Garment quality intelligent detection method in garment processing process

Abstract

The invention relates to the technical field of artificial intelligence and image recognition, and discloses a garment quality intelligent detection method in a garment processing process. The method comprises the following steps: acquiring three-dimensional point cloud data of a target human body and constructing a parameterized human body grid model; obtaining a two-dimensional garment cutting piece digital template, and generating an initial three-dimensional garment model through virtual stitching; simulating dynamic deformation under excitation of gravity, inertia force and respiratory movement in a physical simulation engine to obtain a final clothing model in a steady wearing state; calculating a local gap distribution field with the human body model, and extracting gap mean values, variances and gradient features of key areas such as shoulders, a chest and a waist; and inputting the features into a pre-trained multi-layer perceptron model, and outputting a 0-100-score stereo fit degree quantitative score. According to the method, real person try-on is not needed, high-precision, automatic and repeatable fit degree evaluation is realized, and the garment quality inspection efficiency and the model optimization capability are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and image recognition technology, specifically relating to an intelligent method for detecting garment quality during garment processing. Background Technology

[0002] With the deep penetration of intelligent manufacturing and artificial intelligence technologies into the textile and apparel industry, quality inspection in the garment production process is rapidly evolving from traditional manual visual inspection to automation and intelligence. Current mainstream visual inspection systems are mostly based on high-resolution two-dimensional images or fixed-viewpoint three-dimensional point cloud data, possessing a certain ability to identify static defects such as surface flaws, seam misalignment, and color differences. However, as a flexible product conforming to the curves of the human body, one of the core quality indicators of clothing—three-dimensional fit—essentially depends on the dynamic deformation behavior of the fabric in multi-dimensional space under wearing conditions, including complex mechanical responses such as shoulder sagging, waist and abdomen wrinkles, and armhole tightness. Two-dimensional images, lacking depth information and pose freedom, cannot recreate the spatial topology of clothing when actually worn; while existing three-dimensional scanning technologies are usually performed on static suspension or standard mannequins, ignoring the fabric stretching, slippage, and rebound effects caused by human movement, resulting in a serious disconnect between inspection results and actual wearing experience.

[0003] Intelligent assessment of clothing fit urgently requires the integration of physical perception and spatial modeling capabilities. Fit involves not only static size matching but also dynamic adaptability, that is, the ability of clothing to maintain a comfortable fit without excessive restriction or redundant wrinkles during human activity. This characteristic is highly dependent on the coupling effect of multiple factors such as fabric elastic modulus, shear stiffness, and pattern cutting method, and is difficult to accurately quantify through isolated geometric measurements or empirical rules.

[0004] In existing technologies, some solutions attempt to introduce pressure-sensing clothing or inertial motion capture systems to assist in evaluation. However, the former is costly and interferes with normal wearing conditions, while the latter can only acquire human motion data and cannot simultaneously reflect the deformation response of the clothing itself. Other methods utilize finite element simulation to model clothing deformation, but these are difficult to integrate into real-time production line inspection processes due to difficulties in material parameter calibration and high computational complexity. More critically, current inspection systems generally separate the two dimensions of "visual morphology" and "physical deformation": 3D vision can capture external geometry but lacks internal stress distribution; flexible electronics can record local strain but lack global spatial semantics. This data silo phenomenon prevents the system from constructing a complete deformation mapping of clothing in dynamic wearing scenarios, making it difficult to accurately identify potential fit problems caused by pattern design defects or cutting errors. Therefore, in the clothing processing process, there is an urgent need for an intelligent inspection method that can simultaneously integrate dynamic deformation perception and 3D visual reconstruction to achieve high-fidelity, quantifiable, and scenario-based evaluation of three-dimensional fit. Summary of the Invention

[0005] This invention provides an intelligent method for detecting garment quality during garment processing, aiming to solve the technical problem that the three-dimensional fit of garments is difficult to evaluate using two-dimensional images. This method constructs a multi-scale geometric matching relationship between a three-dimensional dynamic human body model and a digital garment pattern, and combines this with a physical simulation engine to perform high-fidelity simulation of the deformation behavior of garments under actual wearing conditions. This allows for an objective, quantitative, and automated evaluation of the three-dimensional fit of garments without the need for real-person fitting.

[0006] The intelligent garment quality detection method in the garment processing process of the present invention includes the following steps: acquiring three-dimensional scanning point cloud data of the target human body; constructing a parameterized three-dimensional human body mesh model based on the three-dimensional scanning point cloud data; acquiring two-dimensional pattern data of the garment to be inspected; converting the two-dimensional pattern data of the garment to be inspected into an initial three-dimensional garment mesh model through a virtual sewing algorithm; loading the initial three-dimensional garment mesh model into a physical simulation engine and placing it on top of the three-dimensional human body mesh model; In the physical simulation engine, a preset gravity field, inertial force field, and breathing motion excitation are applied to drive the initial three-dimensional clothing mesh model to perform dynamic deformation simulation, generating the final three-dimensional clothing mesh model in a steady-state wearing state; the local gap distribution field between the final three-dimensional clothing mesh model and the human body three-dimensional mesh model is calculated; based on the local gap distribution field, the gap mean, gap variance, and gap gradient features of multiple preset key regions are extracted; the gap mean, gap variance, and gap gradient features are input into a pre-trained fit scoring model, and the three-dimensional fit measurement score of the clothing on the target human body is output.

[0007] Furthermore, the acquisition of the three-dimensional scanning point cloud data of the target human body specifically includes: using a multi-view structured light scanning device to perform a surround scan of the entire human body to obtain a depth image sequence of no less than 8 views; performing point cloud registration and fusion processing on the depth image sequence to generate a complete and denoised three-dimensional point cloud dataset; the spatial resolution of the three-dimensional point cloud dataset is no less than 0.5 mm, and the point density is no less than 400 points per square centimeter.

[0008] Furthermore, the construction of a parameterized human 3D mesh model based on the 3D scanned point cloud data specifically includes: performing Poisson surface reconstruction on the 3D point cloud dataset to generate an initial closed surface mesh; using the Laplacian smoothing algorithm to perform geometric optimization on the initial closed surface mesh to eliminate high-frequency noise; performing non-rigid registration between the optimized mesh model and the standard human template mesh to establish vertex correspondence; and based on the vertex correspondence, extracting key human body size parameters, including shoulder width, chest circumference, waist circumference, hip circumference, arm length, leg length, and neck circumference, to form a parameterized human body model vector.

[0009] Furthermore, the acquisition of the two-dimensional pattern data of the garment to be inspected specifically includes: exporting a vector graphic file containing all the outlines, seam lines, dart lines and alignment marks of the garment from the garment computer-aided design system; the vector graphic file is expressed using a unified coordinate system, and the logical connection relationship between each garment is established through the seam line identifier; the outline of each garment is defined by no less than 100 ordered control points, and the distance between adjacent control points does not exceed 2 mm.

[0010] Furthermore, the process of converting the two-dimensional pattern data into an initial three-dimensional garment mesh model using a virtual sewing algorithm specifically includes: assigning an initial planar position to each two-dimensional pattern piece, making its normal vector parallel to the Z-axis of the world coordinate system; identifying the pairs of edges to be sewn based on the seam line identifiers between the pattern pieces; performing an edge alignment operation on each pair of edges to be sewn, so that the corresponding control points coincide in three-dimensional space; applying tension constraints to the overall structure after sewing using a spring-mass physical model, so that each pattern piece naturally unfolds and initially fits onto the outer surface of the human body three-dimensional mesh model, forming an initial three-dimensional garment mesh model; the number of vertices in the initial three-dimensional garment mesh model is on the same order of magnitude as the number of vertices in the human body three-dimensional mesh model, and each garment vertex is associated with its original pattern piece source information.

[0011] Furthermore, the physical simulation engine adopts a position constraint solution framework based on position dynamics, with a time step set to 1 millisecond; the clothing material properties defined in the physical simulation engine include Young's modulus, Poisson's ratio, shear modulus, areal density, and damping coefficient; the Young's modulus ranges from 10 kPa to 500 kPa, the Poisson's ratio ranges from 0.2 to 0.4, and the areal density ranges from 100 g / m² to 300 g / m². The gravitational field strength is set to The direction is along the negative Z-axis of the world coordinate system; the inertial force field simulates walking by applying sinusoidal periodic acceleration excitation to the center of the torso of the human model, with a frequency of 1.5 Hz and an amplitude of [missing value]. The respiratory motion excitation is achieved by applying radial periodic displacement to the grid vertices of the thoracic region, with a respiratory cycle of 4 seconds and a maximum radial displacement of 10 millimeters.

[0012] Furthermore, the generation of the final three-dimensional clothing mesh model under steady-state wearing conditions specifically includes: running the physical simulation engine until the system energy converges, and determining the convergence condition as the maximum displacement change of the clothing mesh vertices within 500 consecutive time steps being less than 0.1 mm; recording the coordinates of the clothing mesh vertices at the convergence moment to form the final three-dimensional clothing mesh model; and performing topological consistency verification on the final three-dimensional clothing mesh model to ensure that there are no self-intersections, tears, or penetrations.

[0013] Furthermore, the calculation of the local gap distribution field between the final 3D clothing mesh model and the human body 3D mesh model specifically includes: for each vertex on the human body 3D mesh model, calculating its shortest Euclidean distance to the surface of the final 3D clothing mesh model; the calculation of the shortest Euclidean distance adopts an accelerated nearest point search algorithm, which is implemented based on an octree space partitioning structure; mapping the shortest Euclidean distance values ​​of all vertices back to the human body surface to form a continuous gap scalar field; the gap scalar field is stored in the form of vertex attributes of the human body mesh.

[0014] Furthermore, the preset key areas include the shoulder area, chest area, waist area, hip area, armhole area, and crotch area; the shoulder area is defined as a band-shaped area extending 5 cm outward and 3 cm inward from the line connecting the left and right acromion points; the chest area is defined as a circular area with a radius of 10 cm centered on the nipple point; the waist area is defined as an annular area extending 8 cm upward and downward from the navel point; the hip area is defined as an elliptical area with a radius of 12 cm centered on the midpoint of the gluteal cleft; the armhole area is defined as a spherical projection area with a radius of 6 cm around the armpit point; and the crotch area is defined as a rectangular area extending 10 cm forward and backward and 5 cm left and right around the perineum point.

[0015] Furthermore, the extraction of the gap mean, gap variance, and gap gradient features of multiple preset key regions specifically includes: for all vertices within each preset key region, calculating the arithmetic mean of their gap values ​​as the gap mean of that region; calculating the standard deviation of the gap values ​​within that region as the gap variance; performing surface gradient calculation on the gap scalar field to obtain the gradient vector at each vertex; calculating the average value of the gradient vector magnitude within that region as the gap gradient feature; the surface gradient calculation is implemented using the cotangent weight formula in discrete differential geometry.

[0016] Furthermore, the pre-trained fit scoring model is a multilayer perceptron neural network with 18 nodes in the input layer, corresponding to the mean gap, variance gap, and gradient features of the gaps in the six key regions. The hidden layer consists of two layers with 32 and 16 nodes respectively, and the activation function is a modified linear unit. The output layer is a single node with an output value range of 0 to 100, representing the three-dimensional fit score. The training dataset of the fit scoring model consists of no less than 500 sets of real-person fitting experiment data. Each set of data includes the three-dimensional scan results of the same garment on different body types, expert fit scores, and corresponding gap feature vectors. The expert fit scores are obtained by taking the average of the scores given independently by three senior garment craftsmen. The scoring criteria are formulated based on the national standard "Clothing Size" and industry fit evaluation specifications.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention abandons the indirect evaluation mode that relies on two-dimensional images to infer fit, and for the first time constructs a full-link digital simulation and evaluation system from two-dimensional cut pieces to three-dimensional dynamic wearing status. Through high-precision human body modeling, physically realistic clothing deformation simulation, and quantitative analysis based on key area gap features, it achieves objective, repeatable, and high-precision automated detection of clothing three-dimensional fit.

[0018] This method eliminates the need for real-person fitting, significantly shortening the quality inspection cycle and reducing labor costs. Simultaneously, its output fit score is a continuous numerical value, accurately reflecting the garment's adaptability to different body types, providing reliable data support for garment pattern optimization, personalized customization, and intelligent manufacturing. Furthermore, the gap distribution field analysis method employed in this invention can precisely locate issues of excessive looseness or tightness in key areas such as the shoulders, chest, waist, and hips, guiding process improvements and significantly enhancing the wearing comfort and aesthetics of finished garments. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the fit evaluation based on multi-scale geometric matching and physical simulation driven by a three-dimensional dynamic human body model and a digital clothing pattern in this invention. Figure 3 This is a flowchart illustrating the key modeling stages of this invention: constructing a parameterized 3D human body mesh model from 3D point cloud data and generating an initial 3D clothing mesh model from 2D pattern pieces. Figure 4 This is a simulation process framework diagram of the physical simulation engine driving the dynamic deformation simulation of clothing to generate a three-dimensional clothing model in a steady-state wearing state in this invention. Figure 5 This is a flowchart of the quantitative analysis logic for calculating the local gap distribution field and extracting the gap features of key areas in this invention. Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the terminal system and the cloud-based integration scoring model in this invention. Detailed Implementation

[0020] This invention provides an intelligent method for detecting garment quality during garment manufacturing, aiming to solve the technical problem that the three-dimensional fit of garments is difficult to evaluate using two-dimensional images. This method constructs a multi-scale geometric matching relationship between a three-dimensional dynamic human body model and a digital garment pattern, and combines this with a physical simulation engine to perform high-fidelity simulation of the deformation behavior of garments under actual wearing conditions. This allows for an objective, quantitative, and automated evaluation of the three-dimensional fit of garments without the need for real-person fitting. The following will be illustrated in conjunction with the appendix... Figure 1 To be continued Figure 6 This section provides a detailed implementation description of each functional module of the system, expanding upon it layer by layer.

[0021] The intelligent garment quality detection method during garment processing includes the following steps: acquiring three-dimensional scanning point cloud data of the target human body; constructing a parameterized three-dimensional human body mesh model based on the three-dimensional scanning point cloud data; acquiring two-dimensional pattern data of the garment to be inspected; converting the two-dimensional pattern data of the garment into an initial three-dimensional garment mesh model through a virtual stitching algorithm; loading the initial three-dimensional garment mesh model into a physical simulation engine and placing it on top of the three-dimensional human body mesh model. In the physical simulation engine, a preset gravity field, inertial force field, and breathing motion excitation are applied to drive the initial three-dimensional clothing mesh model to perform dynamic deformation simulation, generating the final three-dimensional clothing mesh model in a steady-state wearing state; the local gap distribution field between the final three-dimensional clothing mesh model and the human body three-dimensional mesh model is calculated; based on the local gap distribution field, the gap mean, gap variance, and gap gradient features of multiple preset key regions are extracted; the gap mean, gap variance, and gap gradient features are input into a pre-trained fit scoring model, and the three-dimensional fit measurement score of the clothing on the target human body is output.

[0022] The acquisition of 3D scanning point cloud data of the target human body specifically includes: using a multi-view structured light scanning device to perform a surround scan of the entire human body to obtain a depth image sequence from no less than 8 views; performing point cloud registration and fusion processing on the depth image sequence to generate a complete and denoised 3D point cloud dataset; the spatial resolution of the 3D point cloud dataset is no less than 0.5 mm, and the point density is no less than 400 points per square centimeter. The point cloud registration process uses an iterative nearest-point algorithm, using the depth image from one view as the reference coordinate system, and aligning the point cloud data from the other views to this reference coordinate system through a rigid body transformation matrix; the point cloud fusion uses a weighted average strategy, performing position averaging and normal vector consistency checks on points in the overlapping area, and removing outliers caused by occlusion or reflection; the denoising process uses a bilateral filter to suppress high-frequency noise while preserving geometric edge features, ensuring that the point cloud surface is smooth and the details are complete.

[0023] The specific steps of constructing a parameterized 3D human body mesh model based on the 3D scanned point cloud data include: performing Poisson surface reconstruction on the 3D point cloud dataset to generate an initial closed surface mesh; using the Laplacian smoothing algorithm to perform geometric optimization on the initial closed surface mesh to eliminate high-frequency noise; performing non-rigid registration between the optimized mesh model and the standard human body template mesh to establish vertex correspondence; and extracting key human body size parameters, including shoulder width, chest circumference, waist circumference, hip circumference, arm length, leg length, and neck circumference, based on the vertex correspondence to form a parameterized human body model vector.

[0024] Poisson surface reconstruction transforms point clouds and their normal vectors into an implicit scalar field by solving the Poisson equation. Then, a traveling cubes algorithm is used to extract isosurfaces, generating a topologically consistent triangular mesh. Laplace smoothing moves each vertex to the weighted average position of its neighboring vertices in each iteration, with the weights determined by the reciprocal of the side length. The number of iterations is controlled to within 10 to avoid over-smoothing. Non-rigid registration employs thin-plate spline interpolation, using the control points of the standard template mesh as the source points and the corresponding anatomical landmarks of the optimized mesh as the target points to construct a global deformation field, ensuring precise alignment of the two meshes on the anatomical structure. Key dimensional parameters are obtained by locating standard anatomical landmarks on the registered mesh and calculating their Euclidean distance or perimeter. All parameters are stored as floating-point numbers in the parameterized human model vector for subsequent body type classification and adaptation analysis.

[0025] The acquisition of the two-dimensional digital pattern data of the garment to be inspected specifically includes: exporting a vector graphic file containing all the outlines, seam lines, dart lines, and alignment markers of the garment pieces from the garment computer-aided design system; the vector graphic file uses a unified coordinate system, and the pieces are logically connected through seam line identifiers; the outline of each piece is defined by no less than 100 ordered control points, with the distance between adjacent control points not exceeding 2 mm. The vector graphic file format is Extensible Markup Language (XML), and each piece is represented by an independent path element. The path node contains coordinates, type (corner point or curve control point), and connection relationship attributes; the seam line identifier is embedded in the metadata of the boundary nodes of the piece in string form, and boundary nodes with the same identifier will be forcibly aligned in the subsequent virtual sewing stage; the dart lines are represented by dashed paths, and their start and end points are respectively associated with two alignment markers inside the piece, used to guide the folding operation during three-dimensional unfolding; the alignment markers are marked with circular symbols, and their coordinate accuracy is retained to three decimal places in millimeters to ensure that the seam alignment error is less than 0.1 mm.

[0026] The process of converting the two-dimensional pattern data into an initial three-dimensional garment mesh model using a virtual sewing algorithm specifically includes: assigning an initial planar position to each two-dimensional pattern piece, making its normal vector parallel to the Z-axis of the world coordinate system; identifying the edge pairs to be sewn based on the seam line identifiers between the pattern pieces; performing an edge alignment operation on each pair of edges to be sewn, so that the corresponding control points coincide in three-dimensional space; applying tension constraints to the overall structure after sewing using a spring-mass physical model, so that each pattern piece naturally unfolds and initially fits onto the outer surface of the human body three-dimensional mesh model, forming an initial three-dimensional garment mesh model; the number of vertices in the initial three-dimensional garment mesh model is on the same order of magnitude as the number of vertices in the human body three-dimensional mesh model, and each garment vertex is associated with its original pattern piece source information.

[0027] The initial planar position allocation follows the garment manufacturing layout rules. The main pattern pieces, such as the front, back, and sleeve pieces, are arranged sequentially along the X-axis, with a spacing of no less than 20% of the maximum width of the pattern piece to avoid collisions. The edge alignment operation first resamples the control points of the two seam edges to ensure that their number is equal and their arc length parameterization is consistent. Then, the optimal rigid body transformation is solved using the least squares method to minimize the point-to-point distance between the two boundaries in three-dimensional space. The spring-mass model discretizes each pattern piece into a mass network, with the mass points connected by linear springs. The original length of the spring is equal to the side length of the pattern piece in the two-dimensional state, and the spring constant is set to 100 Newtons per meter. Tension constraints are achieved by applying radial contraction forces around the garment mesh. The magnitude of the force is proportional to the surface normal vector of the human body mesh, driving the garment mesh to move closer to the human body surface. The bonding process adopts a projection mapping strategy, projecting the garment mass points in the opposite direction of the surface normal to the nearest human body vertex and recording the projection distance as the initial gap value. The pattern piece source information is stored in the custom attribute field of each garment vertex in the form of integer tags for subsequent material attribute allocation and fault tracing.

[0028] The physical simulation engine employs a position constraint solution framework based on position dynamics, with a time step set to 1 millisecond. The clothing material properties defined in the physical simulation engine include Young's modulus, Poisson's ratio, shear modulus, surface density, and damping coefficient. The Young's modulus ranges from 10 kPa to 500 kPa, the Poisson's ratio ranges from 0.2 to 0.4, and the surface density ranges from 100 g / m² to 300 g / m². The gravitational field strength is set to... The direction is along the negative Z-axis of the world coordinate system; the inertial force field simulates walking by applying sinusoidal periodic acceleration excitation to the center of the torso of the human model, with a frequency of 1.5 Hz and an amplitude of [missing value]. The respiratory motion excitation is achieved by applying radial periodic displacement to the grid vertices of the thoracic region, with a respiratory cycle of 4 seconds and a maximum radial displacement of 10 millimeters.

[0029] The positional dynamics framework significantly improves numerical stability by iteratively solving positional constraints instead of force equilibrium equations. Three constraint relaxation iterations are performed within each time step, with constraint types including distance constraints (maintaining the internal structure of the garment pieces), bending constraints (resisting curvature changes), and collision constraints (preventing penetration of the human body). Material properties are loaded from a preset material library based on the garment piece source information; different garment pieces can be assigned different properties to simulate blended or lining structures. A gravitational field acts as a globally constant acceleration on all garment mass points. An inertial force field is achieved by modifying the motion state of the human body model; the acceleration at the center point of the torso is updated according to a sine function, causing the entire human body mesh to sway periodically. Respiratory motion excitation is achieved by defining an ellipsoidal influence domain in the chest cavity region. The displacement of vertices within the domain decays radially, the displacement direction is along the local normal vector, and the displacement function is a cosine waveform, ensuring a smooth and continuous breathing process.

[0030] The process of generating the final three-dimensional clothing mesh model under steady-state wearing conditions specifically includes: running the physics simulation engine until the system energy converges, and determining the convergence condition as the maximum displacement change of the clothing mesh vertices within 500 consecutive time steps being less than 0.1 mm; recording the coordinates of the clothing mesh vertices at the convergence moment to form the final three-dimensional clothing mesh model; and performing topological consistency verification on the final three-dimensional clothing mesh model to ensure that there are no self-intersections, tears, or penetrations.

[0031] System energy is defined as the sum of kinetic and potential energy. Kinetic energy is calculated from the particle velocity, and potential energy includes elastic potential energy and gravitational potential energy. Convergence detection is performed at the end of each time step, maintaining a sliding window of length 500 and recording the maximum displacement change within the window. Once the convergence condition is met, the simulation loop is terminated immediately. The vertex coordinates of the final model are stored as double-precision floating-point numbers, retaining micron-level precision. Topology verification uses ray casting to detect self-intersections, emitting rays from each edge and counting the number of intersections with other faces; if the number is greater than one, it is marked as a self-intersection. Penetration detection is achieved by calculating the signed distance from the clothing vertex to the human body mesh; a negative value indicates penetration, requiring a rollback to the previous valid state and adjustment of collision constraint parameters for resimulation.

[0032] The calculation of the local gap distribution field between the final 3D clothing mesh model and the human body 3D mesh model specifically includes: for each vertex on the human body 3D mesh model, calculating its shortest Euclidean distance to the surface of the final 3D clothing mesh model; the calculation of the shortest Euclidean distance adopts an accelerated nearest point search algorithm, which is implemented based on an octree space partitioning structure; mapping the shortest Euclidean distance values ​​of all vertices back to the human body surface to form a continuous gap scalar field; the gap scalar field is stored in the form of vertex attributes of the human body mesh.

[0033] The octree is constructed with the clothing mesh bounding box as the root node, recursively segmented to leaf nodes containing no more than 8 triangular faces; the nearest point search starts from the root node, prioritizing the traversal of the child nodes closest to the human body vertex, and using the triangle-to-point distance formula to calculate the precise distance; to improve efficiency, only human body vertices located within the clothing bounding box are fully searched, and external vertices are directly assigned a preset maximum gap value of 50 mm; the gap scalar field is smoothed by Gaussian filtering, and the kernel radius is set to twice the average side length of the human body mesh to eliminate local noise; all gap values ​​are stored in millimeters in the vertex color channel or custom attributes of the human body mesh for easy visualization and subsequent analysis.

[0034] The preset key areas include the shoulder area, chest area, waist area, hip area, armhole area, and crotch area; the shoulder area is defined as a band-shaped area extending 5 cm outward and 3 cm inward from the line connecting the left and right acromion points; the chest area is defined as a circular area with a radius of 10 cm centered on the nipple point; the waist area is defined as an annular area extending 8 cm upward and downward from the navel point; the hip area is defined as an elliptical area with a radius of 12 cm centered on the midpoint of the gluteal cleft; the armhole area is defined as a spherical projection area with a radius of 6 cm around the armpit point; and the crotch area is defined as a rectangular area extending 10 cm forward and backward and 5 cm left and right around the perineum point.

[0035] Anatomical landmarks are automatically located by pre-setting offsets on the parametric human model. The acromion is located 5 mm below the lateral end of the clavicle, the nipple is located on the midline of the fourth intercostal space, the umbilicus is located 3 cm above the midpoint of the line connecting the iliac crests, the midpoint of the gluteal cleft is located at the level of the lower edge of the sacrum, the axillary point is located at the junction of the inner side of the upper arm and the trunk, and the perineum is located below the pubic symphysis. Region extraction uses Boolean operations in spherical or cylindrical coordinate systems to filter human mesh vertices into corresponding regions according to their distance and angle from the landmarks. Each region independently stores a vertex index list to ensure that subsequent feature calculations do not interfere with each other.

[0036] The extraction of the gap mean, gap variance, and gap gradient features of multiple preset key regions specifically includes: for all vertices within each preset key region, calculating the arithmetic mean of their gap values ​​as the gap mean of that region; calculating the standard deviation of the gap values ​​within that region as the gap variance; performing surface gradient calculations on the gap scalar field to obtain the gradient vector at each vertex; and calculating the average magnitude of the gradient vector magnitude within that region as the gap gradient feature. The surface gradient calculation is implemented using the cotangent weight formula in discrete differential geometry. The gap mean reflects the overall tightness of the region; a small value indicates excessive tightness, and a large value indicates excessive looseness. The gap variance characterizes the uniformity of the gap distribution; a high variance indicates local wrinkles or stretching. The surface gradient vector points in the direction of the fastest gap growth, and its magnitude reflects the degree of drastic gap change; high gradient values ​​often appear at the end of a provincial highway or at the turning point of a suture line. The cotangent weight formula uses the diagonal cotangent value of the first-order neighborhood triangle as the weight for each vertex, and weights the gap difference vector of the neighborhood vertices to obtain the gradient estimate of that vertex. All feature values ​​are normalized and scaled to the interval of 0 to 1 to eliminate the influence of dimensions.

[0037] The pre-trained fit scoring model is a multilayer perceptron neural network. Its input layer has 18 nodes, corresponding to the mean gap, variance gap, and gradient features of the gaps in six key regions. The hidden layer consists of two layers, with 32 and 16 nodes respectively, and the activation function is a modified linear unit. The output layer is a single node with an output value range of 0 to 100, representing the 3D fit score. The training dataset for the fit scoring model consists of no fewer than 500 sets of real-person fitting experiment data. Each set of data includes the 3D scan results of the same garment on different body types, expert fit scores, and corresponding gap feature vectors. The expert fit scores are obtained by averaging the scores independently given by three senior garment craftsmen, and the scoring criteria are based on the national standard "Clothing Size" and industry fit evaluation specifications.

[0038] The neural network uses the mean squared error loss function, the optimizer is the adaptive moment estimate, the initial learning rate is 0.001, and the batch size is 32. The training process implements an early stopping strategy, terminating training when the validation set loss does not decrease for 20 consecutive rounds. The input feature vectors are arranged in regional order to ensure that the model learns the dependencies between regions. The output score is mapped to the 0 to 100 range using a sigmoid function to ensure numerical stability. Cross-validation is performed before model deployment, and the average coefficient of determination of the five-fold validation is not less than 0.85 to ensure generalization ability.

[0039] The above-described methods and steps constitute the core technical solution of this invention. Through a closed-loop process of high-precision modeling, physical simulation, and quantitative feature analysis, it achieves automated, objective, and refined evaluation of garment fit. This method does not rely on real-person fitting and is applicable to quality control throughout the entire garment design, pattern making, and production chain, significantly improving testing efficiency and consistency.

Claims

1. A method for intelligent detection of garment quality during garment processing, characterized in that, include: Acquire 3D scan point cloud data of the target human body; A parameterized 3D human body mesh model is constructed based on the 3D scan point cloud data. Obtain the digital pattern data of the two-dimensional cut pieces of the garment to be inspected; The two-dimensional cut piece digital pattern data is converted into an initial three-dimensional garment mesh model using a virtual sewing algorithm; The initial 3D clothing mesh model is loaded into the physical simulation engine and then placed on top of the human body 3D mesh model; In the physical simulation engine, a preset gravity field, inertial force field, and breathing motion excitation are applied to drive the initial three-dimensional clothing mesh model to perform dynamic deformation simulation and generate the final three-dimensional clothing mesh model in a steady-state wearing state. Calculate the local gap distribution field between the final 3D clothing mesh model and the 3D human body mesh model; Based on the local gap distribution field, the gap mean, gap variance, and gap gradient features of multiple preset key regions are extracted. The gap mean, gap variance, and gap gradient features are input into a pre-trained fit rating model, which outputs a three-dimensional fit metric score for the garment on the target human body.

2. The intelligent garment quality detection method during garment processing according to claim 1, characterized in that, The acquisition of the three-dimensional scan point cloud data of the target human body includes: A multi-view structured light scanning device is used to perform a surround scan of the entire human body to obtain a depth image sequence with no less than 8 views. The depth image sequence is subjected to point cloud registration and fusion processing to generate a complete and denoised 3D point cloud dataset. The spatial resolution of the three-dimensional point cloud dataset is no less than 0.5 mm, and the point density is no less than 400 points per square centimeter.

3. The intelligent garment quality detection method during garment processing according to claim 2, characterized in that, The construction of a parameterized 3D human body mesh model based on the 3D scan point cloud data includes: Poisson surface reconstruction is performed on the three-dimensional point cloud dataset to generate an initial closed surface mesh; The initial closed surface mesh is geometrically optimized using the Laplace smoothing algorithm to eliminate high-frequency noise. The optimized mesh model is non-rigidly registered with the standard human body template mesh to establish vertex correspondence. Based on the vertex correspondence, key human body size parameters are extracted, including shoulder width, chest circumference, waist circumference, hip circumference, arm length, leg length, and neck circumference, forming a parameterized human body model vector.

4. The intelligent garment quality detection method during garment processing according to claim 3, characterized in that, The process of acquiring the two-dimensional cut pattern data of the garment to be inspected includes: Export a vector graphic file containing all pattern outlines, seam lines, dart lines, and alignment marks from the garment computer-aided design system; The vector graphics file uses a unified coordinate system, and logical connections are established between each piece of fabric through seam line identifiers. The outline of each piece is defined by no fewer than 100 ordered control points, with the distance between adjacent control points not exceeding 2 millimeters.

5. The intelligent garment quality detection method during garment processing according to claim 4, characterized in that, The process of converting the two-dimensional pattern data into an initial three-dimensional garment mesh model using a virtual sewing algorithm includes: Assign an initial planar position to each 2D pattern piece, such that its normal vector is parallel to the Z-axis of the world coordinate system; Identify the pairs of edges that need to be sewn together based on the seam markings between the fabric pieces; For each pair of edges that need to be stitched, perform an edge alignment operation to make the corresponding control points coincide in three-dimensional space; A spring-mass physical model is used to apply tension constraints to the overall structure after sewing, so that each piece of fabric unfolds naturally and initially fits the outer surface of the human body three-dimensional mesh model, forming an initial three-dimensional clothing mesh model. The number of vertices in the initial three-dimensional clothing mesh model is on the same order of magnitude as the number of vertices in the human body three-dimensional mesh model, and each clothing vertex is associated with its original cut piece source information.

6. The intelligent garment quality detection method during garment processing according to claim 5, characterized in that, The physical simulation engine adopts a position constraint solution framework based on position dynamics, with a time step set to 1 millisecond. The physical simulation engine defines clothing material properties including Young's modulus, Poisson's ratio, shear modulus, surface density, and damping coefficient. The Young's modulus ranges from 10 kPa to 500 kPa, the Poisson's ratio ranges from 0.2 to 0.4, and the areal density ranges from 100 g / m² to 300 g / m². The gravitational field strength is set to The direction is along the negative Z-axis of the world coordinate system; The inertial force field simulates walking by applying sinusoidal periodic acceleration excitation to the center of the torso of the human model, with a frequency of 1.5 Hz and an amplitude of 0.3 Hz. ; The respiratory motion excitation is achieved by applying radial periodic displacement to the grid vertices of the thoracic region, with a respiratory cycle of 4 seconds and a maximum radial displacement of 10 millimeters.

7. The intelligent garment quality detection method during garment processing according to claim 6, characterized in that, The generation of the final 3D clothing mesh model under steady-state wearing conditions includes: Run the physics simulation engine until the system energy converges. The convergence condition is that the maximum displacement change of the clothing mesh vertices is less than 0.1 mm within 500 consecutive time steps. Record the vertex coordinates of the clothing mesh at the convergence moment to form the final 3D clothing mesh model; The final three-dimensional clothing mesh model is subjected to topological consistency verification to ensure that there are no self-intersections, tears, or penetrations.

8. The intelligent garment quality detection method during garment processing according to claim 7, characterized in that, The calculation of the local gap distribution field between the final 3D clothing mesh model and the 3D human body mesh model includes: For each vertex on the human body 3D mesh model, calculate the shortest Euclidean distance from it to the surface of the final 3D clothing mesh model; The calculation of the shortest Euclidean distance adopts an accelerated nearest point search algorithm, which is implemented based on an octree space partitioning structure; Map the shortest Euclidean distance values ​​of all vertices back to the human body surface to form a continuous gap scalar field; The gap scalar field is stored in the form of vertex attributes of the human body mesh.

9. The intelligent garment quality detection method during garment processing according to claim 8, characterized in that, The preset key areas include the shoulder area, chest area, waist area, hip area, armhole area, and crotch area. The shoulder area is defined as a strip-shaped area extending 5 cm outward and 3 cm inward from the line connecting the left and right acromion points; The chest area is defined as a circular area with a radius of 10 centimeters centered on the nipple point; The waist area is defined as a ring-shaped area extending 8 centimeters above and below the navel. The buttock region is defined as an elliptical region with a radius of 12 centimeters centered at the midpoint of the gluteal cleft; The armhole area is defined as a spherical projection area with a radius of 6 cm around the armpit point; The crotch area is defined as a rectangular area extending 10 cm forward and backward and 5 cm to the left and right around the perineum.

10. The intelligent garment quality detection method during garment processing according to claim 9, characterized in that, The extraction of the gap mean, gap variance, and gap gradient features of multiple preset key regions includes: For all vertices within each preset key region, calculate the arithmetic mean of their gap values, and use this as the average gap value for that region. Calculate the standard deviation of the gap values ​​within this region, which is used as the gap variance; Perform surface gradient calculations on the gap scalar field to obtain the gradient vector at each vertex; Calculate the average magnitude of the gradient vector within this region, and use it as the gap gradient feature; The surface gradient calculation is implemented using the cotangent weight formula in discrete differential geometry.

Images