System for coping with ai virtual fitting of sweater and intelligent switching generation in multiple scene backgrounds

Abstract

The application discloses an AI virtual fitting and multi-scene background intelligent switching generation system for a sweater, relates to the technical field of AI virtual generation, and comprises an acquisition module, a fitting configuration module, a background switching module and a virtual generation module.The technical key points are as follows:an emulation file of a target sweater is acquired;in response to a configuration input operation of the emulation file, fitting effect data groups of the target sweater are acquired in a first virtual scene; wherein the fitting effect data groups comprise an input associated template layer identifier and a scene matching identifier; in response to a background switching action of the first virtual scene, a second virtual scene is obtained, scene parameters of the second virtual scene are analyzed, and the scene parameters are injected into the emulation file; in response to an AI generation operation of the emulation file, fitting rendering scheduling files are generated according to the fitting effect data groups, the rendering scheduling files and the current emulation file are combined to generate effective fitting samples; and the application is adapted to multi-scene background switching, and the precision of AI virtual fitting is improved.

Description

Technical Field

[0001] This invention relates to the field of AI virtual generation technology, specifically to an AI virtual try-on and multi-scene background intelligent switching generation system for sweaters. Background Technology

[0002] AI virtual try-on technology API is an interface service that integrates multiple capabilities such as artificial intelligence image recognition, human posture detection, clothing fitting, and visual rendering. One approach is to acquire human point clouds through a 3D scanner and then simplify and extract features using 3D point cloud processing technology. Another approach relies on large-scale samples and neural network training methods to predict human joints in uploaded 2D human images, reconstruct the human body for feature extraction, and thus achieve AI virtual try-on. For clothing like sweaters, although AI virtual try-on technology has made significant progress overall, existing solutions still have the following limitations: First, existing solutions typically use standardized simulated human body models for try-on simulations, lacking precise adaptation to key dimensions such as shoulder width, chest circumference, and waist circumference of the target person, greatly reducing the try-on effect. Even if some systems support body shape parameter adjustments, it is difficult to accurately match the sweater with the human body curves, resulting in unnatural sweater texture details and a strong AI feel. Second, existing systems mostly replace static images with backgrounds, lacking an intelligent adaptation mechanism between clothing and background. For example, when the scene changes from commuting to a beach vacation, the lighting and shadow effects of the sweater are not adjusted synchronously, resulting in obvious light and shadow discontinuities between the clothing and the background, reducing the overall visual effect. In addition, all scene parameters need to be reloaded when changing scene backgrounds, which greatly increases latency to some extent and cannot meet the user's need for smooth interaction. Summary of the Invention

[0003] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an AI-powered virtual try-on and multi-scene background intelligent switching generation system for sweaters, comprising an acquisition module, a try-on configuration module, a background switching module, and a virtual generation module. By acquiring a simulation file of the target sweater and responding to configuration input operations in the simulation file, the system acquires the try-on effect data set corresponding to the associated input template layer identifier and scene matching identifier in a default first virtual scene. Then, in response to scene switching actions, a second virtual scene different from the first virtual scene is generated, scene parameters are parsed, and accurately injected into the simulation file to achieve multi-scene adaptation and improve the overall visual effect. In response to AI generation operations, a try-on rendering schedule file is generated based on the try-on effect data set. The rendering schedule file is then merged with the current simulation file to generate valid try-on samples matching multiple measurement points, supporting multi-scene virtual try-on display and solving the problems mentioned in the background technology.

[0004] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: This application provides an AI virtual try-on and multi-scene background intelligent switching generation system for sweaters, the system including: an acquisition module for acquiring simulation files of the target sweater; The fitting configuration module responds to the configuration input operation of the simulation file and obtains the fitting effect data set of the target sweater in the first virtual scene; the fitting effect data set includes the associated sample layer identifier and scene matching identifier. The background switching module responds to the background switching action of the first virtual scene, obtains the second virtual scene, parses the scene parameters of the second virtual scene, and injects them into the simulation file; The virtual generation module responds to the AI ​​generation operation of the simulation file, generates a fitting rendering schedule file based on the fitting effect data set, and combines the rendering schedule file with the current simulation file to generate a valid fitting sample.

[0005] Furthermore, it also includes: importing target personnel images, extracting pixel coordinates of the target personnel images and applying them to the template layer through the fitting rendering schedule file, combining effective fitting samples to generate a target dataset, training the target dataset with a target neural network, and outputting AI virtual fitting images.

[0006] Furthermore, in response to the configuration input operation of the simulation file, the fitting effect data set of the target sweater in the first virtual scene is displayed, including: Display at least one input control unit, including an associated template layer input control and a scene-matching input control: The template layer input control has shape constraints and gap constraints, and there is a one-to-one correspondence between the template layer input control and the template layer identifier; the scene matching input control corresponds to a one-to-one correspondence between the scene matching identifier; In response to an input operation on the input control unit, display the fitting effect data set entered into the input control unit.

[0007] Furthermore, the simulation files also include pre-set simulated human body models.

[0008] Furthermore, morphological constraints include: Extract the size parameters and texture features of the target sweater; A two-dimensional coordinate system is established, and a two-dimensional edge curve for the dimensional parameters is fitted to form a two-dimensional template plane. Through mesh generation, the two-dimensional template is discretized into a mesh model composed of multiple mass points, and an initial coordinate index is assigned to each mass point to locate the corresponding measurement point in the mesh model. Simultaneously, the instantaneous coordinates of the mesh model during the fitting process are monitored, the displacement vector between the instantaneous coordinates and the initial coordinate index is calculated, and the deformation deviation is generated. The conformal factors of multiple measurement points are determined by combining texture features, and linear weighted analysis is performed on the deformation deviation and conformal factors at the measurement points to generate the morphological regression gradient. The gradient threshold of ±wc% is set according to the morphological regression gradient, and the cases where the morphological regression gradient deviates from the gradient threshold are screened to adjust the template layer. Among them, the conformal factor is a dynamic quantity.

[0009] Furthermore, gap constraints include: Import a preset simulated human body model, monitor the distance between the mesh model and the simulated human body model along the first direction of the simulated human body model, and collect them into a distance sequence; where the first direction represents the normal direction; Clustering of distance sequence rectangles to identify the first and second anomalous regions; Evaluation parameters, including the area of ​​the abnormal region and the knitting density, are extracted based on the first and second abnormal regions. Based on the normalization of evaluation parameters, evaluation duones are constructed, and an evaluation feature space is constructed based on the evaluation duones. The standard score of the evaluation duotuple in the evaluation feature space for the preset period is used as the offset. Cases with offsets greater than or equal to the offset threshold are filtered out, and the corresponding mesh is matched to update the template to adjust the template layer.

[0010] Furthermore, responding to the background switching action of the first virtual scene includes: Analyze the background switching action to obtain the background semantic identifier; Based on background semantic identifiers, the background source files of the second virtual scene are retrieved from a pre-defined multi-scene background library; Parse the background source file to obtain the scene parameters of the second virtual scene, and generate the second virtual scene; The scene parameters include the light field vector and the spatial depth distribution.

[0011] Furthermore, the second virtual scene includes: Establish the association between scene parameters and simulation files, including lighting and shadow linkage and spatial pose association: Light and shadow linkage: Based on the light field vector, the mesh model is retrieved, and a first semantic label is created for the mesh model according to the background semantic identifier, forming a linkage between the light field vector and the mesh model; Spatial pose association: Based on spatial depth distribution, retrieve the simulated human body model, and create a second semantic label for the simulated human body model according to the background semantic identifier, forming an association and linkage between spatial depth distribution and simulated human body model; Based on the association relationship and combined with the scene matching identifier, the classification rendering range of the background source file for each scene is obtained.

[0012] Furthermore, in response to the AI ​​generation operation of the simulation file, a fitting rendering schedule file is generated based on the fitting effect data set, including: Retrieve the grid update template corresponding to the pattern layer identifier in the fitting effect data group; Add the scene matching identifier and category rendering range in the virtual fitting effect data group to the corresponding mesh update template to obtain the rendering template corresponding to each virtual fitting effect data group; The rendering template is used to add the scene parameters corresponding to the scene matching identifier of the simulated human body model to the mesh model corresponding to the template layer identifier; based on each rendering template, the fitting rendering schedule file is obtained.

[0013] Furthermore, the classification rendering range corresponds to the first semantic tag and the second semantic tag.

[0014] (III) Beneficial Effects This invention provides an AI virtual try-on system for sweaters and an intelligent multi-scene background switching generation system, which has the following beneficial effects: 1. This invention obtains virtual fitting effect data sets through template layer input controls and scene matching input controls. In the template layer input controls, shape constraints and gap constraints are set. To a certain extent, through shape constraints, using size parameters and texture features as analysis objects, the magnitude of the shape-preserving factor is quantified by calculating the degree of deformation deviation, thereby linearly weighting and generating a shape regression gradient. Through gap constraints, using distance sequences as analysis objects, the fluctuation and amplitude variation characteristics are identified. Through specific rectangular clustering, the first and second abnormal regions of the measurement points are identified, and evaluation parameters are extracted, thereby generating an offset. By adjusting the template layer through the shape regression gradient and offset, the accuracy and realism of virtual fitting are improved, greatly enhancing the adaptability to key points. 2. By setting a classification rendering range and using high-precision and low-precision rendering, this invention eliminates the need to load all scene parameters when the scene background changes, greatly reducing latency and meeting the user's need for smooth interaction. Through the matching and binding of background semantic tags, scene parameters and simulation files form a one-to-one correlation and linkage relationship, ensuring that lighting effects and spatial poses automatically adapt with scene changes, thereby improving the realism of AI virtual try-on. 3. This invention imports target personnel images, extracts pixel coordinates from the target personnel images using a virtual try-on rendering schedule file, applies them to a template layer, combines effective virtual try-on samples to generate a target dataset, and inputs it into a target neural network for training. This enables collaborative rendering of the target sweater with multiple scenes. The trained model can quickly generate personalized, high-fidelity, multi-scene adapted AI virtual try-on images to meet user needs. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the modules of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0017] The core of this invention lies in: acquiring a simulation file of the target sweater, responding to the configuration input operation of the simulation file, and in the default first virtual scene, acquiring the fitting effect data group corresponding to the associated input template layer identifier and scene matching identifier; then, responding to the scene switching action to generate a second virtual scene different from the first virtual scene, parsing the scene parameters, and accurately injecting them into the simulation file to achieve multi-scene adaptation and improve the overall visual effect; responding to the AI ​​generation operation, generating a fitting rendering schedule file based on the fitting effect data group, merging the rendering schedule file with the current simulation file, generating effective fitting samples matching multiple measurement points, and supporting multi-scene virtual fitting display.

[0018] Example: This invention provides an AI virtual try-on system for sweaters and an intelligent multi-scene background switching generation system. Figure 1 This is a schematic diagram of the modules of the present invention; please refer to it. Figure 1 The system includes: an acquisition module, a fitting configuration module, a background switching module, and a virtual generation module, and the acquisition module, fitting configuration module, background switching module, and virtual generation module are interconnected. The following is an explanation of each module involved: Acquisition module: Acquires the simulation file of the target sweater; It should be noted that the simulation file is used to directly drive the entire process of subsequent virtual try-on configuration, background switching, AI generation of the target sweater, and target neural network training. That is, it is a structured file of core data for AI virtual try-on of the target sweater and multi-scene adaptation. The file format adopts GLB format, which is compatible with the virtual try-on rendering engine and is easy to adapt, call and store. For various scenarios of virtual try-on for target sweaters, simulation files for the target sweaters can be categorized into different types according to the applicable scenario and sweater characteristics. Different types of simulation files differ in their core data. For example, based on sweater characteristics, they can be divided into basic texture sweaters and complex texture sweaters, corresponding to different texture features. Basic texture sweaters primarily feature simple knit structures like plain knit and rib, corresponding to low texture complexity, while complex texture sweaters primarily feature complex knit structures, corresponding to high texture complexity. The simulation file of the target sweater to be uploaded can be obtained through a data acquisition and configuration terminal. This involves selecting real sweater data captured using a high-definition image acquisition component from a local folder, including but not limited to front, side, and texture detail images, and inputting the physical parameters of the real sweater, including but not limited to material labels, pattern labels, and nominal sizes, to generate the simulation file. Additionally, the simulation file includes a preset simulated human body model. It should be noted that the simulated human body model can be set according to standard male and female human body models. This model has a built-in posture library for different scenarios, and the corresponding posture parameters can be synchronously called when switching scenes, allowing the sweater try-on effect to naturally blend with the scene movements.

[0019] The fitting configuration module responds to the configuration input operation of the simulation file and obtains the fitting effect data set of the target sweater in the first virtual scene; the fitting effect data set includes the associated sample layer identifier and scene matching identifier. In response to configuration input operations in the simulation file, display the fitting effect data set of the target sweater in the first virtual scene, including: displaying at least one input control unit, including an associated template layer input control and a scene matching input control: The template layer input control is set with shape constraints and gap constraints, and the template layer input control corresponds one-to-one with the template layer identifier. The template layer identifier in the same fitting effect data group is used to determine the locking of the geometric mesh in the simulation space. Shape constraints include: extracting the size parameters and texture features of the target sweater: performing image segmentation on the uploaded target sweater, including segmenting the front, side and texture detail images of the target sweater to obtain an interference-free target sweater foreground mask, calling the HRNet model to locate key size points within the foreground mask, including but not limited to the shoulder peaks, chest height points, the widest points of the waist on both sides, the cuff endpoints and the lowest point of the neckline, outputting the pixel coordinates of each key size point, and obtaining size parameters by calculating the Euclidean distance between different pixel coordinates, and the size parameters include but are not limited to the front chest circumference, front hem circumference, back chest circumference, back hem circumference, front length and back length; A two-dimensional coordinate system is established with the lowest point of the neckline as the origin, the horizontal axis pointing to the right (width direction of the target sweater), and the vertical axis pointing downwards (length direction of the target sweater). A two-dimensional edge curve is fitted to the dimensional parameters to form a two-dimensional template plane. A structured quadrilateral mesh is used to divide the two-dimensional template plane into a mesh model composed of multiple mass points, with the geometric center of each mesh unit being a mass point. An initial coordinate index (X0, Y0) is assigned to each mass point, allowing the location of corresponding measurement points within the mesh model, including but not limited to… Limited to the first and second shoulder height points, the first and second chest height points, the first and second back height points, the first and second lateral waist points; simultaneously, the instantaneous coordinates (Xt, Yt) of the mesh model during the fitting process are monitored, and the displacement vector (Xt-X0, Yt-Y0) between the instantaneous coordinates and the initial coordinate indices is calculated. Through standardization, the displacement vector is converted into a scalar form of deformation deviation. That is, for a certain measurement point, its displacement vector is first converted into a scalar form, and the square of the horizontal coordinate of the displacement vector (Xt-X0) is calculated respectively. 2 And the square of the ordinate (Yt-Y0). 2 The absolute displacement length of the measurement point is obtained by adding the two squared values ​​and taking the square root. Then, the absolute displacement length is divided by the straight-line distance from the initial coordinates of the particle to the origin of the coordinate system to obtain the dimensionless deformation deviation. Here, X0 represents the initial spatial abscissa, Y0 represents the initial spatial ordinate, Xt represents the abscissa fitted at time t during the fitting process, and Yt represents the ordinate fitted at time t during the fitting process. It should be noted that the deformation deviation ranges from 0 to 1. The closer the value is to 1, the greater the deformation of the target sweater. The closer the value is to 0, the less deformation the target sweater has. First, a pre-defined mapping rule table is invoked using texture features to establish a hierarchical mapping rule between texture features and conformability factors, and the initial conformability factors for multiple measurement sites are automatically determined; among them, texture features include texture complexity and knitted structure; It should be noted that the shape-preserving factor is a dynamic quantity. In the subsequent determination of the shape-preserving factor, all texture features of each measurement point of the extracted target sweater are obtained. These texture features are then ordered according to the row and column indices of the mesh model and collected into a first feature set. Each element in the first feature set corresponds one-to-one with a measurement point in the mesh model. The dimensions of the elements within the set include the knitting structure and texture complexity. Based on the mesh model, a row-first approach is used to extract values ​​from every two mass points, and a column-first approach is used to extract values ​​from every two mass points in the first feature set. The core texture features of each measurement point are extracted, redundant texture features are removed, and the extracted features are then... The core texture features are bound to the row and column indices of the corresponding measurement points and collected into a second feature set. Each element in the second feature set is input into a preset mapping rule table. Higher texture complexity is assigned a higher shape preservation factor, and lower texture complexity is assigned a lower shape preservation factor. More complex knitted structures are assigned a higher shape preservation factor, and simpler knitted structures are assigned a lower shape preservation factor. For example: plain knit: shape preservation factor < vertical ribbing: shape preservation factor < horizontal ribbing: shape preservation factor < cable knit: shape preservation factor < jacquard: shape preservation factor < braided knit. This ensures that the shape preservation factor of each measurement point corresponds one-to-one with the texture feature it monitors. Linear weighted analysis was performed on the deformation deviation and conformity-preserving factor at the measurement sites. This involved multiplying the deformation deviation by the conformity-preserving factor, subtracting the conformity-preserving factor from 1, and then multiplying by the deformation deviation again. The results of these two multiplications were then summed to obtain the morphological regression gradient. During the linear weighted analysis, deformation deviation data generated during the mesh model fitting process for each measurement site were acquired. This data was matched with the conformity-preserving factor according to row and column indices. A linear weighting method was used to measure the linear correlation between deformation deviation and the conformity-preserving factor, automatically assigning higher conformity-preserving factors to measurement sites with larger deformation deviations and lower conformity-preserving factors to measurement sites with smaller deformation deviations. Simultaneously, a standard morphological regression gradient was set, and the difference between the morphological regression gradient and the standard morphological regression gradient was calculated. A gradient threshold of ±wc% was set for this difference. The process involves calculating gradient thresholds where the difference is greater than ±wc%, indicating that the current difference is not within the ±wc% threshold. This filters out cases where the morphological regression gradient deviates from the threshold and adjusts the template layer accordingly. Here, wc is a constant, and ±wc% represents the percentage fluctuation range based on wc, with wc greater than 0. For example, setting wc to 5 means ±5% is an upward and downward adjustment of 5% based on the difference. It's important to note that a larger morphological regression gradient indicates a greater deviation of the measurement point from its initial coordinates, requiring a corresponding shift back towards the initial coordinates. Measurement points with morphological regression gradients within the threshold maintain their current coordinates and undergo smoothing. After adjusting all measurement points, the edge curves are refitted to obtain the calibrated template layer. By using morphological constraints and taking size parameters and texture features as the analysis objects, the magnitude of the shape preservation factor is quantified by calculating the degree of deformation deviation, thereby generating a morphological regression gradient through linear weighting. This achieves automated and accurate correction of the geometric contour of the template layer. To a certain extent, it provides a standardized template layer for subsequent AI virtual try-on, improving the accuracy and realism of virtual try-on. The gap constraint includes: importing a preset simulated human body model, monitoring the distance between the mesh model and the simulated human body model along the first direction of the simulated human body model, and aggregating it into a distance sequence; wherein, the first direction represents the normal direction; by following the normal direction, the straight-line distance from each mass point of the mesh model to the surface of the simulated human body model is obtained. It should be noted that a distance greater than 0 indicates the existence of a gap, and a distance less than or equal to 0 indicates the existence of clipping. The rectangular clustering of distance sequences includes: statistical analysis of the distance sequences to identify the variation characteristics of distance fluctuations and amplitudes within the corresponding time period. The degree of fluctuation reflects the intensity of the sliding of the mesh model on the simulated human body model, and the degree of amplitude variation reflects the distance deviating from the preset gap. Fluctuation types are classified according to the fluctuation characteristics. By calculating the average value μ and standard deviation σ of the sampled sequence over the preset time period, those falling within the interval [μ-σ, μ+σ] are identified as medium waveforms, those less than [μ-σ, μ+σ] as small waveforms, and those greater than [μ-σ, μ+σ] as large waveforms. Amplitude types are also classified according to the amplitude variation characteristics. For example, the maximum amplitude of the sampled sequence within the preset time period is denoted as M; [0, 0.3M] is classified as a small amplitude type, (0.3M, 0.7M) as a medium amplitude type, and (0.7M, M) as a large amplitude type. It should be noted that parameter settings such as 0.3 and 0.7 are only examples and should be set according to the actual situation. They will not be elaborated here. The horizontal axis interval is defined by the fluctuation type and the vertical axis interval by the amplitude type. The horizontal and vertical axis intervals form a rectangular interval. Within the rectangular interval, multiple combination types are divided by cross-combination of fluctuation type and amplitude type. At the same time, the eight measurement points, namely the first and second shoulder height points, the first and second chest height points, the first and second back height points, the first and second side waist points, and the second side waist points, are expanded to form eight regions: left and right shoulders, left and right chest, left and right back, and left and right waist and abdomen. The regions of measurement points with large fluctuation + large amplitude type or medium fluctuation + large amplitude type that are prone to clipping are marked as the first abnormal region. The regions of measurement points with small fluctuation + large amplitude type or small fluctuation + medium amplitude type that are prone to collapse and are marked as the second abnormal region. Evaluation parameters, including the area of ​​the abnormal region and the knitting density, are extracted based on the first and second abnormal regions. It should be noted that the area of ​​the abnormal region is the total number of grid cells mapped to the mesh model, and the knitting density is obtained by calling texture features. Evaluation parameters are normalized, and evaluation tuples are constructed. An evaluation feature space is then built based on these evaluation tuples. Each evaluation tuple of an abnormal region is used as a feature point and mapped one by one into this evaluation feature space. The standard score of the evaluation tuple in the evaluation feature space for a preset period is used as the offset. Cases where the offset is greater than or equal to an offset threshold are filtered out. Based on the quadrant of the evaluation tuple's position in the evaluation feature space, the corresponding mesh update template is matched to adjust the template layer. Source of offset threshold: Based on historical statistics, offset data from past operations is collected, and statistical analysis is performed on the collected data to determine the average and standard deviation of the offsets when the offset is less than the offset threshold. The offset threshold is set according to the statistical results, and is set as the average of the offsets when the offset is less than the offset threshold plus 2 or 3 times the standard deviation. It should be noted that the value of the multiple is just an example, and the specific setting should be based on the actual situation, which will not be elaborated here. By using gap constraints and distance sequences as the analysis object, the fluctuation characteristics and amplitude characteristics of the distance sequences are identified. Through specific rectangular clustering, the first and second abnormal regions of the measurement sites are identified, and evaluation parameters are extracted to generate offsets. This enables the grid update of the template layer, achieving accurate adaptation of the measurement sites and greatly improving the fitting effect.

[0020] Each scene matching input control corresponds one-to-one with a scene matching identifier. The scene matching identifier is used to identify the scene background associated data that matches the template layer, which facilitates intelligent switching between multiple scene backgrounds in the future. In response to input operations on the input control unit, the fitting effect data group input on the input control unit is displayed.

[0021] Background switching module: Responds to the background switching action of the first virtual scene, obtains the second virtual scene, parses the scene parameters of the second virtual scene, and injects them into the simulation file; Responding to the background switching action of the first virtual scene includes: The background switching action is analyzed to obtain background semantic identifiers. It should be noted that background switching actions can be text input, voice input, or UI interaction input. Text input refers to the background switching command entered by the user in the UI text box, such as "switch to a vacation beach background." Voice input refers to the voice command recorded by the user through the microphone, such as "change my background to the beach." UI interaction input refers to the user clicking a preset scene selection button, such as "office" or "beach vacation." The obtained background switching action is bound to the current user ID and operation timestamp, encapsulated into a data packet, and stored in a temporary cache. Based on the background switching action, NLP technology is used to extract background semantic identifiers, assigning different background type codes to different scenes. NLP semantic feature extraction is a commonly used technique and will not be elaborated upon here. Based on background semantic identifiers, background type codes are extracted. The background source files of the second virtual scene are retrieved from a pre-set multi-scene background library, including visual source files and scene parameter files. The visual source files describe the sequence frames of the virtual scene background, and the scene parameter files are stored in JSON format and correspond one-to-one with the visual source files. The background source files are parsed to obtain the scene parameters of the second virtual scene. A pre-set rendering engine is called to terminate the rendering of the template layer by the first virtual scene. The parsed scene parameters of the second virtual scene are injected into the simulation file to generate the second virtual scene, completing the background switching from the first virtual scene to the second virtual scene. The scene parameters include light field vectors and spatial depth distribution. The light field vectors are used to determine the direction, intensity, and color temperature of the light source, and the spatial depth distribution is used to determine the distance distribution of objects in the scene. Obtaining a second virtual scene includes: establishing the association between scene parameters and simulation files, including lighting and shadow linkage and spatial pose association. Light and shadow linkage: Based on the light field vector, the mesh model is retrieved, and a first semantic label is created for the mesh model according to the background semantic identifier, forming a linkage between the light field vector and the mesh model; Spatial pose association: Based on spatial depth distribution, retrieve the simulated human body model, and create a second semantic label for the simulated human body model according to the background semantic identifier, forming an association and linkage between spatial depth distribution and simulated human body model; Based on the correlation relationship and combined with scene matching identifiers, for each scene background, the first correlation degree of light and shadow linkage is extracted, and the second correlation degree of spatial pose correlation is extracted. Cases where the first and second correlation degrees are greater than or equal to the standard correlation degree threshold are selected, and the scene parameters and simulation files are marked as highly correlated. All measurement points in this case are located, and cluster analysis is performed on the measurement points to generate rendering type clusters, including regular rendering ranges and core rendering ranges. The classification rendering range of the background source file for each scene type is obtained, and high-precision rendering is performed on scenes determined to be within the regular rendering range. For example: High-precision rendering: for mesh models: texture... The image resolution has been increased from 2K to 8K, clearly displaying the yarn direction and knit texture; the number of model faces has increased from 100,000 to 500,000, with no jagged edges; for the background source file: the HDR background image has been increased from 1080P to 4K, with no blurring of the background texture and complete details; low-precision rendering is performed on scenes determined to be within the core rendering range, for example: low-precision rendering: for mesh models: the texture resolution has been reduced from 2K to 1K, retaining only the basic colors and large textures, and the yarn and knit are simplified; the number of model faces has been reduced from 100,000 to 50,000, allowing slight jagged edges; for the background source file: the HDR background image has been reduced from 720P, and distant textures are blurred; The classification rendering range corresponds to the first semantic tag and the second semantic tag; through the matching and binding of background semantic tags, scene parameters and simulation files form a one-to-one association and linkage relationship, ensuring that the lighting effects and spatial poses automatically adapt to scene switching without manual intervention.

[0022] The virtual generation module responds to the AI ​​generation operation of the simulation file, generates a fitting rendering schedule file based on the fitting effect data set, and combines the rendering schedule file with the current simulation file to generate a valid fitting sample. In response to the AI ​​generation operation of the simulation file, a fitting room rendering schedule file is generated based on the fitting room effect data set. This includes: automatically parsing all fitting room effect data sets after responding to the AI ​​generation operation of the simulation file, obtaining the mesh update template corresponding to the pattern layer identifier in the fitting room effect data set; adding the scene matching identifier and category rendering range in the fitting room effect data set to the corresponding mesh update template to obtain the rendering template corresponding to the fitting room effect data set; wherein, the rendering template is used to add the scene parameters corresponding to the scene matching identifier of the simulated human body model to the mesh model corresponding to the pattern layer identifier; based on the rendering template, the fitting room rendering schedule file is obtained; it should be noted that the fitting room rendering schedule file is a preset instruction that the rendering engine can directly execute, and after AI generation, it can directly drive the rendering engine to complete the fitting room effect generation; the rendering process is a conventional technique and will not be described in detail here. Import target personnel images, which are real-life images used to generate virtual try-on effects. These are user-uploaded full-body or half-body photos. The target personnel images undergo preprocessing, including background subtraction, grayscale normalization, and pose correction. A try-on rendering schedule file containing extraction rules for multiple measurement points is used. Based on the preprocessed target personnel images, at least the pixel coordinates containing the aforementioned measurement points are extracted and applied to the template layer. Combined with valid try-on samples, a target dataset is generated. The target dataset is then input into a target neural network, which is used to train the target dataset. The target neural network can be built based on a GAN architecture, including a generator. The generator consists of an input layer and a backbone network. The input layer contains the target dataset. The backbone network uses a U-Net structure combined with a Transformer encoder to output AI virtual try-on images adapted to multiple scenes. The discriminator uses a PatchGAN structure, taking the generated AI virtual try-on images and real sample images as input, and outputting a realism score for the AI ​​virtual try-on images, set between 0 and 1. Through training iterations until the loss function is minimized, collaborative rendering of the target sweater with multiple scenes is achieved. The trained model can quickly generate personalized, high-fidelity AI virtual try-on images adapted to multiple scenes, meeting user needs.

[0023] In the application, the various formulas mentioned are all calculated by removing dimensions and taking their numerical values. The formulas are derived from the most recent real-world situation by collecting a large amount of data and conducting software simulations. The formulas are set by those skilled in the art according to the actual situation.

[0024] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0025] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0026] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An AI virtual try-on and multi-scene background intelligent switching generation system for sweaters, characterized in that, The system includes: an acquisition module for acquiring the simulation file of the target sweater; The fitting configuration module responds to the configuration input operation of the simulation file and obtains the fitting effect data set of the target sweater in the first virtual scene; wherein, the fitting effect data set includes the associated sample layer identifier and scene matching identifier; The background switching module responds to the background switching action of the first virtual scene, obtains the second virtual scene, parses the scene parameters of the second virtual scene, and injects them into the simulation file; The virtual generation module responds to the AI ​​generation operation of the simulation file, generates a fitting rendering schedule file based on the fitting effect data set, and combines the rendering schedule file with the current simulation file to generate a valid fitting sample.

2. The AI ​​virtual try-on and multi-scene background intelligent switching generation system for sweaters according to claim 1, characterized in that, Also includes: Import the target person's image, extract the pixel coordinates of the target person's image through the virtual try-on rendering schedule file and apply them to the template layer, combine with valid virtual try-on samples to generate the target dataset, train the target dataset with the target neural network, and output AI virtual try-on images.

3. The AI ​​virtual try-on and multi-scene background intelligent switching generation system for sweaters according to claim 1, characterized in that, In response to the configuration input operation of the simulation file, display the fitting effect data set of the target sweater in the first virtual scene, including: Display at least one input control unit, including an associated template layer input control and a scene-matching input control: The template layer input control has shape constraints and gap constraints, and there is a one-to-one correspondence between the template layer input control and the template layer identifier; the scene matching input control corresponds to a one-to-one correspondence between the scene matching identifier; In response to an input operation on the input control unit, display the fitting effect data set entered into the input control unit.

4. The AI ​​virtual try-on and multi-scene background intelligent switching generation system for sweaters according to claim 3, characterized in that, The simulation files also include a pre-set simulated human body model.

5. The AI ​​virtual try-on and multi-scene background intelligent switching generation system for sweaters according to claim 4, characterized in that, Shape constraints include: Extract the size parameters and texture features of the target sweater; A two-dimensional coordinate system is established, and a two-dimensional edge curve for the dimensional parameters is fitted to form a two-dimensional template plane. The two-dimensional template is discretized into a grid model composed of multiple mass points through grid division, and an initial coordinate index is assigned to each mass point to locate the corresponding measurement point in the grid model. Simultaneously, the instantaneous coordinates of the mesh model during the fitting process are monitored, the displacement vector between the instantaneous coordinates and the initial coordinate index is calculated, and the deformation deviation is generated. The conformal factors of multiple measurement points are determined by combining texture features, and linear weighted analysis is performed on the deformation deviation and conformal factors at the measurement points to generate the morphological regression gradient. The gradient threshold of ±wc% is set according to the morphological regression gradient, and the cases where the morphological regression gradient deviates from the gradient threshold are screened to adjust the template layer. Among them, the conformal factor is a dynamic quantity.

6. The AI ​​virtual try-on and multi-scene background intelligent switching generation system for sweaters according to claim 5, characterized in that, Clearance constraints include: Import a preset simulated human body model, monitor the distance between the mesh model and the simulated human body model along the first direction of the simulated human body model, and collect them into a distance sequence; where the first direction represents the normal direction; Clustering of distance sequence rectangles to identify the first and second anomalous regions; Evaluation parameters, including the area of ​​the abnormal region and the knitting density, are extracted based on the first and second abnormal regions. Based on the normalization of evaluation parameters, evaluation duones are constructed, and an evaluation feature space is constructed based on the evaluation duones. The standard score of the evaluation duotuple in the evaluation feature space for the preset period is used as the offset. Cases with offsets greater than or equal to the offset threshold are filtered out, and the corresponding mesh is matched to update the template to adjust the template layer.

7. The AI ​​virtual try-on and multi-scene background intelligent switching generation system for sweaters according to claim 1, characterized in that, Responding to the background switching action of the first virtual scene includes: Analyze the background switching action to obtain the background semantic identifier; Based on background semantic identifiers, the background source files of the second virtual scene are retrieved from a pre-defined multi-scene background library; Parse the background source file to obtain the scene parameters of the second virtual scene, and generate the second virtual scene; The scene parameters include the light field vector and the spatial depth distribution.

8. The AI ​​virtual try-on and multi-scene background intelligent switching generation system for sweaters according to claim 7, characterized in that, In obtaining the second virtual scene, the following are included: Establish the association between scene parameters and simulation files, including lighting and shadow linkage and spatial pose association: Light and shadow linkage: Based on the light field vector, the mesh model is retrieved, and a first semantic label is created for the mesh model according to the background semantic identifier, forming a linkage between the light field vector and the mesh model; Spatial pose association: Based on spatial depth distribution, retrieve the simulated human body model, and create a second semantic label for the simulated human body model according to the background semantic identifier, forming an association and linkage between spatial depth distribution and simulated human body model; Based on the association relationship and combined with the scene matching identifier, the classification rendering range of the background source file for each scene is obtained.

9. The AI ​​virtual try-on and multi-scene background intelligent switching generation system for sweaters according to claim 8, characterized in that, In response to the AI ​​generation operation of the simulation file, a fitting room rendering schedule file is generated based on the fitting room effect data set, including: Retrieve the grid update template corresponding to the pattern layer identifier in the fitting effect data group; Add the scene matching identifier and category rendering range from the fitting effect data group to the corresponding mesh update template to obtain the rendering template corresponding to the fitting effect data group; the rendering template is used to add the scene parameters of the simulated human body model corresponding to the scene matching identifier to the mesh model corresponding to the corresponding template layer identifier. Based on the rendering template, obtain the fitting room rendering schedule file.

10. The AI ​​virtual try-on and multi-scene background intelligent switching generation system for sweaters according to claim 9, characterized in that, The category rendering range corresponds to the first semantic tag and the second semantic tag.

Images