Clothing modeling-oriented fabric shearing property intelligent prediction method and device

Abstract

The invention discloses a clothing modeling-oriented fabric shearing property intelligent prediction method and device. The fabric shearing property intelligent prediction device collects an image corresponding to a detected fabric; the image is preprocessed through an intelligent fabric shearing prediction method and then input into a target segmentation model for target segmentation, a threshold value is dynamically adjusted according to the local brightness of the image in the target segmentation model for binaryzation, and the accuracy of target segmentation of the target segmentation model is improved; meanwhile, a feature embedding model is used for extracting embedded features from the segmentation result; carrying out dimensionality reduction on the embedded features by adopting principal component analysis, and fusing the embedded features after dimensionality reduction with the collected three-dimensional features; inputting the fusion result into an error correction model to obtain an error correction value; according to the error correction value and a shear stiffness theoretical value obtained through the shear stiffness mechanism model, a final shear stiffness prediction value is obtained, it is guaranteed that the physical significance of a prediction result can be explained, and the prediction precision of the shear stiffness can be improved.

Description

Technical Field

[0001] This invention belongs to the field of clothing testing technology, specifically relating to a method and device for intelligent prediction of fabric shear properties for clothing design. Background Technology

[0002] The human body surface is composed of multiple complex three-dimensional curved surfaces. To make fabric conform to the concavity, convexity, curvature, and slope of the human body's surface, it is usually necessary to use surface shaping techniques such as darts, pleats, seams, and gathers. The implementation of these techniques depends on the fabric's good shear properties. Fabrics with poor shear deformation are prone to oblique arching when subjected to external forces, thus affecting the aesthetics of the garment. Therefore, fabric shear properties play a crucial role in garment shaping. Since the 1960s, numerous scholars at home and abroad have conducted extensive research on this topic and developed various testing methods, such as the KES-F test method, the off-axis tensile test method, and the picture frame shear test method.

[0003] Although various testing methods exist for fabric shear properties, most are costly or complex. Current fabric shear property testing primarily relies on specialized equipment such as the KES-FB1-A style meter, which suffers from the following technical drawbacks: KES-type instruments are expensive, complex to operate, require professional maintenance, and are difficult to popularize among small and medium-sized apparel companies and fabric testing institutions; each test requires the preparation of samples of specific specifications, and cannot visually represent the fabric shear morphology, requiring complex data processing to obtain shear stiffness, which is time-consuming and labor-intensive. Furthermore, current shear property research only focuses on the warp and weft directions of the fabric. In reality, fabric mechanical properties exhibit significant anisotropy, with significant differences in shear properties across different directions. Moreover, in actual wear, fabrics are subjected to shear forces from various directions; therefore, focusing solely on the warp and weft shear properties is insufficient.

[0004] To address these issues, the industry urgently needs a low-cost, multi-directional, and visualized fabric shear testing technology. Simultaneously, it needs to establish an efficient oblique shear stiffness prediction model to reduce the workload of full-angle testing and improve the efficiency of fabric performance evaluation. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and device for intelligent prediction of fabric shear properties for clothing design, so as to achieve low-cost, multi-directional fabric shear property detection and high-precision oblique shear stiffness prediction.

[0006] In a first aspect, the present invention provides a method for intelligent prediction of fabric shear properties for clothing design, the method comprising:

[0007] The fabric to be tested is cut into a test sample and suspended by clamping one corner of it; the fabric image and three-dimensional features of the test sample when it is clamped are obtained.

[0008] A target segmentation model is constructed and used to segment the fabric and background in a fabric image to obtain a binarized image. The target segmentation model includes an encoder, a decoder, a bottleneck layer, and an output layer. The encoder includes multiple sub-encoding modules. The decoder includes multiple sub-decoding modules. The number of sub-encoding modules and sub-decoding modules are the same and they correspond one-to-one. The corresponding sub-encoding modules and sub-decoding modules are connected by skip connections. The output layer processes the output feature map of the last sub-decoding module through a cascaded activation function and a dual-threshold calibration module to obtain the output result of the target segmentation model. In the dual-threshold calibration module, the input feature map is divided into multiple local blocks, and the average probability value of each local block is calculated. A pixel threshold is set according to the average probability value. The part of the local block with a probability greater than or equal to the pixel threshold is regarded as the target, and the rest is regarded as the background.

[0009] Construct a feature embedding model and use it to extract embedded features from the binarized image; fuse the embedded features and 3D features to obtain a fused feature vector;

[0010] A fabric shear stiffness mechanism model was constructed, and the theoretical shear stiffness was obtained based on the fabric shear stiffness mechanism model.

[0011] An error correction model is constructed, and the fused feature vector is input into the error correction model to obtain the error correction value; the final predicted value of shear stiffness is obtained based on the theoretical shear stiffness and the error correction value.

[0012] Preferably, in the encoder, all sub-encoding modules except the last one process their output feature maps through a max pooling layer before inputting them into the next sub-encoding module; the last sub-encoding module processes its output feature map through a max pooling layer before inputting it into the bottleneck layer.

[0013] In the decoder, except for the first sub-decoding module, the input of the other sub-decoding modules is the concatenation result of the output feature map of the previous sub-decoding module processed by the deconvolution enhancement module and the output feature map of the corresponding sub-encoding module; the input of the first sub-decoding module is the concatenation result of the output feature map of the bottleneck layer processed by the deconvolution enhancement module and the output feature map of the corresponding sub-encoding module.

[0014] Preferably, the deconvolution enhancement module includes a series of deconvolution layers and an edge alignment module; the deconvolution layers are used to perform upsampling operations on the input feature map; the edge alignment module is used to offset the output feature map of the deconvolution layers according to the offset of the max pooling layer.

[0015] Preferably, both the sub-encoding module and the sub-decoding module include a first convolutional block, a channel attention module, and a second convolutional block connected in sequence; the channel attention module adopts an SE module; the first convolutional block and the second convolutional block each include a convolutional layer, a batch normalization layer, and an activation function connected in sequence.

[0016] Preferably, the feature embedding model includes a multi-layer convolutional neural network; each convolutional neural network includes multiple convolutional layers, multiple max pooling layers and a global average pooling layer connected in sequence.

[0017] Preferably, the embedded features after feature selection are fused with the three-dimensional features to obtain a fused feature vector; the feature selection method is principal component analysis.

[0018] Preferably, the three-dimensional features include overlap length, longitudinal height difference, intersection-lowest point length, front angle, and side angle; the overlap length is the length of the overlapping portion hanging down at both ends of the fabric; the longitudinal height difference is the vertical distance between the fabric intersection point and the lowest point in the side view; and the intersection-lowest point length is the horizontal distance between the fabric intersection point and the lowest point.

[0019] Preferably, the error correction model employs a genetic algorithm to optimize the backpropagation neural network.

[0020] Preferably, the error correction model uses the unknown parameters in the shear stiffness model as the optimization objective, and the fitness function is to minimize the mean square error between the predicted value and the measured value of the fabric shear stiffness mechanism model.

[0021] Secondly, the present invention provides a fabric shear property intelligent prediction device for clothing design, which is used to execute the above-mentioned fabric shear property intelligent prediction method; the fabric shear property intelligent prediction device includes a base, a support, a reference object, a fixing clip, a data acquisition module, and a shear performance prediction module; the support includes a longitudinal section for vertical fixation on the base and a transverse section fixed at the top of the longitudinal section; the fixing clip is installed on the transverse section of the support for clamping the test sample; the reference object is used for image size calibration; the data acquisition module includes an image data acquisition module and a point cloud data acquisition module; the image data acquisition module is used to acquire two-dimensional images; the point cloud data acquisition module is used to acquire point cloud data; the shear performance prediction module is used to predict the shear performance of the tested fabric based on the data acquired by the data acquisition module.

[0022] The beneficial effects of this invention are:

[0023] 1. This invention introduces a dual-threshold calibration method into the target segmentation model to assist in the division of the target and the background. By dynamically adjusting the threshold through local brightness of the image for binarization, it solves the problem that the traditional fixed threshold cannot adapt to the uneven brightness of the fabric image, which leads to misjudgment, and improves the accuracy of the target segmentation model in target segmentation.

[0024] 2. This invention introduces an edge alignment module into the decoder. The edge alignment module fine-tunes the decoded feature map according to the offset at the encoding end, and then concatenates the encoded feature map and the decoded feature map through channels. This solves the problem of feature space misalignment caused by existing skip connections, thereby achieving accurate segmentation of the segmentation edge by the target segmentation model.

[0025] 3. This invention integrates mechanistic models and data-driven models. The mechanistic model ensures that the physical meaning of the prediction results can be interpreted, while the error correction model can improve the prediction accuracy, thus solving the inherent defects of pure mechanistic models and pure data-driven models. At the same time, the testing device of this invention is low in cost, does not require professional personnel to operate, can be easily popularized by small and medium-sized enterprises, can cover shear performance testing at any angle, and is more in line with actual wearing needs.

[0026] 4. This invention is compatible with various fabrics such as cotton, linen, silk, wool, and chemical fibers, and can intuitively present the fabric cutting shape. It can be applied to multiple scenarios such as fabric research and development, clothing selection, and quality inspection. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall structure of Embodiment 1 of the present invention.

[0028] Figure 2 This is a schematic diagram of the target segmentation model in Embodiment 2 of the present invention.

[0029] Figure 3 This is a flowchart of the feature processing of the edge alignment module in Embodiment 2 of the present invention.

[0030] Figure 4 This is a schematic diagram of three-dimensional features in Embodiment 2 of the present invention.

[0031] Figure 5 This is a schematic diagram of the genetic algorithm optimization backpropagation neural network prediction process in Embodiment 2 of the present invention.

[0032] Figure 6 These are front views of a portion of the fabric in different directions.

[0033] Figure 7 This is a comparison between the actual and predicted values ​​of shear stiffness.

[0034] Reference numerals: 1. Base; 2. Support; 3. Reference object; 4. Fixing clip; 5. Test sample. Detailed Implementation

[0035] The present invention will be further described below with reference to the accompanying drawings.

[0036] Example 1

[0037] like Figure 1 As shown, a smart fabric shear property prediction device for clothing design includes a base 1, a support 2, a reference object 3, a fixing clip 4, a data acquisition module, and a shear property prediction module. The base 1 is a horizontal support structure, measuring 30cm × 30cm × 5cm, made of 6061 aluminum alloy, with a 2cm thick anti-slip rubber pad attached to the bottom to ensure no displacement during testing and to ensure device stability. The support 2 includes a longitudinal section for vertical fixation to the base 1 and a transverse section fixed to the top of the longitudinal section. The longitudinal section uses a telescopic rod structure, allowing for height adjustment of the transverse section. The fixing clip 4 is installed in the middle of the transverse section of the support 2 to hold one corner of the test sample 5, allowing the sample to hang naturally. The reference object 3 is installed on the side of the transverse section away from the longitudinal section for image size calibration. The test sample 5 is square, with a side length of 15cm to 25cm. The data acquisition module includes an image data acquisition module and a point cloud data acquisition module. The shear property prediction module is used to predict the shear properties of the tested fabric based on the data acquired by the data acquisition module.

[0038] In this embodiment, the adjustment range of the bracket 2 is 50cm~80cm; the adjustment accuracy is 1cm.

[0039] In this embodiment, the camera mount is compatible with 1 / 4-inch screw interfaces.

[0040] In this embodiment, the clamping surface of the fixing clip 4 is 2cm×2cm; the clamping force of the fixing clip 4 is 5N~10N, and a silicone pad is pasted on the inner side of the fixing clip 4 to avoid damaging the fabric and ensure stable clamping.

[0041] In this embodiment, reference object 3 is circular in shape and has a diameter of 1 cm.

[0042] In this embodiment, the size of the test sample 5 is 20cm × 20cm.

[0043] Example 2

[0044] A method for intelligent prediction of fabric shear properties for garment design employs the intelligent fabric shear prediction device described in Example 1. This method includes the following steps:

[0045] Step 1: Building the dataset

[0046] A labeled dataset containing 5,000 images of different fabrics (including cotton, linen, silk, wool, and chemical fibers), at different angles and under different lighting conditions was constructed and divided into training, validation, and test sets in an 8:1:1 ratio.

[0047] Step 2: Construct the target segmentation model

[0048] like Figure 2 As shown, the target segmentation model adopts a U-Net structure, which includes an encoder, decoder, bottleneck layer, and output layer. The encoder input is a 1920×1080×3 RGB image. The encoder includes four sub-encoding modules and four max-pooling layers (2×2, stride 2). Except for the last sub-encoding module, each sub-encoding module processes its output feature map through a max-pooling layer before inputting it into the next sub-encoding module; the last sub-encoding module processes its output feature map through a max-pooling layer before inputting it into the bottleneck layer. Except for the first sub-encoding module, the input feature map of the remaining sub-encoding modules is the result of processing the output feature map of the previous sub-encoding module through a max-pooling layer; the input feature map of the first sub-encoding module is the target image. The decoder includes four sequentially connected sub-decoding modules and four deconvolution enhancement modules. Except for the first sub-encoding module, the input of the remaining sub-encoding modules is the fusion result of the output feature map of the corresponding sub-decoding module and the output feature map of the previous sub-encoding module processed by the deconvolution enhancement modules. The input to the first sub-encoding module is the output feature map of the corresponding sub-decoding module, which is fused with the output feature map of the bottleneck layer after processing by the deconvolution enhancement module. The output layer processes the output feature map of the last sub-decoding module through a cascaded sigmoid activation function and a double threshold calibration module to obtain the output result of the target segmentation model.

[0049] The sub-encoding and sub-decoding modules have the same structure, both consisting of a first convolutional block, a channel attention module, and a second convolutional block connected in sequence. Both the first and second convolutional blocks include a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function connected in sequence.

[0050] like Figure 3 As shown, skip connections require splicing the outputs of the encoder and decoder, and max pooling and deconvolution upsampling processes can lead to misalignment of feature space locations. In fabric segmentation, yarn edges are key features (affecting subsequent yarn width measurements), and misalignment can cause blurred and jagged segmentation edges. Therefore, a deconvolution enhancement module is used to process the output feature map of the sub-encoding module. The deconvolution enhancement module includes a cascaded deconvolution layer and an edge alignment module. In the edge alignment module, based on the offset of the max pooling layer, the input feature map is fine-tuned by 1 pixel horizontally / vertically, i.e., the entire map is shifted by 1 pixel, and even-numbered rows and columns are filled in, with an error ≤ 1 pixel, thus making the segmentation sharper.

[0051] Because fabric images contain a large amount of interference (such as white background reflections on light-colored fabrics and black background dust on dark-colored fabrics), traditional convolutional blocks often misclassify this noise as fabric features. To address the weakness of traditional convolutional blocks in distinguishing between "fabric features" and "background noise," a channel attention module is added to perform "channel weight learning" on the initially extracted feature maps. This strengthens fabric-related features (such as yarn texture) and suppresses background noise (such as dust and reflections), essentially "filtering out effective features." The channel attention module employs a squeeze-and-excitation (SE) structure.

[0052] Fixed threshold binarization cannot adapt to the uneven brightness of fabric images, leading to misclassification. A dual threshold calibration module is introduced to dynamically adjust the threshold based on local image brightness, ensuring segmentation integrity. In the dual threshold calibration module, the input feature map is divided into 32×32 local blocks, and the average probability value P_avg of each local block is calculated. Based on the average probability value P_avg, a pixel threshold T = max(0.3, min(0.5, P_avg × 0.8)) is set, and the portion of the local block with a probability greater than or equal to the pixel threshold T is considered fabric, while the remaining portion is considered background.

[0053] Step 3: Training the model

[0054] The target segmentation model was trained using training and test sets. During training, the Dice loss function and Adam optimizer (learning rate 0.001, weight decay 1e-5) were used to guide the model parameter updates. The model was iterated for 200 rounds until the Dice coefficient on the validation set was ≥0.98.

[0055] Step 4: Cut the fabric to be tested into test samples and place them in a standard atmospheric environment (temperature 18-22℃, relative humidity 60-70%) for 24 hours. Take front and side views of the test samples using a camera, with the camera 50cm away from the fabric, shutter speed 1 / 200s, and ISO 100-400. Use a black background for light-colored samples and a white background for dark-colored samples to avoid ambient light interference. Perform adaptive histogram equalization on the acquired fabric images to improve image contrast; use Gaussian filtering (5×5 kernel, standard deviation 1.2) to remove image noise and ensure the stability of feature extraction. Input the processed fabric images into the trained target segmentation model to obtain binarized images.

[0056] Step 5: Construct a fusion feature vector for shear performance prediction

[0057] 5-1. Extracting Embedded Features

[0058] 5-1-1. Pretreatment

[0059] Morphological closing operations (3×3 rectangular structuring elements) are performed on the binarized image to fill the tiny holes, and connected component analysis is used to remove noise regions with an area of ​​less than 50 pixels to obtain an accurate fabric region mask.

[0060] 5-1-2. Extracting Embedded Features

[0061] A feature embedding model is constructed, taking a preprocessed binarized image (1920×1080×1) as input and outputting 128-dimensional embedded features corresponding to the binarized image. These features cover the fabric's morphological characteristics (such as contour complexity and curvature distribution) and texture characteristics (such as yarn interweaving texture). The feature embedding model consists of a six-layer convolutional neural network (CNN). Each CNN layer comprises four convolutional layers (3×3 kernels, stride 1, padding=1, ReLU activation function) connected in sequence, two max-pooling layers (2×2, stride 2), and one global average pooling layer.

[0062] 5-1-3. Feature Filtering

[0063] Principal component analysis (PCA) was used to reduce the dimensionality of the 128-dimensional embedding features, and the top 20 features with a cumulative variance contribution rate of ≥95% were retained as the final embedding features, which reduced the amount of computation and preserved the core information.

[0064] 5-2. Extracting 3D Features

[0065] like Figure 4 As shown, a laser scanner or structured light scanner (scanning accuracy ≥ 0.045 mm) was used in handheld fast mode (scanning accuracy 0.045 mm, point cloud generation speed 1.5 million points / second) to scan around the test sample 5 to obtain three-dimensional point cloud data. The point cloud data was imported into Geomagic Wrap software for global alignment (with the base plane as the XY plane and the line connecting the clamping points as the X-axis), removal of redundant point cloud data from the base, and filling of holes. Based on the actual size of the reference object, the three-dimensional features of the fabric were extracted from the three-dimensional point cloud data and the front and side views. These features are: overlap length TC (length of the overlapping part of the two ends of the fabric in the front view (cm)), longitudinal height difference TH (vertical distance between the intersection point and the lowest point of the fabric in the side view (cm)), intersection point-lowest point length TL (horizontal distance between the intersection point and the lowest point of the fabric in the side view (cm)), front angle Tα1, and side angle Tα2 (front / side angle (°) (cm) based on the three-dimensional point cloud).

[0066] 5-3. The 20-dimensional embedded feature is concatenated with the 5 three-dimensional features to obtain a 25-dimensional fused feature vector, which is used for subsequent shearing performance prediction.

[0067] Step Six: Shear Performance Prediction

[0068] 6-1. Constructing a model of fabric shearing mechanism

[0069] Based on yarn slip theory and interlacing point friction mechanics, the shear stiffness model is established as follows:

[0070]

[0071] Where G(θ) is the theoretical value of shear stiffness; E eq d is the equivalent elastic modulus of the yarn; eq ρ is the equivalent fineness of the yarn; μ is the interlacing point density; μ is the coefficient of friction at the yarn interlacing point (tested by the YG151 yarn friction meter when warp and weft yarns are in cross contact); F n The normal pressure at the interlacing point (needs optimization); θ is the test angle, i.e., the angle between the warp direction (0°) and the warp yarns; k1 represents the contribution weight of yarn structural parameters (E, d, ρ) to shear stiffness, reflecting the conversion ratio of the fabric's "inherent structural stiffness", k1 > 0; k2 represents the frictional effect at the interlacing point (μ×F). n The contribution weight to shear stiffness reflects the proportion of frictional resistance between yarns that is converted into overall shear resistance, k2 > 0.

[0072] The methods for obtaining different parameters in the shear stiffness model are as follows:

[0073] (a) Equivalent elastic modulus of yarn

[0074] The elastic modulus E of the yarn was determined using an electronic single-yarn tensile tester (such as the YG061 model). Pretension was set (0.5 cN / tex for cotton and linen yarns, 1 cN / tex for synthetic fiber yarns), and the stretching speed was 50 mm / min. Tensile tests were performed on the yarn to obtain a stress-strain curve. During the initial linear phase of the stress-strain curve, the slope was calculated, which is the elastic modulus E of the yarn. If the fabric is a blended yarn, the elastic modulus of each component yarn was tested separately, and then weighted according to the blending ratio to obtain the equivalent elastic modulus of the entire yarn. Equivalent elastic modulus E eq The calculation is as follows: The equivalent elastic modulus is the weighted arithmetic mean of the elastic moduli of the warp and weft yarns, weighted by the proportion of warp and weft yarns. It is expressed as:

[0075]

[0076] Among them, P j Warp density (threads / cm) refers to the number of warp threads per centimeter of width; E j P represents the elastic modulus of the warp yarn (GPa), obtained by measuring the slope of the initial linear segment of the warp yarn stress-strain curve using an electronic single-yarn strength meter;w Weft density (threads / cm) refers to the number of weft yarns per centimeter of length; E w The elastic modulus (GPa) of the weft yarn is tested using the same method as that of the warp yarn.

[0077] Weighted correction for special scenarios

[0078] 1. Blended fabrics (warp / weft yarns are made of different materials)

[0079] If the warp and weft yarns are made of different materials (e.g., warp yarn is polyester, weft yarn is cotton), in addition to the weighting of the number of yarns, the weighting of the material proportion can be added, and the formula is modified as follows:

[0080]

[0081] Where, ω j ω w The material weight ratio of warp and weft yarns (e.g., if polyester accounts for 60%, then ωj = 0.6) ensures that the mechanical properties of different materials are reasonably reflected.

[0082] 2. Cases where there is a significant difference in warp / weft yarn diameter.

[0083] If the diameters of the warp and weft yarns differ greatly (e.g., the warp yarns are thicker and the weft yarns are thinner), then the volume ratio is used as the weight. The volume of yarn per unit length (1cm) = cross-sectional area of ​​a single yarn × number of yarns. The cross-sectional area of ​​a single yarn, S = π × r 2 (where r is the yarn radius, which can be converted from fineness and fiber density), the formula for the equivalent elastic modulus is:

[0084]

[0085] (b) Equivalent fineness of yarn

[0086] Yarn fineness (d) is an indicator of the thickness of yarn, measured in tex (tex), which is the weight in grams of 1000m of yarn at a standard moisture regain. Equivalent yarn fineness (d) eq The calculation uses the proportion of warp and weft yarn counts as weights. Since yarn fineness (tex) characterizes the thickness of a single yarn, its contribution to the overall fabric is positively correlated with the yarn count.

[0087]

[0088] Where, d j The warp fineness (tex) refers to the standard moisture regain mass of 1000m of warp yarn; d w The fineness of the weft yarn (tex) is tested using the same method as the warp yarn.

[0089] (c) Interlacing point density

[0090] Interlacing density ρ refers to the number of interlacing points between warp and weft yarns per unit area of ​​fabric, reflecting the tightness of the fabric structure. The unit is usually interlacing points / cm². The method for obtaining this density is as follows:

[0091] First, test the warp density (P) of the fabric using a fabric density microscope (or a microscope). j He Weimi P w The unit is roots / cm. Then, calculate the interlacing point density according to the tissue:

[0092] For plain weave fabrics: each warp and weft yarn interweaves once, and the interweaving density ρ = Pj × Pw × 2 (because each warp and weft yarn in a plain weave interweaves in an "up and down" manner, the number of warp and weft yarns crossing per unit area is warp density × weft density × 2).

[0093] For twill fabrics: Calculate based on the number of cycles in the twill weave. For example, for 2 / 1 twill, the number of interlacing times within the weave cycle is 3, and the interlacing point density ρ = Pj × Pw × (number of interlacing times within the weave cycle or number of yarns in the weave cycle).

[0094] For satin fabrics: the calculation is based on the number of warp and weft yarns and the weave repeat. The core logic is "the actual number of warp and weft yarn crosses per unit area." Taking a 5-end, 3-warp satin weave as an example, the specific calculation method for the interlacing density ρ is as follows:

[0095] (1) Determine the basic parameters (measured or known)

[0096] Assuming the test involves 5 warp-faced satin fabrics with a warp density of P... j =40 threads / cm (40 warp threads per centimeter of width), weft density P w =30 threads / cm (30 weft yarns per centimeter). Weave repeat R=5, number of interlacing points within the repeat = 5 (5 warp yarns × 1 warp weave point / thread).

[0097] (2) Calculate the number of tissue cycles per unit area (1cm×1cm).

[0098] The total number of warp yarns per unit area (num) j Total number of weft yarns (num) w They are respectively:

[0099] num j =P j ×1cm=40 pieces

[0100] num w =P w ×1cm=30 pieces

[0101] Based on the total number of warp yarns (num) j Total number of weft yarns (num) wGet the number of repeats C in the warp direction j Number of repeats C in the weft direction w They are respectively:

[0102] C j = num j ÷R = 40 ÷ 5 = 8 cycles

[0103] C w = num w ÷R = 30 ÷ 5 = 6 cycles

[0104] Total number of cycles per unit area C sum for:

[0105] C sum =C j ×C w = 8 × 6 = 48

[0106] (3) Calculate the total number of interlacing points per unit area.

[0107] There are 5 interlacing points within each weave cycle (5×5 warp and weft yarn grid), therefore: the total number of interlacing points in a 1cm×1cm area = the total number of cycles × the number of interlacing points per cycle = 48×5=240.

[0108] 6-2. Constructing the dataset

[0109] One hundred fabrics with different structural parameters (covering different yarn fineness, interlacing density, and elastic modulus) were selected, and their shear stiffness at different angles, such as 0°, 45°, and 90°, was measured using a KES-FB1-A style meter and used as labels.

[0110] 6-3. Constructing the objective function

[0111] For each fabric i, define the prediction error Error. i for:

[0112]

[0113]

[0114]

[0115] in, This is the theoretical value of shear stiffness; This represents the measured value of shear stiffness.

[0116] 6-4. The unknown parameters k1, k2, μ, and F in the shear stiffness model... nAs the optimization objective of the GA-BP neural network (genetic algorithm optimized backpropagation neural network), the parameter search range is μ∈[0.1,0.8], F n ∈[0.5N,5N]; where k1 and k2 are globally shared scaling factors, reflecting the dimensional transformation relationship between fabric structural stiffness and friction enhancement effect; μ and F n These are individual parameters for each type of fabric. The input layer of the GA-BP neural network consists of fabric structural parameters (E...). eq d eq The output layer has unknown parameters (k1, k2, μ, F). n The hidden layers are set to 2 layers (15 nodes in the first layer and 10 nodes in the second layer), and the activation function is the Sigmoid function.

[0117] 6-5. The initial weights and thresholds of the BP neural network are optimized using a genetic algorithm. The population size is 50, the crossover probability is 0.6, the mutation probability is 0.01, and the number of iterations is 100. The fitness function is "minimum mean square error between the predicted value and the measured value of the mechanism model".

[0118] 6-6. Optimize k1, k2, μ, F n Substituting into the shear stiffness model, we obtain the optimized shear stiffness model G0'.

[0119] Step 7: Construct a hybrid prediction model

[0120] Using a 25-dimensional fused feature vector as input, a GA-BP neural network error correction model is constructed, such as... Figure 5 As shown, the error ΔG between the predicted and measured values ​​from the output mechanism model is represented. The final predicted shear stiffness value G = G0' + ΔG retains the theoretical interpretability of the mechanism model while correcting errors through data-driven modeling.

[0121] In this embodiment, the model training parameters are as follows: GA-BP neural network input layer nodes 25, hidden layer nodes 20, output layer nodes 1; learning rate 0.01, momentum factor 0.9, number of iterations 1000, error threshold 0.001; genetic algorithm population size 50, crossover probability 0.6, mutation probability 0.01, number of iterations 100.

[0122] In this embodiment, the prediction accuracy is: model determination coefficient R² ≥ 0.96, and relative error ≤ 3.0%.

[0123] Step 8: Method Evaluation

[0124] To explore general patterns, 18 common fabrics were selected, which differ in raw materials, weave, density, etc., including common fabrics such as cotton, linen, silk, and wool. Their basic specifications are shown in Table 1.

[0125] Table 1 Fabric Specifications

[0126]

[0127] According to FZ / T 01113-2012 "Test Method for Shear Properties of Fabrics with Small Deformation", the fabrics were tested in a standard atmospheric environment. Fabrics from Table 1 were cut into 20 cm × 20 cm samples along 0° (warp), 15°, 30°, 45°, 60°, 75°, 90° (weft), 105°, 120°, 135°, 150°, and 165° (since 0° and 180° are identical, 180° was not tested). The shear stiffness of the fabric at different angles and the shear hysteresis moments (i.e., 2HG and 2HG5) at 0.5° and 5° were tested using a KES-FB1-A shear strength tester. Three samples were cut for each fabric type, and the tests were conducted three times. The average value was taken, and the results are shown in Table 2. The elastic modulus E of the warp and weft yarns of the fabric was measured using an electronic single-yarn strength tester (such as the YG061 model), and E was calculated using the method described above. eq and measure d eq and ρ.

[0128] Table 2 Shear stiffness of fabrics at different angles (unit: gf·(cm·deg)) -1 )

[0129]

[0130] Various fabrics were tested using this invention, and front views of some fabrics are shown below. Figure 6 As shown. After preprocessing, the front view was segmented using a target segmentation model (Dice coefficients ≥ 0.98) and embedded features were extracted. Principal component analysis (PCA) was used to reduce the dimensionality of the embedded features, retaining the first 20 dimensions. Point cloud data of the front view of the fabric at different angles using the one-point suspension method were obtained using an EinScan Pro 2X Plus 3D scanner. After processing with Geomagic Wrap software, five 3D features were extracted: overlap length TC, longitudinal height difference TH, length TL between intersection point and lowest point, front angle Tα1, and side angle Tα2.

[0131] E eq d eq Substituting ρ, θ, and the shear stiffness G measured by KES at each 15° into the mechanism model The unknown parameters k1, k2, μ, and F of the mechanism model were optimized using a GA-BP neural network. n The optimized k1, k2, μ, and F nSubstituting into the shear stiffness mechanism model, we obtain the theoretical value G0'.

[0132] Using a 25-dimensional fused feature vector as input, the GA-BP error correction model outputs an error correction value ΔG; the final predicted shear stiffness value is G = G0' + ΔG. Taking fabrics #17 and #18 as examples, the actual and theoretical shear stiffness values ​​at 12 angles are as follows: Figure 7 As shown. From Figure 7 It can be seen that the theoretical value of shear stiffness obtained by using the present invention has good consistency with the actual value measured by KES.

Claims

1. A method for intelligent prediction of fabric shear properties for clothing design, characterized in that: The method includes: The fabric to be tested is cut into a test sample and suspended by clamping one corner of it; the fabric image and three-dimensional features of the test sample when it is clamped are obtained. A target segmentation model is constructed and used to separate the fabric and background in a fabric image, resulting in a binarized image. The target segmentation model includes an encoder, a decoder, a bottleneck layer, and an output layer. The encoder includes multiple sub-encoding modules, and the decoder includes multiple sub-decoding modules. The number of sub-encoding modules and sub-decoding modules are the same and they correspond one-to-one. The corresponding sub-encoding modules and sub-decoding modules are connected by skip connections. The output layer processes the output feature map of the last sub-decoding module through a cascaded activation function and a dual-threshold calibration module to obtain the output result of the target segmentation model. In the dual-threshold calibration module, the input feature map is divided into multiple local blocks, and the average probability value of each local block is calculated. A pixel threshold is set according to the average probability value, and the part of the local block with a probability greater than or equal to the pixel threshold is regarded as the target, and the rest is regarded as the background. Construct a feature embedding model and use it to extract embedded features from the binarized image; fuse the embedded features and 3D features to obtain a fused feature vector; A fabric shear stiffness mechanism model was constructed, and the theoretical shear stiffness was obtained based on the fabric shear stiffness mechanism model. An error correction model is constructed, and the fused feature vector is input into the error correction model to obtain the error correction value; the final predicted value of shear stiffness is obtained based on the theoretical shear stiffness and the error correction value.

2. The intelligent prediction method for fabric shear properties for clothing design according to claim 1, characterized in that: In the encoder, all sub-encoding modules except the last one process their output feature maps through a max pooling layer before inputting them into the next sub-encoding module; the last sub-encoding module processes its output feature map through a max pooling layer before inputting it into the bottleneck layer. In the decoder, except for the first sub-decoding module, the input of the other sub-decoding modules is the concatenation result of the output feature map of the previous sub-decoding module processed by the deconvolution enhancement module and the output feature map of the corresponding sub-encoding module; the input of the first sub-decoding module is the concatenation result of the output feature map of the bottleneck layer processed by the deconvolution enhancement module and the output feature map of the corresponding sub-encoding module.

3. The intelligent prediction method for fabric shear properties for clothing design according to claim 2, characterized in that: The deconvolution enhancement module includes a series of deconvolution layers and an edge alignment module; the deconvolution layers are used to perform upsampling operations on the input feature map; the edge alignment module is used to offset the output feature map of the deconvolution layers according to the offset of the max pooling layer.

4. The intelligent prediction method for fabric shear properties for clothing design according to claim 1, characterized in that: Both the sub-encoding module and the sub-decoding module include a first convolutional block, a channel attention module, and a second convolutional block connected in sequence; the channel attention module adopts an SE module; the first convolutional block and the second convolutional block each include a convolutional layer, a batch normalization layer, and an activation function connected in sequence.

5. The intelligent prediction method for fabric shear properties for clothing design according to claim 1, characterized in that: The feature embedding model includes a multi-layer convolutional neural network; each convolutional neural network includes multiple convolutional layers, multiple max pooling layers and a global average pooling layer connected in sequence.

6. The intelligent prediction method for fabric shear properties for clothing design according to claim 1, characterized in that: The embedded features after feature selection are fused with the three-dimensional features to obtain a fused feature vector; the feature selection method is principal component analysis.

7. The intelligent prediction method for fabric shear properties for clothing design according to claim 1, characterized in that: The three-dimensional features include overlap length, longitudinal height difference, intersection-lowest point length, front angle, and side angle; the overlap length is the length of the overlapping portion hanging down at both ends of the fabric; the longitudinal height difference is the vertical distance between the fabric intersection point and the lowest point in the side view; the intersection-lowest point length is the horizontal distance between the fabric intersection point and the lowest point.

8. The intelligent prediction method for fabric shear properties for clothing design according to claim 1, characterized in that: The error correction model uses a genetic algorithm to optimize the backpropagation neural network.

9. The intelligent prediction method for fabric shear properties for clothing design according to claim 8, characterized in that: The error correction model uses the unknown parameters in the shear stiffness model as the optimization objective, and the fitness function is to minimize the mean square error between the predicted and measured values ​​of the fabric shear stiffness mechanism model.

10. A smart fabric shear prediction device for clothing design, characterized in that: The device is used to perform a smart fabric shear property prediction method for clothing design as described in claim 1. The smart fabric shear property prediction device includes a base (1), a support (2), a reference object (3), a fixing clip (4), a data acquisition module, and a shear property prediction module. The support (2) includes a longitudinal section for vertical fixation on the base (1) and a transverse section fixed at the top of the longitudinal section. The fixing clip (4) is installed on the transverse section of the support (2) and is used to hold the test sample (5). The reference object (3) is used for image size calibration. The data acquisition module includes an image data acquisition module and a point cloud data acquisition module. The image data acquisition module is used to acquire two-dimensional images. The point cloud data acquisition module is used to acquire point cloud data. The shear property prediction module is used to predict the shear property of the tested fabric based on the data acquired by the data acquisition module.

Images