. S43 performs Fourier shape descriptor extraction on the boundary contour of the pattern. The number of boundary contour sampling points is set to 256, and the Fourier transform order is set to 15. The first 10 low-frequency coefficients are used as the shape descriptor. The rotational symmetry score is calculated by measuring the contour of the pattern as it rotates 360 degrees around the center. The maximum percentage of overlapping area is obtained. The mirror symmetry score is obtained by calculating the average percentage of overlapping area of ​​the outline after the pattern is mirrored along the horizontal and vertical axes. S44 performs a weighted linear combination of texture contrast, energy feature value, color harmony index, rotational symmetry score and mirror symmetry score. The five weighting coefficients are set to 0.2, 0.15, 0.3, 0.2 and 0.15 respectively. After weighted summation, a quantitative score sequence of each pattern is generated. The score value is accurate to two decimal places.

[0038] Preferred, such as Figure 5 As shown, S5 includes the following sub-steps: S51, parsing the pattern corresponding to each score in the quantitative scoring sequence and extracting the multimodal fusion weight vector corresponding to the pattern with the highest score; S52, performing gradient-guided coarse adjustment on the current color correction intensity based on the color correction intensity parameter of the pattern with the highest score; S53, removing primitives with scores below the elimination threshold from the recombined pattern primitive set and supplementing new primitives by affine mutation based on the morphological characteristics of the primitives with the highest scores; S54, encapsulating the updated fusion weight vector, color correction intensity parameter, and recombined pattern primitive set into a new generation generation strategy parameter for the AI ​​agent to call and execute the next iteration.

[0039] Specifically, in step S5, S51 analyzes and quantifies the pattern corresponding to each score in the scoring sequence, extracts the multimodal fusion weight vector corresponding to the pattern with the highest score, which includes two components: text modality weight and visual modality weight. The sum of the two components is always equal to 1. The text modality weight ranges from 0.2 to 0.8, and the visual modality weight ranges from 0.2 to 0.8. At the same time, the color correction intensity parameter corresponding to the pattern with the highest score is extracted. This parameter includes two components: color gamut mapping intensity and drift compensation amplitude. The color gamut mapping intensity ranges from 0.5 to 1.0, and the drift compensation amplitude ranges from 0.1 to 0.5. S52 performs a gradient-guided coarse adjustment operation on the current color correction intensity based on the color correction intensity parameter of the pattern with the highest score. The gradient direction of the coarse adjustment operation is the direction of the difference between the intensity parameter of the pattern with the highest score and the current intensity parameter. The coarse adjustment step size is set to 0.3 times the difference, and the maximum value of the single coarse adjustment step size is limited to 0.05. S53 reconstructs the pattern primitive set. Primitives with scores below the elimination threshold are removed. The elimination threshold is set to 0.6 times the median of the current quantized score sequence. The number of primitives removed is determined based on the actual number of primitives in the set that are below the threshold, and this number accounts for 25% to 40% of the total number of primitives in the set. New primitives are supplemented by affine mutation based on the morphological characteristics of the top 20% of primitives in the score. Affine mutation includes translation parameter mutation, rotation parameter mutation, and scaling parameter mutation. The translation parameter mutation range is from -8 pixels to +8 pixels, the rotation parameter mutation range is from -12 degrees to +12 degrees, and the scaling parameter mutation range is from 0.85 to 1.15. The number of new primitives after mutation is equal to the number of primitives removed. S54 encapsulates the updated fusion weight vector, color correction intensity parameters, and recombined pattern primitive set into a new generation generation strategy parameter. The encapsulation operation combines the three into a single parameter object according to a predetermined data structure. This parameter object includes a weight field, an intensity field, and a primitive set field, which are called by the AI ​​agent to execute the next round of iteration.

[0040] like Figure 6As shown, an AI-based clothing pattern design system is applied to an AI-based clothing pattern design method. The system includes: a pattern primitive deconstruction and reconstruction engine, which receives source pattern data, internally performs a morphological atom decomposition protocol and a Gaussian receptive domain convolution extraction process, and outputs a reconstructed pattern primitive set; a multimodal fusion layout generator, whose input is coupled to the output of the pattern primitive deconstruction and reconstruction engine, receives the reconstructed pattern primitive set, and internally includes a text encoding channel and a visual encoding channel. The outputs of the two channels perform tensor splicing and nonlinear activation operations in a common embedding space, outputting a multimodal pattern layout candidate set; and a color gamut drift correction processor, whose input is coupled to the output of the multimodal fusion layout generator, internally embedding a base chromaticity constraint module and a directional color difference compensation module, performing color gamut correction on the multimodal pattern layout candidate set. The system outputs a color-corrected pattern after mapping and drift correction. A pattern quality evaluator, with its input coupled to the output of the color gamut drift correction processor, includes a texture statistics calculation unit, a color distribution analysis unit, and a shape symmetry measurement unit. The weighted summation of the outputs from each unit generates a quantitative scoring sequence. An agent-based strategy iteration controller, with its input coupled to the output of the pattern quality evaluator and its output coupled to the control ports of the pattern primitive deconstruction and reconstruction engine, the multimodal fusion layout generator, and the color gamut drift correction processor, dynamically adjusts the internal operating parameters of each engine, generator, and processor based on the quantitative scoring sequence. A pattern output cache interface, with its input coupled to the termination determination port of the agent-based strategy iteration controller, receives and stores the final garment pattern when the iteration termination condition is met.

[0041] The AI-based method and system for designing clothing patterns organically integrates pattern deconstruction and reconstruction, multimodal fusion generation, color space correction, and a quantitative evaluation platform through an AI agent. This transforms the process from isolated, sequential steps into a closed-loop iterative architecture based on quantitative scoring feedback. The agent dynamically adjusts the composition of the reconstructed primitive set, the allocation of multimodal fusion weights, and the parameter values ​​for color correction intensity based on the scoring sequence generated by the evaluation platform. This allows the structural features of the deconstruction stage to guide the layout decisions in the generation stage, while the color gamut constraints of the correction stage can in turn adjust the style fusion tendency in the generation stage. This effectively eliminates the shortcomings of existing technologies where errors accumulate at each stage and cannot be compensated, thus improving the overall coordination and design quality of the final pattern.

[0042] This invention addresses the shortcomings of existing color correction methods that only perform global coarse-grained stretching while ignoring the optical properties of the fabric substrate. It introduces a directional color difference compensation mechanism targeting the chromaticity parameters and bidirectional reflectance distribution characteristics of the target garment fabric substrate during the color gamut mapping process. This mechanism applies non-rigid drift correction and local neighborhood smoothing interpolation to the chromaticity values ​​of each pixel in the pattern, ensuring that the corrected pattern's color rendering on the real fabric highly matches the virtual design, significantly suppressing color deviation between screen preview and physical implementation. Simultaneously, a multi-dimensional quantitative evaluation platform provides precise iterative guidance signals to the intelligent agent from the perspectives of texture contrast, color harmony, and shape symmetry. This enables the system to quickly converge to the optimal design scheme that combines aesthetic quality and engineering adaptability within a limited number of iterations, overcoming the technical bottleneck of existing solutions lacking a collaborative feedback mechanism.

[0043] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0044] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for designing clothing patterns based on AI intelligent agents, characterized in that, Includes the following steps: S1, responding to design requirements, invokes a pattern element deconstruction and recombination algorithm via an AI agent to perform morphological atom decomposition on the source pattern, generating a set of recombined pattern primitives with texture gradient distribution characteristics; S2, based on the recombined pattern primitive set, calls the multimodal adaptation generation model of clothing patterns, maps text description features and visual style features to the common embedding space, and outputs a multimodal pattern layout candidate set; S3, for the candidate set of multimodal pattern layout, calls the color space correction algorithm for clothing patterns. Based on the base chromaticity parameters and light reflection characteristics of the target clothing fabric, it performs color gamut mapping and non-rigid color drift compensation on each pattern in the candidate set, and outputs the color-corrected pattern. S4. Input the color-corrected pattern into the pattern design effect quantitative evaluation platform. By calculating the pattern complexity, color contrast and structural symmetry indicators, a quantitative scoring sequence for each pattern is generated. S5, the AI ​​agent performs joint optimization and iteration on the recombined pattern primitive set, multimodal fusion weight and color correction intensity based on the quantitative scoring sequence, and updates the pattern generation strategy parameters; S6, repeat S2 to S5 until the iteration termination condition is met, and output the final clothing pattern.

2. The clothing pattern design method based on AI intelligent agent according to claim 1, characterized in that, The traditional pattern element deconstruction and recombination algorithm includes the following expression: ； in, This represents the set of recombined pattern primitives. Indicates the first A recombined pattern primitive, Indicates the domain of the source pattern. Represents the morphological atom decomposition operator. Indicates the spatial location of the source pattern The gray intensity function at that location, Indicated by Center position parameter and with The first of the scale parameters A Gaussian receiver domain function Domain Lebesgue measure differential on, This indicates the total number of recombined primitives; subsequently, for each primitive... Apply deconstruction and reconstruction mapping: ； in, This represents the optimized result after deconstruction and recombination. A recombination element, Represents the gradient operator. Indicates the first Texture transformation functions, Indicates the first The curvature control parameters corresponding to the transformation, This represents the total number of transformation functions. express Norm, This indicates the product operation.

3. The clothing pattern design method based on AI intelligent agent according to claim 1, characterized in that, The multimodal adaptation and generation model for clothing patterns includes the following expression: ； in, This represents the fusion latent feature matrix in the common embedding space. This represents the hyperbolic tangent activation function. This represents the text modality projection weight matrix. This represents the text description feature vector. Represents the visual modality projection weight matrix. Represents visual style feature vectors. Represents the bias vector; based on the fused latent feature matrix Further, a multimodal pattern layout candidate set is generated: ； in, This represents the candidate set of multimodal pattern layouts. Indicates the first A matrix of candidate layout patterns, Indicates the generation of the first The number of fusion atoms used when selecting candidate patterns. Indicates the first The candidate pattern is the first one Attention weight coefficients of each fused atom This represents the natural exponential function. Indicates the first The candidate pattern is the first one The cumulative affinity matrix of the fused atoms, Indicating continuous variables The affinity density function at that point, Denotes the upper limit constant of integration. , .

4. The clothing pattern design method based on AI intelligent agent according to claim 1, characterized in that, The color space correction algorithm for clothing patterns includes the following expression: ； in, This represents the color matrix of the pattern output after color correction. , Indicates the total number of cascaded calibration stages. Indicates the first Stage color gamut mapping diagonal matrix, This indicates the input pattern color matrix. This represents the Hadamard product operation. Indicates the first The spectral reflectance compensation matrix is ​​constructed based on the base colorimetric parameters of the target apparel fabric. Indicates the first Stage-based non-rigid color drift compensation bias matrix.

5. The clothing pattern design method based on AI intelligent agent according to claim 1, characterized in that, The AI ​​agent includes a graphics processor cluster, a tensor processing unit array, and a high-bandwidth storage unit. The source patterns are derived from a pattern database in vector graphics format. The design requirement instructions include pattern theme tags, clothing category tags, and style preference tags. The multimodal fusion weights are dynamically adjusted by the AI ​​agent based on a confidence allocation mechanism. The adjustment step size decreases exponentially with the number of iterations, and the decay factor ranges from 0.92 to 0.

98. The color correction intensity is controlled by a color gamut boundary constraint coefficient, which is 0.75 in the early stage of iteration and gradually converges to 0.25 in the later stage of iteration. The iteration termination condition is that the sum of the absolute values ​​of the changes in two adjacent iterations of the quantized scoring sequence is lower than a preset threshold, which ranges from 0.015 to 0.

025. The upper limit of the number of iterations is set to 80 to 120. After each iteration, the recombined pattern primitive set retains the top 60% to 75% of the primitives in terms of score from high to low to participate in clothing pattern generation, and discards the primitives in the bottom 25% to 40%.

6. The clothing pattern design method based on AI intelligent agent according to claim 1, characterized in that, S2 includes: The geometric contours and texture density distributions of each primitive in the recombined pattern primitive set are extracted as structural descriptor sequences; the structural descriptor sequences are aligned with textual descriptive features across modal attention to generate an attention weight distribution map; the visual style features are weighted and aggregated based on the attention weight distribution map to obtain a stylized structural fusion tensor. The stylized structure is fused with a tensor and input into a multilayer perceptron decoder to output a multimodal pattern layout candidate set.

7. The clothing pattern design method based on AI intelligent agent according to claim 1, characterized in that, S3 includes: Extract the principal components of the base colorimetry and the characteristic parameters of the bidirectional reflectance distribution function of the target clothing fabric under a standard light source; construct a color gamut boundary constraint surface based on the principal components of the base colorimetry, and project the colorimetry values ​​of each pixel in the pattern candidate set into the interior of the constraint surface; The color difference compensation vector field under different viewing angles is calculated based on the characteristic parameters of the bidirectional reflectance distribution function. The directional drift correction is applied to the projected chromaticity values. Local neighborhood smoothing interpolation is performed on the drift-corrected pattern color matrix to output the color-corrected pattern.

8. The method for designing clothing patterns based on AI intelligent agents according to claim 1, characterized in that, S4 includes: Gray-level co-occurrence matrix statistics are performed on the color-corrected patterns to extract texture contrast and energy feature values; The pattern is converted to a uniform color space, and the variance and covariance of each color channel are calculated to quantify the harmony index. Fourier shape descriptor extraction is performed on the boundary contour of the pattern, and rotational symmetry score and mirror symmetry score are calculated; The texture contrast, energy feature value, color harmony index, rotational symmetry score and mirror symmetry score are weighted and linearly combined to generate a quantitative score sequence for each pattern.

9. The method for designing clothing patterns based on AI intelligent agents according to claim 1, characterized in that, S5 includes the following steps: The pattern corresponding to each score in the quantitative scoring sequence is analyzed, and the multimodal fusion weight vector corresponding to the pattern with the highest score is extracted. Based on the color correction intensity parameter of the pattern with the highest score, a gradient-guided coarse adjustment operation is performed on the current color correction intensity; primitives with scores below the elimination threshold in the recombined pattern primitive set are removed, and new primitives are supplemented by affine mutation based on the morphological characteristics of the primitives with the highest scores. The updated fusion weight vector, color correction intensity parameters, and recombined pattern primitive set are encapsulated into a new generation generation strategy parameter, which is then called by the AI ​​agent to execute the next iteration.

10. A clothing pattern design system based on AI intelligent agents, characterized in that, The system is applied to the clothing pattern design method based on AI intelligent agents as described in claim 1, comprising: The pattern primitive deconstruction and reconstruction engine receives source pattern data, internally runs a morphological atom decomposition protocol and a Gaussian receptive domain convolution extraction process, and outputs a reconstructed pattern primitive set. The multimodal fusion layout generator, with its input coupled to the output of the pattern primitive deconstruction and reconstruction engine, receives the reconstructed pattern primitive set and includes a text encoding channel and a visual encoding channel. The outputs of the two channels perform tensor splicing and nonlinear activation operations in a common embedding space, and output a multimodal pattern layout candidate set. The color gamut drift correction processor, with its input coupled to the output of the multimodal fusion layout generator, internally embeds a base color constraint module and a directional color difference compensation module. It performs color gamut mapping and drift correction on the multimodal pattern layout candidate set and outputs a color-corrected pattern pattern. The pattern quality evaluator, with its input coupled to the output of the color gamut drift correction processor, includes a texture statistical calculation unit, a color distribution analysis unit, and a shape symmetry measurement unit. The weighted summation of the outputs from each unit generates a quantitative scoring sequence. The agent strategy iteration controller, with its input coupled to the output of the pattern quality evaluator and its output coupled to the control ports of the pattern primitive deconstruction and reconstruction engine, the multimodal fusion layout generator, and the color gamut drift correction processor, dynamically adjusts the internal operating parameters of each engine, generator, and processor based on the quantitative scoring sequence. The pattern output buffer interface, with its input coupled to the termination determination port of the agent strategy iteration controller, is used to receive and store the final garment pattern when the iteration termination condition is met.

Images