Intimate apparel fitting system based on anonymized parameter mapping and physical deformation feedback

Abstract

This invention provides a body-fitting clothing try-on system based on anonymized parameter mapping and physical deformation feedback, comprising: an anonymized parameter mapping and retrieval module for receiving discrete physiological parameters, retrieving a baseline 3D avatar mesh from an anonymized template library, and generating a 3D avatar mesh; a rule-based geometric feature interference module for triggering geometric correction on the 3D avatar mesh according to clothing style features; generating artificial scratches by performing negative numerical offset on contact path pixels on the depth map; an image synthesis module for denoising using a diffusion model in conjunction with a control network; and generating contact shadows by inducing changes in depth gradient; and a visual enhancement module for controlling skin texture adjustment, achieving high-fidelity semi-transparent interference of materials using bipartite masking and soft mask gradient fusion technology. This system retrieves anonymized templates from physiological parameters and constructs a high-fidelity digital avatar to simulate the physical pressure and shaping effect of body-fitting clothing.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and more specifically, to a body-fitting clothing try-on system based on anonymized parameter mapping and physical deformation feedback. Background Technology

[0002] With the rapid development of e-commerce and digital identity technology, the online interactive experience for apparel products has become crucial for improving conversion rates. While existing virtual try-on and AI-powered clothing technologies can initially simulate the coverage of clothing on the human body, significant gaps and limitations remain in highly segmented and privacy-sensitive areas such as intimate apparel, including regular underwear, sexy lingerie, and swimwear. These limitations are specifically manifested in the following aspects:

[0003] (i) The conflict between privacy and user resistance: Existing AI-based (AIGC) dress-up technologies typically require users to upload real photos of themselves or facial images to extract biometric features for image synthesis. However, intimate apparel involves highly sensitive personal privacy, and users have strong psychological resistance and security concerns about uploading such real images. This makes it difficult for existing "real-person image-driven" technologies to achieve large-scale commercial application in this category.

[0004] (II) Distortion of General Modeling and Realistic Physiological Details: The visual effect of close-fitting clothing relies heavily on the physical interaction between the fabric and the soft tissue of the human body (such as compression, wrapping, and support). Existing general 3D human body templates are mostly idealized standard body shapes based on statistical mean, failing to reproduce the local anatomical features of real users (such as the distribution of accessory breast tissue, physiological abdominal protrusions, and waist fat). Furthermore, existing models lack modeling of the mechanical feedback of skin under pressure, failing to generate realistic "marks" from bra straps and underwires on the skin. This results in generated fitting images exhibiting a noticeable "clothing floating" effect, lacking realism and making it difficult to assist users in making accurate purchasing decisions.

[0005] (III) Data Collection Threshold in Complex Environments: High-quality AI dress-up usually requires full-body photos taken from specific angles, under specific lighting conditions and backgrounds. In scenarios involving fitting in intimate clothing, guiding users to take and upload photos of their private parts that meet the algorithm's requirements is not only extremely complex and has a low conversion rate, but also crosses the line of data compliance and privacy protection.

[0006] (iv) Lack of targeted industry standard mapping: The intimate apparel category has a strict sizing system (such as cup size and circumference parameters). Existing general dressing techniques often ignore the strong coupling relationship between human body parameters (such as the difference between upper and lower chest circumference) and the geometric deformation of clothing, resulting in a disconnect between the "fitting effect" and the "actual size chart".

[0007] In summary, existing virtual try-on technologies, while ensuring user privacy, struggle to balance the realism of the generated clothing with ease of use. Therefore, there is an urgent need to propose a method for trying on close-fitting clothing based on anonymized parameter mapping and physical deformation feedback. Summary of the Invention

[0008] To overcome the shortcomings and deficiencies of existing technologies, the present invention aims to provide a body-fitting clothing try-on system based on anonymized parameter mapping and physical deformation feedback. This system, without requiring users to upload real photos, retrieves anonymized template libraries through physiological parameters and constructs a high-fidelity digital avatar to simulate the physical pressure and shaping effect of body-fitting clothing on the human body, while solving the problems of privacy leakage risks and data collection thresholds.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a body-fitting clothing try-on system based on anonymized parameter mapping and physical deformation feedback, comprising:

[0010] The anonymized parameter mapping and retrieval module receives discrete physiological parameters input by the user during the trial fitting and retrieves N baseline 3D avatar meshes from the anonymized template library. A target anonymized 3D avatar mesh M is generated using a linear weighted fusion method; local anatomical features are compensated to obtain the 3D avatar mesh. The 3D avatar mesh M' is divided into several semantic regions and bound to a set of physical response parameters.

[0011] The rule-based geometric feature interference module is used to trigger preset geometric corrections in the coordinate region corresponding to the 3D avatar mesh M' based on the style features of the target clothing; identify the contact path between the target clothing and the skin, and perform negative numerical offset on the contact path pixels on the rendered depth map to generate artificial scratches that simulate physical compression effects; and use human semantic segmentation map to decouple the deformed 3D avatar mesh.

[0012] The image synthesis module uses a diffusion model in conjunction with a control network to inject a depth map containing artificial scratches as a core condition into the denoising process; by utilizing depth gradient changes, it induces the generation of contact shadows with spatial depth, thus eliminating the floating effect of the straps from a visual perspective.

[0013] It also includes a visual enhancement module, which controls skin texture adjustment by adjusting the noise weights and perturbation values ​​in the latent space, avoiding the model-like appearance caused by excessive smoothing; it employs bipartite mask and soft mask gradient fusion technology to ensure that transparent or semi-transparent materials achieve high-fidelity semi-transparent interference in the latent space.

[0014] Preferably, the anonymization parameter mapping and retrieval module includes the following units:

[0015] A1. Human body parameter encoding unit:

[0016] The system receives discrete physiological parameters input from users trying on clothes, constructs a human body parameter vector P, and performs max-min normalization based on a large sample statistical range to obtain the final human body parameter vector. ;

[0017] A2. Human Latent Space Mapping Unit:

[0018] A pre-trained variational autoencoder is used as the human parameter encoder network E to process the human parameter vector. This is mapped to the latent vector Z of the human body shape in the latent shape space S;

[0019] A3. Multi-template fusion reconstruction unit:

[0020] In the anonymized template library, the KNN algorithm is used to calculate the Euclidean distance, and the N reference 3D avatar meshes that are closest to the human body shape latent vector Z are retrieved in the shape latent space S. The anonymized template library is built based on massive 3D human body scanning data. It extracts human geometric topological features from the 3D human body scanning data, decouples the human geometric topological features to form a low-dimensional feature shape latent space S, and stores it.

[0021] Calculate N baseline 3D avatar meshes based on the reciprocal of the similarity distance. Fusion weights The target anonymous 3D avatar mesh M is generated through linear weighted fusion; the generated 3D avatar mesh M completely strips away facial textures and identifiable individual biometric information;

[0022] A4. Soft tissue compensation and reconstruction unit:

[0023] Based on the input body fat percentage f and chest-to-waist ratio and local circumference gradient changes Constructing a nonlinear soft tissue compensation function Calculate the displacement compensation amount for each vertex of the 3D avatar network. ;

[0024] Using displacement compensation Compensation was performed on local anatomical features such as accessory breast expansion, abdominal fat accumulation, and sagging to obtain a compensated three-dimensional avatar mesh. ;

[0025] A5. Body Continuity Constraint Unit:

[0026] Establish the energy function for topological continuity constraints of the 3D avatar mesh. By minimizing Limit the rate of change of normal between adjacent vertices to ensure the smooth continuity and natural transition of the skin surface;

[0027] A6. Human semantic region binding unit:

[0028] Using a pre-defined anatomical mapping matrix, perform 3D semantic region segmentation on the 3D avatar mesh M'; for each semantic region Independently bind physical response parameter groups ,in This represents the regional compression coefficient. For soft tissue elasticity parameters, This represents the material contact weight.

[0029] Preferably, in the human body parameter encoding unit, the human body parameter vector P is constructed as follows:

[0030]

[0031] Where h represents height and w represents weight; Indicates upper chest circumference; The symbol indicates underbust circumference; wa indicates waist circumference; hip indicates hip circumference; and f indicates body fat percentage.

[0032] Human body parameter vector for:

[0033]

[0034] in, , These are the minimum and maximum values ​​of the components of the human body parameter vector P within the large sample statistical range, respectively.

[0035] In the human body latent space mapping unit, the latent vector Z of the human body shape is:

[0036]

[0037] In the multi-template fusion reconstruction unit, the fusion weights for:

[0038]

[0039] Among them, Z i Z j These are the latent vectors of the human body shape for the i-th and j-th reference 3D avatar meshes, respectively;

[0040] The target anonymous 3D avatar mesh M is generated using a linear weighted fusion method as follows:

[0041]

[0042] In the soft tissue compensation and reconstruction unit, the displacement compensation amount of each vertex of the three-dimensional avatar network for:

[0043]

[0044] Compensated 3D Avatar Mesh for:

[0045]

[0046] In the body continuity constraint unit, the energy function of the three-dimensional avatar mesh topological continuity constraint :

[0047]

[0048] in, Represents adjacent vertex pairs in a 3D avatar mesh. As vertex The unit normal vector;

[0049] In the human semantic region binding unit, the three-dimensional semantic region segmentation formula is:

[0050]

[0051] in, The total semantic region of the three-dimensional avatar mesh M'; , , , , These are the semantic regions for the chest, shoulder straps, waistband, abdomen, and buttocks, respectively.

[0052] Preferably, a parameter reliability verification unit is further included between the human body parameter encoding unit and the human body latent space mapping unit.

[0053] Calculate the confidence probability density function of the human body parameter vector ;when If the input data is found to contain counterintuitive errors, the output parameter correction vector is used to attenuate the human body parameter vector P or trigger a pop-up correction.

[0054] Preferably, the rule-based geometric feature interference module includes the following units:

[0055] B1, Local geometric correction element:

[0056] The structural topological feature lines of the target garment are pre-established; using a ray casting algorithm, the structural topological feature lines of the target garment are spatially projected along the surface tangent direction of the 3D avatar mesh M', thereby locking the set of compressed vertices on the surface of the 3D avatar mesh. ;

[0057] Based on the tightness coefficient and elastic modulus of the target garment, the set of compression vertices is determined. Apply a negative displacement along the surface normal of the 3D avatar mesh to perform nonlinear anisotropy correction;

[0058] B2, Depth Map Interference Unit:

[0059] The 3D avatar mesh M' after local geometric correction is rendered into a 2D depth map using a preset camera view matrix and orthogonal projection matrix;

[0060] By using coordinate mapping, the corresponding pixel coordinate paths of the structural topological feature lines of the target garment on the two-dimensional depth map are extracted. This serves as the contact path between the target garment and the skin; a Gaussian first-order partial derivative operator is used as the sliding window kernel to analyze the pixel coordinate path. The pixel values ​​within a preset range on both sides of the center line are non-linearly offset in a negative direction; the offset depth section mathematically presents a "U" or "V" shaped spatial structure with smooth ends and a concave middle, thereby generating artificial grooves on the depth map that simulate physical compression effects.

[0061] B3. Semantic Feature Alignment Unit:

[0062] A dense pose estimation network is used to project the 3D avatar mesh obtained after processing the depth map interference unit into a 2D model, generating a human semantic segmentation map that is absolutely aligned with it at the pixel level.

[0063] The 3D avatar mesh region is decoupled into independent channels for each part of the body using semantic tags to create a human semantic segmentation mask.

[0064] Preferably, in the local geometric correction unit, the correction formula for performing nonlinear anisotropy correction is:

[0065]

[0066] in, The original 3D coordinate vectors of the 3D avatar mesh; Let be the unit normal vector of the currently compressed vertex. The tightness factor of the target garment. The soft tissue elasticity parameters in the physical response parameter group bound to the semantic region.

[0067] Preferably, the image synthesis module includes the following units:

[0068] C1. Multidimensional structural constraint element:

[0069] A multi-condition input matrix driven by a control network is constructed. A depth map containing artificial scratches and a human semantic segmentation mask are taken as input. The control network extracts feature maps of the depth map tensor and the human semantic segmentation mask at different network resolution scales to form multi-dimensional structural constraints. Simultaneously, a two-dimensional tiled image of the target garment is used to extract high-dimensional material and texture feature vectors through an independent garment visual encoder, serving as the garment's visual features. The garment visual features and standard Gaussian random noise are interactively fused through a cross-attention mechanism, and then explicitly weighted and fused with the control network output via a skip connection. After fusion, they are fed into a diffusion model for iterative reverse denoising, and finally output through image decoding.

[0070] C2, Physical Shadow Induction Unit:

[0071] Based on latent space trajectory correction without classifier guidance, the diffusion model automatically generates contact shadows and edge dark areas that conform to the logic of real optical scattering at the artificial marks on the straps, steel rings, and skin.

[0072] Preferably, the visual enhancement module includes the following units:

[0073] D1, Skin Texture Fine-Tuning Unit:

[0074] Based on the human semantic segmentation map, in the exposed skin area not covered by clothing, the low-dimensional feature tensor space generated by the diffusion model for image denoising is adjusted by adjusting the variance of random variables and noise weights, serving as the latent space. This induces the high-frequency generation of minute skin textures, pore textures, and minute non-identity biometric features. At the edges of artificial scratches and in the area of ​​skin under pressure, the mean shift of the local feature vectors in the latent space is manipulated to induce the generation of a slight reddish color difference caused by soft tissue pressure and obstructed blood microcirculation.

[0075] D2, Material Semantic Fusion Unit:

[0076] For transparent or semi-transparent materials used in close-fitting clothing, a two-layer control architecture is constructed within the potential space:

[0077] Using a binary mask to absolutely limit the spatial footprint and generation boundary of a completely opaque material; using a soft mask Quantify the physical light transmittance and perforation density of different transparent or semi-transparent materials;

[0078] In the feature mixing layer of the latent space, a soft mask is used. The continuous grayscale weight values ​​are used to adjust the underlying skin feature tensor of the avatar after processing by the skin texture fine-tuning unit. With the texture feature tensor of clothing fabric Perform cascaded linear interpolation mixing.

[0079] Preferably, in the material semantic fusion unit, the pixel values ​​of the binary mask are discrete {0, 1}; the soft mask The pixel values ​​are continuously distributed between [0.0, 1.0].

[0080] The cascaded linear interpolation obtained by the cascaded linear interpolation mixing for:

[0081] .

[0082] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0083] 1. Extremely high privacy protection and security: Using physiological parameters instead of image-driven methods eliminates the risk of leaking users' sensitive biometric features at the source, significantly reducing the barrier to trying on clothes.

[0084] 2. Low-cost physical interaction restoration: By avoiding complex finite element simulation calculations, the innovative depth map manual scratching technology induces the generation algorithm to automatically generate realistic marks and shadows, solving the industry pain point of strap suspension.

[0085] 3. Supports realistic previews of non-ideal body shapes: The anonymized template library covers real physiological conditions such as underarm fat protrusion and slightly chubby abdomen, providing highly valuable "buyer show" level visual feedback.

[0086] 4. User-friendly interaction: Users only need to input basic physiological parameters to get high-fidelity feedback, which greatly shortens the conversion process. Attached Figure Description

[0087] Figure 1 This is a structural block diagram of the body-fitting clothing try-on system based on anonymized parameter mapping and physical deformation feedback of the present invention;

[0088] Figure 2 This is a schematic diagram of the principle of the multidimensional structural constraint unit of the present invention. Detailed Implementation

[0089] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0090] Example

[0091] This embodiment presents a close-fitting clothing try-on system based on anonymized parameter mapping and physical deformation feedback, such as... Figure 1 As shown, it includes:

[0092] The anonymized parameter mapping and retrieval module receives discrete physiological parameters input by the user during the trial fitting and retrieves N baseline 3D avatar meshes from the anonymized template library. ( A target anonymized 3D avatar mesh M is generated through a linear weighted fusion method; local anatomical features are compensated to obtain the 3D avatar mesh. The 3D avatar mesh M' is divided into several semantic regions and bound to a set of physical response parameters.

[0093] The rule-based geometric feature interference module is used to trigger preset geometric corrections in the coordinate region corresponding to the 3D avatar mesh M' based on the style features of the target clothing; identify the contact path between the target clothing and the skin, and perform negative numerical offset on the contact path pixels on the rendered depth map to generate artificial scratches that simulate physical compression effects; use human semantic segmentation map to decouple the deformed 3D avatar mesh, providing spatial position constraints for subsequent guided generation;

[0094] The image synthesis module uses a diffusion model in conjunction with a control network to inject a depth map containing artificial scratches as a core condition into the denoising process; by utilizing depth gradient changes, it induces the generation of contact shadows with spatial depth, thus eliminating the floating effect of the straps from a visual perspective.

[0095] It also includes a visual enhancement module, which controls skin texture adjustment by adjusting the noise weights and perturbation values ​​in the latent space, avoiding the model-like appearance caused by excessive smoothing; it employs bipartite mask and soft mask gradient fusion technology to ensure that transparent or semi-transparent materials achieve high-fidelity semi-transparent interference in the latent space.

[0096] Specifically, the anonymization parameter mapping and retrieval module includes the following units:

[0097] A1. Human body parameter encoding unit:

[0098] Receive discrete physiological parameters input by the user trying on the garment, and construct a human body parameter vector P:

[0099]

[0100] Where h represents height and w represents weight; Indicates upper chest circumference; The symbol indicates underbust circumference; wa indicates waist circumference; hip indicates hip circumference; and f indicates body fat percentage.

[0101] Based on the large sample statistical range, the human body parameter vector P is subjected to min-max normalization to obtain the human body parameter vector. :

[0102]

[0103] in, , These are the minimum and maximum values ​​of the components of the human body parameter vector P within a large sample statistical range, respectively, to eliminate feature bias caused by different units.

[0104] A2. Human Latent Space Mapping Unit:

[0105] A pre-trained variational autoencoder (VAE) is used as the human parameter encoder network E to process the human parameter vector. Mapped to the latent human shape vector Z in the shape latent space S (low-dimensional feature manifold space):

[0106]

[0107] The latent vector Z of the human body shape is used to describe the circumference ratio, fat distribution, chest, waist and hip structure and local curvature relationship of the human body.

[0108] A3. Multi-template fusion reconstruction unit:

[0109] In the anonymized template library, the KNN algorithm is used to calculate the Euclidean distance, and the N reference 3D avatar meshes that are closest to the human body shape latent vector Z are retrieved in the shape latent space S. ( The anonymized template library is built based on massive 3D human body scanning data. It extracts human geometric topological features from the 3D human body scanning data, and uses a variational autoencoder (3D Mesh VAE) to decouple the human geometric topological features to form a low-dimensional feature latent space S composed of height, weight, upper and lower chest circumference, waist-to-hip ratio, and body fat percentage, and stores it.

[0110] Calculate N baseline 3D avatar meshes based on the reciprocal of the similarity distance. Fusion weights :

[0111]

[0112] Among them, Z i Z j These are the latent vectors of the human body shape for the i-th and j-th reference 3D avatar meshes, respectively;

[0113] The target anonymous 3D avatar mesh M is generated using a linear weighted fusion method:

[0114]

[0115] The generated 3D avatar mesh M completely strips away facial textures and identifiable individual biometric information;

[0116] A4. Soft tissue compensation and reconstruction unit:

[0117] To address non-ideal body deformities that traditional parametric skeleton models cannot represent, this study uses input parameters such as body fat percentage (f) and chest-to-waist ratio. and local circumference gradient changes Constructing a nonlinear soft tissue compensation function Nonlinear soft tissue compensation function It is an anisotropic spatial deformation displacement field function based on body fat percentage and geometric curvature constraints, constructed using the Sigmoid / RBF operator. It nonlinearly maps user-input discrete physiological parameters into normal displacements of the vertices of the 3D avatar mesh to accurately represent soft tissue deformation. It calculates the displacement compensation for each vertex of the 3D avatar network. :

[0118]

[0119] Using displacement compensation Compensation was performed on local anatomical features such as accessory breast expansion, abdominal fat accumulation, and sagging to obtain a compensated three-dimensional avatar mesh. :

[0120]

[0121] A5. Body Continuity Constraint Unit:

[0122] To avoid mesh surface tearing or topological fractures caused by nonlinear compensation, a topological continuity constraint energy function for the 3D avatar mesh is established. :

[0123]

[0124] in, Represents adjacent vertex pairs in a 3D avatar mesh. As vertex The unit normal vector;

[0125] By minimizing This limits the rate of change of normal between adjacent vertices, ensuring the smooth continuity and natural transition of the skin surface.

[0126] The topological continuity constraint of the 3D avatar mesh established by the body continuity constraint unit ensures that the 3D avatar mesh after soft tissue compensation has a continuous and smooth manifold surface. This smooth 3D avatar mesh is directly input into the human semantic region binding unit to carry out semantic region binding, and serves as the geometric carrier for ray projection in the local geometry correction unit and the input mesh for high-precision depth map rendering in the depth map interference unit.

[0127] A6. Human semantic region binding unit:

[0128] Using a pre-defined anatomical mapping matrix, perform 3D semantic region segmentation on the 3D avatar mesh M':

[0129]

[0130] in, The total semantic region of the three-dimensional avatar mesh M'; , , , , These are semantic regions for the chest, shoulder straps, waistband, abdomen, and buttocks; each semantic region... Independently bind physical response parameter groups ,in This represents the regional compression coefficient. For soft tissue elasticity parameters, This is the material contact weight, used to drive subsequent dynamic feature perturbations.

[0131] Preferably, a parameter reliability verification unit is also included between the human body parameter encoding unit and the human body latent space mapping unit.

[0132] Based on large-sample statistical distribution, the confidence probability density function of the human body parameter vector is calculated. ;when (like When the value is 0.05, it is determined that there is a counterintuitive error in the input data (such as an extremely abnormal BMI), and the output parameter correction vector is used to attenuate the human body parameter vector P or trigger a pop-up correction.

[0133] The rule-based geometric feature interference module aims to establish cross-modal geometric interaction rules between the 3D digital avatar and the target body-fitting garment, transforming the binding force of the fabric on the human body into spatial constraint signals on a high-order control base map; it includes the following units:

[0134] B1, Local geometric correction element:

[0135] Pre-establish the structural topological feature lines of the target garment (such as the shoulder strap center line, the outline of the underwire, and the path of the side wing straps); use a ray casting algorithm to spatially project the structural topological feature lines of the target garment along the surface tangent direction of the 3D avatar mesh M', thereby locking the set of pressure vertices on the surface of the 3D avatar mesh. ;

[0136] Based on the tightness coefficient and elastic modulus of the target garment, the set of compression vertices is determined. A negative displacement is applied along the normal to the surface of the 3D avatar mesh to perform nonlinear anisotropy correction. The correction formula is as follows:

[0137]

[0138] in, The original 3D coordinate vectors of the vertices on the 3D avatar mesh before compression; Let be the unit normal vector of the currently compressed vertex. The tightness factor of the target garment. Soft tissue elasticity parameters in the physical response parameter set bound to semantic regions (in soft tissue regions such as the waist, abdomen, and chest). (Increase, and attenuate in hard bone areas such as the clavicle and ribs through mesh curvature operators). Through this correction, the three-dimensional avatar mesh M' exhibits local concavities and deformations in three-dimensional space that conform to anatomical logic.

[0139] B2, Depth Map Interference Unit:

[0140] After completing the local geometric correction, the 3D avatar mesh M' is rendered using a preset camera view matrix and orthogonal projection matrix to generate a 2D high-precision depth map (resolution not less than 1024×1024, pixel grayscale values ​​strictly correspond to the Z-Buffer distance in 3D space).

[0141] By using coordinate mapping, the corresponding pixel coordinate paths of the structural topological feature lines of the target garment on the two-dimensional depth map are extracted. This serves as the contact path between the target garment and the skin; a Gaussian first-order partial derivative operator is used as the sliding window kernel to analyze the pixel coordinate path. The pixel values ​​within a preset range on both sides of the center line are non-linearly shifted negatively (reducing the depth gray value, representing the space concave into the human body); the shifted depth section mathematically presents a "U" or "V" shaped spatial structure with smooth ends and a concave middle, thereby generating artificial scratches on the depth map that simulate the physical compression effect.

[0142] B3. Semantic Feature Alignment Unit:

[0143] A dense pose estimation network (such as DensePose) is used to project the 3D avatar mesh obtained after processing the depth map interference unit into a 2D model to generate a human semantic segmentation map that is absolutely aligned with it at the pixel level.

[0144] The 3D avatar mesh region is decoupled into independent channels for human semantic segmentation masks for various body parts (such as left chest, right chest, left shoulder girdle area, right shoulder girdle area, side waist area, back area, etc.) by semantic tags. This mask is not only used to accurately define the spatial boundary between clothing coverage and exposed skin, but also provides strict dynamic region of interest (ROI) spatial location constraints for subsequent local inpainting of diffusion models, preventing geometric interference signals from contaminating unpressed skin areas.

[0145] The image synthesis module, at its core, uses multi-channel conditional tensor guidance to forcibly constrain the latent space denoising trajectory of the Latent Diffusion Model (LDM), enabling the algorithm to perform high-fidelity redrawing in accordance with the physical boundaries constructed by the aforementioned modules. It includes the following units:

[0146] C1. Multidimensional structural constraint element:

[0147] like Figure 2 As shown, a multi-condition input matrix driven by a control network is constructed. A depth map containing artificial scratches and a human semantic segmentation mask are taken as input. The control network extracts the feature maps of the depth map tensor and the human semantic segmentation mask at different network resolution scales to form multi-dimensional structural constraints. Simultaneously, a two-dimensional tiled image of the target garment is used to extract high-dimensional material and texture feature vectors through an independent garment visual encoder, serving as the garment's visual features. The garment visual features and standard Gaussian random noise are interactively fused through a cross-attention mechanism, and then explicitly weighted and fused with the control network output jump connection. After fusion, they are fed into a diffusion model for iterative reverse denoising. Finally, the image is decoded and output to forcibly lock the overall outline of the generated image and the garment's placement.

[0148] C2, Physical Shadow Induction Unit:

[0149] When using the diffusion model to denoise the shape latent space S at the interface between skin and close-fitting clothing, the input depth map exhibits drastic pixel spatial gradient changes at the artificially incised areas (i.e., The gradient signal mathematically guides the network to identify the region as an ambient occlusion (AO) area. Based on latent space trajectory correction using classifier-free guidance (CFG), the diffusion model automatically generates subtle contact shadows and edge dark areas that conform to the logic of real optical scattering at the artificial etched areas of the straps, underwires, and skin. Through this underlying implicit learning and physical condition induction, the visual appearance of the straps or clothing edges is completely eliminated.

[0150] The visual enhancement module focuses on the microscopic representation and material physical interference of generated images, addressing common issues such as over-smoothing, plastic-like appearance, and clipping at the edges of complex transparent materials. It includes the following units:

[0151] D1, Skin Texture Fine-Tuning Unit:

[0152] To recreate realistic skin details on a completely desensitized, biometric-free 3D avatar mesh, non-uniform spatial perturbation noise is introduced in the final critical time step of the diffusion model's reverse denoising process. Based on the human semantic segmentation map, in the exposed skin areas not covered by clothing, the low-dimensional feature tensor space generated by the diffusion model for image denoising is adjusted by adjusting the variance of random variables and noise weights, serving as the latent space. This induces the high-frequency generation of subtle skin textures, pore details, and minor non-identity biometric representations. At the edges of artificial scratches and in areas of compressed skin, the mean shift of local feature vectors in the latent space is manipulated to induce subtle rosy color differences caused by soft tissue pressure and impaired microcirculation, significantly reducing the 3D dummy-like appearance caused by over-smoothing. This yields the avatar's underlying skin feature tensor. ;

[0153] D2, Material Semantic Fusion Unit:

[0154] For transparent or semi-transparent materials such as lace, mesh, and semi-transparent silk used in close-fitting clothing, a two-layer control structure is constructed in the potential space:

[0155] Using a binary mask, the pixel values ​​are discrete {0, 1}, which absolutely limits the spatial occupation and generation boundary of fully opaque materials (such as metal buckles, shaped steel rings, and thick fabrics).

[0156] Using soft masks The pixel values ​​are continuously distributed between [0.0, 1.0], quantifying the physical light transmittance and cutout density of different transparent or semi-transparent materials;

[0157] In the feature mixing layer of the latent space, a soft mask is used. The continuous grayscale weight values ​​are used to adjust the underlying skin feature tensor of the avatar after processing by the skin texture fine-tuning unit. With the texture feature tensor of clothing fabric Perform cascaded linear interpolation mixing; cascaded linear interpolation for:

[0158]

[0159] This technology mandates that the diffusion model, during denoising and generation, retains both the skin color and muscle shadows of the underlying anonymous human body (as a translucent background color), while also overlaying the warp and weft lines of mesh fabric, reflectivity, and lace trim on the surface. This achieves high-fidelity semi-transparent optical interference and multi-layered material interweaving effects without requiring high-performance physical ray tracing rendering.

[0160] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0161] This embodiment provides a method for trying on close-fitting clothing based on anonymized parameter mapping and physical deformation feedback, the steps of which are as follows:

[0162] Step S01, Human Body Parameter Encoding: Receive discrete physiological parameters input by the user during the fitting process via the terminal interface, including height. ,weight Upper chest circumference Underbust ,waistline Hip circumference Body fat percentage Construct a human body parameter vector P; based on a large sample statistical range (e.g., height range [140, 190], weight range [35, 120]), substitute the above discrete parameters into the minimum-maximum normalization formula to convert the physical quantities into a dimensionless standardized input vector. The values ​​of its internal components are strictly limited to the range of [0.0, 1.0].

[0163] Step S02, Parameter reliability verification:

[0164] vector Input the preset joint probability density function A reasonableness check is performed; the body mass index is calculated. and verify The correlation between (chest circumference difference) and body fat percentage f. (If the user maliciously inputs...) and Determine the joint probability from extremely contradictory data. The parameter correction operator is automatically activated to pull back and decay the weights of anomalous features toward the statistical mean, ensuring that the avatar mesh generated subsequently does not suffer from topological distortion.

[0165] Step S03, Human Latent Space Mapping:

[0166] Human body parameter vector The data is fed into a human body parameter encoder network E constructed based on a variational autoencoder (VAE). The VAE network E, through multiple fully connected layers and nonlinear mapping with activation functions, outputs a latent human body shape vector Z in a 32-dimensional continuous manifold space. The latent human body shape vector Z expresses the highly decoupled skeletal proportions, muscle cross-sectional curvature, and local fat enrichment trends, achieving the goal of representing discrete body shapes using continuous mathematical manifolds.

[0167] Step S04, Multi-template Fusion Reconstruction:

[0168] In the anonymized template library (containing no less than 3,000 examples of desensitized 3D human body mesh data), the KNN algorithm is used to retrieve the N templates that have the closest Euclidean distance to the latent vector Z of the human body shape within the 32-dimensional shape latent space S (in this embodiment, we take...). ) Baseline 3D Avatar Mesh Calculate the inverse distances between these N baseline 3D avatar meshes and the target vector, and convert them into normalized Softmax weights. Finally, a linear weighted fusion method was used. The initial three-dimensional avatar mesh M is generated by fusion and has a specific ratio and completely stripped of facial and individual biometric features.

[0169] Step S05, Soft tissue compensation and reconstruction:

[0170] To address the need for accurate body shape reproduction in close-fitting clothing fitting scenarios, body fat percentage (f) and local circumference gradient changes are extracted. A nonlinear soft tissue compensation function is applied to a specific set of vertices belonging to the accessory breast area, waist and abdomen area, and lower buttock line in the 3D avatar mesh M. For example, in the area where the armpit meets the chest, the displacement compensation amount for the outward expansion of the accessory breast is calculated based on the difference in chest circumference and body fat percentage. This is then superimposed onto the original mesh to obtain the compensated 3D avatar mesh M'.

[0171] Step S06, Human semantic region binding:

[0172] Divide the 3D avatar mesh M' into the chest area. shoulder strap Waistband ,abdomen Each semantic region is an independent semantic region. In the data structure, labels are attached to the vertices of each semantic region, and specific physical response parameter sets are bound to them. For example, for the abdominal semantic region. Assign higher soft tissue elasticity parameters Assigning lower soft tissue elasticity parameters to the rib semantic region This is used to drive subsequent physical compression simulations.

[0173] Step S07, Local geometric correction:

[0174] Obtain the structural topological feature lines (shoulder strap center line) of the target garment (such as a lace backless bra). Underwire Using the ray casting algorithm, along the unit normal vector of the surface of the 3D avatar mesh M'. Perform back projection to lock the set of compressed vertices covered by the clothing lines. .

[0175] Read the soft tissue elasticity parameters (such as those of the shoulder strap semantic region) from the physical response parameter group bound to the semantic region. (and combine the garment's own style and tightness factor) , making the vertex set A physical displacement or offset inwards from the body is generated along the negative direction of the normal:

[0176]

[0177] Step S08, Depth Map Interference:

[0178] Configure the virtual camera's view matrix V and orthographic projection matrix Q to render the avatar M' after 3D deformation. A two-dimensional depth map D with a resolution of [resolution missing]. The grayscale value of each pixel in the two-dimensional depth map D is strictly mapped to the Z-Buffer distance.

[0179] Projecting the structural topological feature lines of the target garment onto this depth map determines the pixel coordinate path. Path with pixel coordinates Centered on, using standard deviation Sliding window convolution is performed on the pixel's first-order Gaussian partial derivative operator, along with the pixel coordinate path. The grayscale values ​​within a certain radius on both sides are nonlinearly shifted negatively (reducing the grayscale value), artificially constructing a smooth, concave "V-shaped" cross-section on the two-dimensional control source. Mathematically, this cross-section constitutes an "artificial notch" tensor representing the physical compression effect.

[0180] Semantic feature alignment:

[0181] A dense pose estimation network (such as DensePose) is used to project a 2D image onto the 3D avatar mesh obtained after depth map perturbation, generating a human semantic segmentation map that is absolutely aligned with it at the pixel level.

[0182] Step S09, Multidimensional structural constraints:

[0183] Initiate the denoising loop for the Latent Diffusion Model (LDM), with a total time step of 1. Depth map containing artificial scratches Compared with the human semantic segmentation map generated by the DensePose dense pose estimation network As a spatial control flow injection into the control network ControlNet.

[0184] Simultaneously, the two-dimensional tiled image of the target garment is input into the CLIP-ViT-H / 14 garment visual encoder to extract high-dimensional material and texture feature vectors. As the denoising time step t decreases from 50 to 1, the ControlNet network forcibly imposes structural constraints in the Cross-Attention layer of the U-Net, preventing the ControlNet from randomly altering the core skeletal outline of the human body and the occupational boundaries of the clothing during denoising.

[0185] Step S10, Physical Shadow Induction:

[0186] In the core stage of noise reduction iteration ( Due to the drastic spatial gradient changes caused by artificial markings in the input depth map (), The gradient difference is implicitly identified as an ambient occlusion (AO) feature by the algorithm during the classifier-free guidance (CFG) stage of the main network. The denoising network spontaneously fills in high-probability dark-toned pixels at the edges of the straps, the underwire, and the artificial creases on the skin, thus automatically creating tiny contact shadows that conform to the laws of real optical scattering in the final generated image, completely eliminating the visual floating effect of the clothing edges.

[0187] Step S11, Fine-tuning skin texture:

[0188] In the tail stage of noise denoising ( The system identifies semantic regions of exposed skin not covered by clothing (such as the chest and sides), injects non-uniform high-frequency spatial perturbation noise into the latent space, and forces the network to generate natural skin texture, pores, minor imperfections, and blood vessel patterns. Specifically, the system identifies the coordinate edges of artificial marks and, by manipulating the mean shift of the local latent feature vector, guides the network to generate a low-probability erythema simulation at the mark boundaries, mimicking the obstructed microcirculation of blood flow after soft tissue compression.

[0189] Step S12, Material semantic fusion:

[0190] For transparent or semi-transparent materials in clothing, the system activates dual-layer control in the LatentFeature Blending Layer of the latent space: completely opaque buttons and underwires are categorized into a bipartite mask. Mesh fabrics with continuously varying light transmittance are classified as soft masks. .

[0191] During latent feature fusion, the system performs cascaded linear interpolation:

[0192]

[0193] This allows the mesh area to reveal the underlying anonymous skin color and muscle shadows while clearly presenting the warp and weft weave of lace on the surface, achieving a high-fidelity semi-transparent physical interference effect during the final noise reduction output.

[0194] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A body-fitting clothing try-on system based on anonymized parameter mapping and physical deformation feedback, characterized in that: include: The anonymized parameter mapping and retrieval module receives discrete physiological parameters input by the user during the trial fitting and retrieves N baseline 3D avatar meshes from the anonymized template library. ; A target anonymized 3D avatar mesh M is generated using a linear weighted fusion method; local anatomical features are compensated to obtain the 3D avatar mesh. ; Perform 3D semantic region segmentation on the 3D avatar mesh M' into several semantic regions and bind physical response parameter groups; The rule-based geometric feature interference module is used to trigger preset geometric corrections in the coordinate region corresponding to the 3D avatar mesh M' based on the style features of the target clothing; identify the contact path between the target clothing and the skin, and perform negative numerical offset on the contact path pixels on the rendered depth map to generate artificial scratches that simulate physical compression effects; and use human semantic segmentation map to decouple the deformed 3D avatar mesh. The image synthesis module uses a diffusion model in conjunction with a control network to inject a depth map containing artificial scratches as a core condition into the denoising process; by utilizing depth gradient changes, it induces the generation of contact shadows with spatial depth, thus eliminating the floating effect of the straps from a visual perspective. And a visual enhancement module, used to control skin texture adjustment by adjusting the noise weights and perturbation values ​​in the latent space, avoiding the model-like appearance caused by excessive smoothing; By employing a gradient fusion technique combining a binary mask and a soft mask, high-fidelity semi-transparent interference is achieved in the latent space for transparent or semi-transparent materials.

2. The body-fitting clothing try-on system based on anonymized parameter mapping and physical deformation feedback according to claim 1, characterized in that: The anonymization parameter mapping and retrieval module includes the following units: A1. Human body parameter encoding unit: The system receives discrete physiological parameters input from users trying on clothes, constructs a human body parameter vector P, and performs max-min normalization based on a large sample statistical range to obtain the final human body parameter vector. ; A2. Human Latent Space Mapping Unit: A pre-trained variational autoencoder is used as the human parameter encoder network E to process the human parameter vector. This is mapped to the latent vector Z of the human body shape in the latent shape space S; A3. Multi-template fusion reconstruction unit: In the anonymized template library, the KNN algorithm is used to calculate the Euclidean distance, and the N reference 3D avatar meshes that are closest to the human body shape latent vector Z are retrieved in the shape latent space S. The anonymized template library is built based on massive 3D human body scanning data. It extracts human geometric topological features from the 3D human body scanning data, decouples the human geometric topological features to form a low-dimensional feature shape latent space S, and stores it. Calculate N baseline 3D avatar meshes based on the reciprocal of the similarity distance. Fusion weights The target anonymous 3D avatar mesh M is generated through linear weighted fusion; the generated 3D avatar mesh M completely strips away facial textures and identifiable individual biometric information; A4. Soft tissue compensation and reconstruction unit: Based on the input body fat percentage f and chest-to-waist ratio and local circumference gradient changes Constructing a nonlinear soft tissue compensation function Calculate the displacement compensation amount for each vertex of the 3D avatar network. ; Using displacement compensation Compensation was performed on local anatomical features such as accessory breast expansion, abdominal fat accumulation, and sagging to obtain a compensated three-dimensional avatar mesh. ; A5. Body Continuity Constraint Unit: Establish the energy function for topological continuity constraints of the 3D avatar mesh. By minimizing Limiting the rate of change of normal between adjacent vertices ensures the smooth continuity and natural transition of the skin surface; A6. Human semantic region binding unit: Using a pre-defined anatomical mapping matrix, perform 3D semantic region segmentation on the 3D avatar mesh M'; for each semantic region Independently bind physical response parameter groups ,in This represents the regional compression coefficient. For soft tissue elasticity parameters, This represents the material contact weight.

3. The body-fitting clothing try-on system based on anonymized parameter mapping and physical deformation feedback according to claim 2, characterized in that: In the human body parameter encoding unit, the human body parameter vector P is constructed as follows: ； Where h represents height and w represents weight; Indicates upper chest circumference; The symbol indicates the underbust measurement; wa indicates the waist measurement; hip indicates the hip measurement; and f indicates the body fat percentage. Human body parameter vector for: ； in, , These are the minimum and maximum values ​​of the components of the human body parameter vector P within the large sample statistical range, respectively. In the human body latent space mapping unit, the latent vector Z of the human body shape is: ； In the multi-template fusion reconstruction unit, the fusion weights for: ； Among them, Z i Z j These are the latent vectors of the human body shape for the i-th and j-th reference 3D avatar meshes, respectively; The target anonymous 3D avatar mesh M is generated using a linear weighted fusion method as follows: ； In the soft tissue compensation and reconstruction unit, the displacement compensation amount of each vertex of the three-dimensional avatar network for: ； Compensated 3D Avatar Mesh for: ； In the body continuity constraint unit, the energy function of the three-dimensional avatar mesh topological continuity constraint : ； in, Represents adjacent vertex pairs in a 3D avatar mesh. As vertex The unit normal vector; In the human semantic region binding unit, the three-dimensional semantic region segmentation formula is: ； in, The total semantic region of the three-dimensional avatar mesh M'; , , , , These are the semantic regions for the chest, shoulder straps, waistband, abdomen, and buttocks, respectively.

4. The body-fitting clothing try-on system based on anonymized parameter mapping and physical deformation feedback according to claim 2, characterized in that: Between the human body parameter encoding unit and the human body latent space mapping unit, there is also a parameter reliability verification unit: Calculate the confidence probability density function of the human body parameter vector ;when If the input data is found to contain counterintuitive errors, the output parameter correction vector is used to attenuate the human body parameter vector P or trigger a pop-up correction.

5. The body-fitting clothing try-on system based on anonymized parameter mapping and physical deformation feedback according to claim 1, characterized in that: The rule-based geometric feature interference module includes the following units: B1, Local geometric correction element: The structural topological feature lines of the target garment are pre-established; using a ray casting algorithm, the structural topological feature lines of the target garment are spatially projected along the surface tangent direction of the 3D avatar mesh M', thereby locking the set of compressed vertices on the surface of the 3D avatar mesh. ; Based on the tightness coefficient and elastic modulus of the target garment, the set of compression vertices is determined. Apply a negative displacement along the surface normal of the 3D avatar mesh to perform nonlinear anisotropy correction; B2, Depth Map Interference Unit: The 3D avatar mesh M' after local geometric correction is rendered into a 2D depth map using a preset camera view matrix and orthogonal projection matrix; By using coordinate mapping, the corresponding pixel coordinate paths of the structural topological feature lines of the target garment on the two-dimensional depth map are extracted. This serves as the contact path between the target garment and the skin; a Gaussian first-order partial derivative operator is used as the sliding window kernel to analyze the pixel coordinate path. The pixel values ​​within a preset range on both sides of the center line are non-linearly offset in a negative direction; the offset depth section mathematically presents a "U" or "V" shaped spatial structure with smooth ends and a concave middle, thereby generating artificial grooves on the depth map that simulate physical compression effects. B3. Semantic Feature Alignment Unit: A dense pose estimation network is used to project the 3D avatar mesh obtained after processing the depth map interference unit into a 2D model, generating a human semantic segmentation map that is absolutely aligned with it at the pixel level. The 3D avatar mesh region is decoupled into independent channels for each part of the body using semantic tags to create a human semantic segmentation mask.

6. The body-fitting clothing try-on system based on anonymized parameter mapping and physical deformation feedback according to claim 5, characterized in that: In the local geometric correction unit, the correction formula for performing nonlinear anisotropy correction is: ； in, The original 3D coordinate vectors of the 3D avatar mesh; Let this be the unit normal vector of the currently compressed vertex. The tightness factor of the target garment. The soft tissue elasticity parameters in the physical response parameter group bound to the semantic region.

7. The body-fitting clothing try-on system based on anonymized parameter mapping and physical deformation feedback according to claim 1, characterized in that: The image synthesis module includes the following units: C1. Multidimensional structural constraint element: A multi-condition input matrix driven by a control network is constructed. A depth map containing artificial scratches and a human semantic segmentation mask are taken as input. The control network extracts feature maps of the depth map tensor and the human semantic segmentation mask at different network resolution scales to form multi-dimensional structural constraints. Simultaneously, a two-dimensional tiled image of the target garment is used to extract high-dimensional material and texture feature vectors through an independent garment visual encoder, serving as the garment's visual features. The garment visual features and standard Gaussian random noise are interactively fused through a cross-attention mechanism, and then explicitly weighted and fused with the control network output via a skip connection. After fusion, they are fed into a diffusion model for iterative reverse denoising, and finally output through image decoding. C2, Physical Shadow Induction Unit: Based on latent space trajectory correction without classifier guidance, the diffusion model automatically generates contact shadows and edge dark areas that conform to the logic of real optical scattering at the artificial marks on the straps, steel rings, and skin.

8. The body-fitting clothing try-on system based on anonymized parameter mapping and physical deformation feedback according to claim 1, characterized in that: The visual enhancement module includes the following units: D1, Skin Texture Fine-Tuning Unit: Based on the human semantic segmentation map, in the exposed skin area not covered by clothing, the low-dimensional feature tensor space generated by the diffusion model for image denoising is adjusted by adjusting the variance of random variables and noise weights, serving as the latent space. This induces the high-frequency generation of minute skin textures, pore textures, and minute non-identity biometric features. At the edges of artificial scratches and in the area of ​​skin under pressure, the mean shift of the local feature vectors in the latent space is manipulated to induce the generation of a slight reddish color difference caused by soft tissue pressure and obstructed blood microcirculation. D2, Material Semantic Fusion Unit: For transparent or semi-transparent materials used in close-fitting clothing, a two-layer control architecture is constructed within the potential space: Using a binary mask to absolutely limit the spatial footprint and generation boundary of a completely opaque material; using a soft mask Quantify the physical light transmittance and perforation density of different transparent or semi-transparent materials; In the feature mixing layer of the latent space, a soft mask is used. The continuous grayscale weight values ​​are used to adjust the underlying skin feature tensor of the avatar after processing by the skin texture fine-tuning unit. With the texture feature tensor of clothing fabric Perform cascaded linear interpolation mixing.

9. The body-fitting clothing try-on system based on anonymized parameter mapping and physical deformation feedback as described in claim 8, characterized in that: In the material semantic fusion unit, the pixel values ​​of the binary mask are discrete {0, 1}; the soft mask... The pixel values ​​are continuously distributed between [0.0, 1.0]. The cascaded linear interpolation obtained by the cascaded linear interpolation mixing for: 。

Images