Textile pattern dislocation identification method

Abstract

The invention provides a textile pattern dislocation identification method, which comprises the following steps of: performing geometric transformation analysis on extracted pattern contour information, calculating a relative position relationship of each part of a pattern, establishing a pattern moisture deformation model, and judging whether the pattern is dislocated or deformed by comparing the difference between the current pattern and a reference model; on the basis of the corrected pattern, adding a special waterproof mark on the surface of the wet textile, manufacturing a microscopic mark which does not influence the fabric performance, and capturing the position of the waterproof mark through infrared imaging equipment to serve as an auxiliary reference point for pattern positioning; and based on the waterproof mark position of the pattern, tracking the motion trail of the waterproof fabric, calculating the image distortion degree according to the motion vector field, correcting the moving waterproof mark fabric image in real time, and eliminating the influence of motion blur on waterproof mark identification.

Description

Technical Field

[0001] The invention relates to the field of information technology, and in particular to a method for identifying misalignment of textile patterns. Background Art

[0002] In the textile production process, quality control after pattern printing and dyeing is a key link. However, when textiles are exposed to water or in a humid environment, the pattern is prone to dislocation and deformation, which brings challenges to quality inspection. At the same time, the texture and diverse patterns on the surface of textiles increase the difficulty of recognition. In addition, the high-speed movement of fabrics on the production line also creates obstacles for real-time detection. To solve this problem, it is necessary to develop a dislocation recognition technology that can work stably under humid conditions. This technology should be able to overcome the influence of moisture on the pattern morphology and maintain recognition accuracy when the fabric moves quickly. At the same time, it is also necessary to consider how to add reliable waterproof marks without affecting the original performance and appearance of the fabric. Another technical difficulty is how to achieve rapid segmentation and recognition of large-area complex patterns with limited computing resources. While ensuring the recognition accuracy, meet the real-time requirements of the production line. In addition, how to effectively integrate the information of the segmented sub-regions to obtain the dislocation status of the overall pattern is also a problem to be solved. Summary of the invention

[0003] The present invention provides a method for identifying misalignment of textile patterns, which mainly comprises:

[0004] The textile images in a humid environment are preprocessed, and the pattern features are extracted. A reference model of the normal morphology of the pattern is established according to the pre-trained data set, and the pattern area and the background area are separated from the original image to obtain the pattern contour information.

[0005] Perform geometric transformation analysis on the extracted pattern contour information, calculate the relative position relationship of each part of the pattern, establish a pattern moisture deformation model, and determine whether the pattern is dislocated or deformed by comparing the difference between the current pattern and the reference model;

[0006] If the pattern is found to be misaligned or deformed, a mathematical model of the effect of moisture on the pattern is established based on the material characteristics of the textile and the production process parameters. By simulating the change rules of the pattern under different humidity conditions, the test results are dynamically corrected to eliminate the interference of moisture on the recognition accuracy.

[0007] Based on the corrected pattern, special waterproof marks are added to the surface of wet textiles to make microscopic marks that do not affect the performance of the fabric. The positions of the waterproof marks are captured by infrared imaging equipment as auxiliary reference points for pattern positioning.

[0008] Based on the waterproof mark position of the pattern, the movement trajectory of the waterproof fabric is tracked, the degree of image distortion is calculated according to the motion vector field, and the image of the waterproof mark fabric in motion is corrected in real time to eliminate the influence of motion blur on the waterproof mark recognition;

[0009] Based on the correction of fast-moving textile images, the textile texture is segmented, and the waterproof mark is used as the seed point to expand to the surrounding area to determine the texture boundary. The large area pattern is divided into multiple sub-regions. At the same time, feature extraction and dislocation analysis are performed on each sub-region. The analysis results of each sub-region are summarized to evaluate the dislocation recognition status of the overall textile pattern.

[0010] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0011] The present invention discloses a method for identifying pattern misalignment of textiles. The method establishes a reference model by extracting pattern features, separates the pattern from the background, and obtains contour information. A geometric transformation analysis is performed on the contour, a moisture deformation model is established, and the pattern misalignment or deformation is determined. In combination with material properties and process parameters, a moisture influence model is established for dynamic correction. A waterproof mark is added to the surface of the fabric, and the mark position is captured by infrared imaging as an auxiliary positioning. The movement trajectory of the waterproof fabric is tracked, and the moving image is corrected in real time. The texture is segmented, the boundary is determined using the waterproof mark as a seed point, and the sub-areas are divided and analyzed. The present invention solves the problem of identifying pattern misalignment of textiles under humid conditions, improves the recognition accuracy, and can be applied to fields such as textile quality control. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 The present invention is a flowchart of a method for identifying misalignment of a textile pattern.

[0013] Figure 2 It is a schematic diagram of a textile pattern misalignment recognition method of the present invention.

[0014] Figure 3 This is another schematic diagram of a textile pattern misalignment recognition method according to the present invention. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0016] like Figure 1 -3. In this embodiment, a method for identifying misalignment of a textile pattern may specifically include:

[0017] Step S101 , preprocessing the textile image in a humid environment, extracting pattern features, establishing a reference model of the normal form of the pattern according to a pre-trained data set, separating the pattern area and the background area from the original image, and obtaining pattern contour information.

[0018] The original image of the textile in a humid environment is obtained, and the dark channel prior defogging algorithm is used to eliminate the blur effect of the original image to obtain the defogged image; the defogged image is subjected to adaptive histogram equalization and guided filtering processing, and the optimal threshold is calculated according to the OTSU algorithm for binarization processing to obtain the pre-processed binary image; the pre-processed binary image is subjected to morphological operations, the noise area is removed by opening operation, the internal holes of the pattern are filled by closing operation, the pattern contour is extracted by the Canny edge detection algorithm, and the SIFT feature descriptor of the pattern is extracted by the scale-invariant feature transformation algorithm; according to the pre-trained normal morphological reference model, the SIFT feature is compared with the model feature, and the cosine similarity between the feature vectors is calculated; if the cosine similarity is higher than the preset threshold, it is determined to be a normal pattern, otherwise it is determined to be an abnormal area; the normal pattern area is subjected to regional growth segmentation, the pixel point with the largest pattern edge gradient is selected as the seed point, the color Euclidean distance and gradient direction difference of adjacent pixels are calculated, and if both are less than the preset threshold, the pixel is included in the growth area until no new pixel meets the growth condition or the preset maximum number of iterations is reached, and the separated pattern area and background area are obtained.

[0019] Specifically, the original image of textiles in a humid environment is obtained. The dark channel prior defogging algorithm is used to eliminate the blur effect under the influence of water vapor. The image contrast is adjusted by adaptive histogram equalization. The guided filter is used to remove image noise and retain edge information. The optimal threshold is automatically calculated according to the OTSU algorithm for image binarization to obtain the preprocessed binary image. Morphological operations are performed on the preprocessed binary image. The opening operation is used to remove small noise areas, and the closing operation is used to fill small holes inside the pattern. The pattern contour is extracted by the Canny edge detection algorithm. The local feature descriptor of the pattern is extracted using the scale-invariant feature transform (SIFT) algorithm to obtain the geometric and texture feature description of the pattern. According to the pre-trained normal morphological reference model, the extracted SIFT features are compared with the model features, the cosine similarity between the feature vectors is calculated, and the similarity threshold is set to determine the degree of feature matching. If the matching degree is higher than the threshold, it is considered to be a normal pattern, otherwise it is marked as an abnormal area. The pattern area marked as normal is segmented by region growing. The pixel with the largest pattern edge gradient is selected as the seed point. The color Euclidean distance and gradient direction difference of adjacent pixels are calculated. If both are less than the preset threshold, the pixel is included in the growth region until no new pixel meets the growth condition or the preset maximum number of iterations is reached, and finally the separated pattern area and background area are obtained. After obtaining the original image of the textile in a humid environment, the dark channel prior dehazing algorithm is first applied to eliminate the influence of water vapor. The algorithm is based on the observed phenomenon that the intensity of at least one color channel in most local areas is very low. By estimating the atmospheric illumination and transmittance, a clear image can be restored. Then, adaptive histogram equalization is used to adjust the image contrast. The image is divided into several 8x8 pixel blocks. Each block is histogram equalized, and then the results are merged using bilinear interpolation to avoid block effects. Then, guided filtering is used for denoising. The filter uses a guide image to retain edge information. The filter radius is set to 3 and the regularization parameter is set to 0.1. Finally, the OTSU algorithm is used to automatically calculate the optimal threshold for binarization. The algorithm finds the optimal threshold by maximizing the inter-class variance and divides the image into foreground and background. Morphological operations are performed on the binary image obtained, and 3x3 rectangular structural elements are used for opening and closing operations to remove small noise areas and fill small internal holes, respectively. Then the Canny edge detection algorithm is used to extract the pattern contour, setting the low threshold to 50 and the high threshold to 150. Then the SIFT algorithm is used to extract local feature descriptors. The SIFT algorithm is invariant to scale and rotation. By constructing a scale space, locating key points, assigning directions, and generating feature descriptors, a 128-dimensional feature vector is obtained. The extracted SIFT features are compared with the pre-trained normal morphology reference model, and the cosine similarity between the feature vectors is calculated. The similarity threshold is set to 0.8. If the matching degree is higher than this threshold, it is considered to be a normal pattern, otherwise it is marked as an abnormal area.Finally, the normal pattern area is segmented by region growing. The pixel with the largest pattern edge gradient is selected as the seed point. The Euclidean distance and gradient direction difference of adjacent pixels are calculated. The color distance threshold is set to 10 and the gradient direction difference threshold is set to 30 degrees. If both are less than the preset threshold, the pixel is included in the growth area. The maximum number of iterations is set to 1000, and finally the separated pattern area and background area are obtained.

[0020] Step S102 , performing geometric transformation analysis on the extracted pattern contour information, calculating the relative position relationship of each part of the pattern, establishing a pattern moisture deformation model, and determining whether the pattern is dislocated or deformed by comparing the difference between the current pattern and the reference model.

[0021] The pattern contour information is obtained, the Harris corner point detection operator is used to identify the feature points in the pattern, and the stable feature points are screened out according to the preset corner point response threshold and the non-maximum suppression window size; the Euclidean distance and angle between the stable feature points are calculated, the geometric structure description of the pattern is constructed, and the relative position relationship of each part of the pattern is obtained; for the stable feature points, an elastic grid deformation model is established, the pattern is divided into grids of a preset size, and the displacement of the internal points of the grid is calculated by bicubic interpolation to obtain a nonlinear deformation field; the geometric structure description of the current pattern is compared with that of the reference model, the local feature descriptor is extracted by the accelerated robust feature algorithm, and the feature matching is performed using the nearest neighbor ratio method. If the Euclidean distance difference between the matching point pairs exceeds the preset threshold, it is determined that the area is misaligned; the thin plate spline interpolation algorithm is used to correct the misaligned area, and the matching point pairs that have not been deformed are selected as control points, a radial basis function is constructed, and the interpolation coefficient is solved. The misaligned area is resampled to obtain the corrected pattern contour information.

[0022] Specifically, based on the extracted pattern contour information, the Harris corner detection operator is used to identify the feature points in the pattern. The corner response threshold is set to 0.01, the non-maximum suppression window size is 3x3 pixels, and stable feature points are screened out. The Euclidean distance and angle between the feature points are calculated to construct the geometric structure description of the pattern and obtain the relative position relationship of each part of the pattern. For the identified feature points, an elastic grid deformation model is established, and the pattern is divided into a 16x16 grid, with each grid node as a control point. The displacement of the internal points of the grid is calculated using bicubic interpolation to obtain a nonlinear deformation field. By minimizing the objective function of deformation energy and feature point matching error, the optimal deformation parameters are solved to obtain the pattern deformation model under a humid environment. The geometric structure description of the current pattern is compared with that of the reference model, and the accelerated robust feature (SURF) algorithm is used to extract local feature descriptors. The nearest neighbor ratio method is used for feature matching, and the matching threshold is set to 0.7. The Euclidean distance difference between the matching point pairs is calculated. If it exceeds 5 pixels, the area is judged to be misaligned. At the same time, the affine transformation matrix of the matching point pairs is calculated. If the ratio of the singular values ​​of the transformation matrix exceeds 1.5, the pattern is judged to be deformed. According to the dislocation and deformation judgment results, combined with the moisture deformation model, the actual deformation degree of each part of the pattern is calculated. The thin plate spline interpolation algorithm is used to correct the deformed area, and the matching point pairs that have not been deformed are selected as control points to construct the radial basis function. The interpolation coefficient is solved, and the deformed area is resampled to obtain the corrected pattern contour information for subsequent pattern recognition and quality assessment. Starting from the pattern contour information, the Harris corner detection operator is used to identify feature points, and the corner response threshold is set to 0.01 to ensure that the detected corner points have sufficient significance. A 3x3 pixel non-maximum suppression window is used to screen out the local optimal corner points, usually 100-200 stable feature points. The Euclidean distance and angle between these feature points are calculated to construct a geometric structure description matrix. The matrix size is n×n, where n is the number of feature points, and each element contains the distance and angle information between two feature points. Based on these feature points, a 16x16 elastic mesh deformation model is established with a total of 256 control points. Bicubic interpolation is used to calculate the displacement of the internal points of the mesh, and the interpolation kernel size is 4x4. By minimizing the objective function of deformation energy and feature point matching error, the gradient descent method is used for iterative optimization, the learning rate is set to 0.01, and the maximum number of iterations is 1000, and the nonlinear pattern deformation model under humid environment is obtained. When comparing the current pattern with the reference model, the SURF algorithm is used to extract local feature descriptors, the Hessian matrix threshold is set to 300, and a 64-dimensional feature vector is extracted. The nearest neighbor ratio method is used for feature matching, and the matching threshold is set to 0.7, which usually obtains 50-100 pairs of matching points. The Euclidean distance difference between the matching point pairs is calculated. If it exceeds 5 pixels, the area is judged to be misaligned.At the same time, the RANSAC algorithm is used to estimate the affine transformation matrix of the matching point pairs, and the number of iterations is set to 1000, and the internal point threshold is 2 pixels. If the ratio of the singular values ​​of the transformation matrix exceeds 1.5, the pattern is judged to be deformed. According to the misalignment and deformation judgment results, combined with the nonlinear deformation model obtained above, the actual deformation degree of each part of the pattern is calculated. Finally, the thin plate spline interpolation algorithm is used for correction, and at least 4 matching point pairs that have not been deformed are selected as control points to construct the radial basis function f(r) = r2log(r). Solve the interpolation coefficient, resample the deformed area, and the grid size is twice that of the original image to obtain the corrected pattern contour information. This process is achieved through iterative optimization, setting the maximum number of iterations to 100 and the convergence threshold to 0.001 pixels.

[0023] Step S103, if the pattern is found to be misaligned or deformed, a mathematical model of the influence of moisture on the pattern is established according to the material characteristics of the textile and the production process parameters. By simulating the change law of the pattern under different humidity conditions, the detection result is dynamically corrected to eliminate the interference of moisture on the recognition accuracy.

[0024] The material water absorption parameters and environmental humidity values ​​are obtained, and a multi-layer perceptron neural network is used to establish a nonlinear mapping relationship between the material water absorption rate, environmental humidity and pattern deformation degree; a mathematical model of the influence of moisture on the pattern is constructed based on the nonlinear mapping relationship. Real-time humidity data is collected using an environmental humidity sensor, and a time series prediction model is constructed in combination with the dynamic moisture absorption characteristics of the material; the time series prediction model uses a long short-term memory network to process the humidity data sequence and predict the change of moisture content inside the material over time. The predicted moisture content is input into the multi-layer perceptron neural network to calculate the theoretical deformation state of the pattern under the current humidity conditions; the theoretical deformation state is compared with the actual detected pattern misalignment or deformation, and the deviation matrix between the two is calculated. A state space model is constructed, in which the state variable is the deviation of the geometric features of the pattern, and the observed variable is the actual detection result; the Kalman filter algorithm is applied for dynamic estimation and correction, and the process noise covariance matrix and the observation noise covariance matrix are obtained by offline estimation of historical data. The pattern geometric features are corrected using the filtered state estimation values; the corrected pattern geometric features are matched with the original reference model, and the affine transformation matrix is ​​calculated using an image registration algorithm based on mutual information; the mutual information objective function is optimized using a gradient descent method, and the transformation parameters are iteratively updated until convergence or the maximum number of iterations is reached.

[0025] Specifically, according to the material characteristics and production process parameters of textiles, a mathematical model of the influence of moisture on patterns is constructed, and a multi-layer perceptron neural network is used to establish a nonlinear mapping relationship between material water absorption, environmental humidity and pattern deformation. The network is trained by experimental data. The input layer contains material parameters and humidity values, the hidden layer uses the ReLU activation function, and the output layer predicts the change in the geometric features of the pattern. The real-time humidity data is collected by the environmental humidity sensor, and the time series prediction model is constructed in combination with the dynamic moisture absorption characteristics of the material. The humidity data sequence is processed using the long short-term memory network (LSTM) to predict the change of the moisture content inside the material over time. The predicted moisture content is input into the multi-layer perceptron trained in the previous step to calculate the theoretical deformation state of the pattern under the current humidity conditions. The theoretical deformation state is compared with the actual detected pattern misalignment or deformation, and the deviation matrix between the two is calculated. A state space model is constructed, in which the state variable is the deviation of the geometric features of the pattern, and the observed variable is the actual detection result. The Kalman filter algorithm is applied for dynamic estimation and correction, and the process noise covariance matrix and the observation noise covariance matrix are obtained by offline estimation of historical data. The geometric features of the pattern are corrected using the filtered state estimation value. The corrected pattern geometric features are matched with the original reference model, and the affine transformation matrix is ​​calculated using the mutual information-based image registration algorithm. The mutual information objective function is optimized using the gradient descent method, and the transformation parameters are iteratively updated until convergence or the maximum number of iterations is reached. Finally, the inverse matrix of the obtained affine transformation matrix is ​​applied to the detection results for reverse transformation to eliminate the interference of moisture on the pattern recognition accuracy and obtain the corrected pattern recognition results. When constructing a mathematical model of the effect of moisture on the pattern, a multilayer perceptron neural network with 3 hidden layers is used, each containing 64, 32 and 16 neurons respectively. The input layer receives 5 material parameters (such as fiber type, fabric density) and 1 humidity value, and the output layer predicts the deformation and rotation angle of the pattern in the x and y directions. The Adam optimizer is used with a learning rate of 0.001 and 500 rounds of training on 10,000 samples. The environmental humidity sensor collects data every 5 seconds, and an LSTM network with 12 time steps is constructed with a hidden layer size of 32 to predict the change in moisture content inside the material in the next hour. The predicted moisture content sequence is input into the multi-layer perceptron to obtain 24 sets of theoretical deformation states. The actual detection uses image matching based on SURF features to extract 200 feature points and calculate the Euclidean distance deviation from the theoretical state. A 6-dimensional state space model (x, y displacement, rotation angle, and its rate of change) is constructed, and the observed variables are 4-dimensional (x, y displacement, rotation angle, area change rate). In the Kalman filter, the diagonal elements of the process noise covariance matrix are set to 0.01, and the diagonal elements of the observation noise covariance matrix are set to 0.1. The Kalman filter is iterated 50 times for each frame of the image to obtain the corrected pattern geometric features.Image registration uses a mutual information-based method, using 32×32 pixel image blocks to calculate mutual information, and applying the L-BFGS-B algorithm to optimize affine transformation parameters. The maximum number of iterations is set to 100, and the convergence threshold is 1e-6. Finally, the inverse matrix of the 3×3 affine transformation matrix is ​​applied to the detection result to perform a reverse transformation to obtain the corrected pattern recognition result. The entire process is implemented on the GPU, and the single frame processing time is controlled within 50 milliseconds.

[0026] According to different fabric material characteristics and production process parameters, a mathematical model of the influence of moisture on patterns is constructed. This model is used to quantitatively describe the geometric distortion effect of changes in moisture content on the fabric surface pattern, so as to predict the distortion degree and distortion direction parameters. While identifying the fabric surface pattern, the recognition result is subjected to real-time geometric correction based on the predicted pattern distortion degree and distortion direction.

[0027] A database of fabric material characteristics and production process parameters is constructed, and a mapping relationship between moisture content and pattern geometric distortion is established by using a fully connected neural network, wherein the fully connected neural network includes an input layer, a hidden layer and an output layer; a capacitive sensor array is used to collect moisture content data on the fabric surface, and the moisture content data is used as an input of a long short-term memory network; a moisture content prediction sequence is obtained according to the output of the long short-term memory network, and the moisture content prediction sequence is input into the fully connected neural network; the output of the fully connected neural network is obtained to obtain a prediction value sequence of the degree and direction of pattern distortion; a scale-invariant feature conversion algorithm is used to extract actual pattern feature points and descriptors on the fabric surface, and a time-varying thin plate spline transformation matrix is ​​constructed according to the prediction value sequence of the degree and direction of pattern distortion; if the time-varying thin plate spline transformation matrix is ​​constructed, non-rigid correction is performed on the actual pattern feature points; a non-rigid registration algorithm based on mutual information is used to compare the corrected pattern with a standard pattern; and the transformation parameters are optimized by a gradient descent method to obtain a pattern recognition result.

[0028] Specifically, a database was constructed based on the fabric material characteristics and production process parameters. A five-layer fully connected neural network was used to establish the mapping relationship between moisture content and pattern geometric distortion. The input layer contained 10 material parameters and 1 moisture content. The hidden layers were 64, 32, and 16 neurons, respectively. The ReLU activation function was used. The output layer predicted 3 distortion degree parameters and 2 direction parameters. The network was trained by the stochastic gradient descent method, and the learning rate was set to 0. The moisture content data on the fabric surface was collected in real time using a capacitive sensor array, and the environmental parameters were obtained by combining the temperature and humidity sensors. The moisture content change trend in the next 30 minutes was predicted by the long short-term memory network (LSTM). The LSTM network contained 2 hidden layers, each with 64 units. The predicted moisture content sequence was input into the trained fully connected neural network to obtain the predicted value sequence of the pattern distortion degree and direction. At the same time, the scale-invariant feature transform (SIFT) algorithm was used to extract the actual pattern feature points and descriptors on the fabric surface. Based on the predicted distortion parameter sequence, a time-varying thin plate spline transformation matrix was constructed, and the extracted SIFT feature points were subjected to real-time non-rigid correction. During the correction process, the fast approximate nearest neighbor algorithm was used to accelerate the feature point matching. The corrected pattern was compared with the standard pattern using a non-rigid registration algorithm based on mutual information. The transformation parameters were optimized by the gradient descent method. The number of iterations was set to 100 to obtain the final pattern recognition result. In the implementation process, a database containing 1,000 combinations of fabric materials and process parameters was first constructed, with 20,000 samples of distorted data for each combination at 20 different moisture contents. A five-layer fully connected neural network was used, with 11 neurons in the input layer, 64, 32, and 16 neurons in the three hidden layers, and 5 neurons in the output layer. The Adam optimizer was used with a learning rate of 0.001, 500 rounds of training on 16,000 training samples, and 4,000 samples in the validation set. An 8x8 capacitive sensor array was arranged on the fabric surface, with each sensor covering an area of ​​1 cm2 and a sampling frequency of 10 Hz. The sampling frequency of the ambient temperature and humidity sensor was 1 Hz. The LSTM network received data from the past 10 minutes, predicted the moisture content for the next 30 minutes, and updated the prediction results every 5 seconds. The SIFT algorithm extracts features on a 1024x1024 pixel image, sets the key point response threshold to 0.04, and extracts a maximum of 1000 feature points. The thin plate spline transform uses 20 control points and dynamically updates the transformation matrix based on the predicted distortion parameters. The fast approximate nearest neighbor algorithm uses a kd tree structure and sets the maximum number of searches to 50. The non-rigid registration algorithm uses a 32x32 pixel image block to calculate mutual information, uses the L-BFGS-B optimizer, has a maximum number of iterations of 100, and a convergence threshold of 1e-6. The entire processing flow is implemented on a GPU, and the average processing time for a single pattern recognition and correction is 50ms, meeting the real-time requirements.The non-rigid registration algorithm can adapt to a variety of fabrics with moisture contents ranging from 5% to 40%, with a pattern recognition accuracy of 98% and a geometric correction error of less than 2 pixels on average.

[0029] Step S104, based on the corrected pattern, special waterproof marks are added to the surface of the wet textile to make microscopic marks that do not affect the performance of the fabric, and the positions of the waterproof marks are captured by infrared imaging equipment as auxiliary reference points for pattern positioning.

[0030] The geometric information of a microscopic waterproof mark is obtained, and the geometric information is generated by AutoCAD software. The size of the microscopic waterproof mark is controlled within the range of 50-100 microns, and the shape is an irregular hexagon. According to the geometric information, a nano-scale waterproof material is used to prepare a mark, and the waterproof material solution is atomized into droplets with a diameter of 20-30 microns by piezoelectric inkjet technology. The droplets are positioned and sprayed on the surface of a wet textile by a piezoelectric inkjet print head with a resolution of 1200dpi, and the spray frequency is set to 10kHz to obtain a microscopic waterproof mark. The microscopic waterproof mark is imaged using a multispectral imaging system, and the position information and geometric features of the waterproof mark are extracted by wavelet transformation and morphological operations. It is judged whether the position information and geometric features match the original pattern. If they match, the RANSAC algorithm is used to screen the matching point pairs; the affine transformation matrix is ​​calculated according to the matching point pairs, and an enhanced pattern positioning model containing the waterproof mark is established.

[0031] Specifically, according to the corrected pattern contour information, the geometry of the microscopic waterproof mark is generated using AutoCAD software. The mark size is controlled within the range of 50-100 microns, and the shape is designed as an irregular hexagon with a side length ratio of 1:1.2:0.8:1:1.1:0.9 to ensure coordination with the textile fiber structure. The mark is prepared using nano-scale waterproof materials, and the waterproof material solution is atomized into droplets with a diameter of 20-30 microns through piezoelectric inkjet technology. The piezoelectric inkjet print head with a resolution of 1200dpi is used to position the spray on the wet textile surface, and the spray frequency is set to 10kHz to form a microscopic waterproof mark. The multispectral imaging system is used to image the textile surface after spraying, with a wavelength range of 1-14 microns, including 6 bands, and a spatial resolution of 10 microns / pixel. The position information and geometric features of the waterproof mark are extracted through wavelet transform and morphological operations. The extracted waterproof mark position information is registered with the original pattern, and the RANSAC algorithm is used to screen the matching point pairs, calculate the affine transformation matrix, set the number of iterations to 1000, and the internal point threshold to 2 pixels. An enhanced pattern positioning model containing waterproof marks is established for subsequent pattern recognition and positioning. In the actual implementation process, the corrected pattern contour information is first imported using AutoCAD2022 software, and the geometric figures of the microscopic waterproof marks are generated using a parametric design method. The mark is designed as an irregular hexagon, and the side length is set at a ratio of 1:1.2:0.8:1:1.1:0.9, with the longest side being 100 microns and the shortest side being 80 microns. An ethanol solution containing 10% hydrophobic nanosilica is used as a waterproof material, and the mark is made using a Fujifilm Dimatix DMP-2850 piezoelectric inkjet printer. The print head nozzle diameter is 21.5 microns, the droplet volume is 10 picoliters, and the jet frequency is set to 10kHz. The printing resolution is set to 1200 dpi, corresponding to a spacing of about 21.2 microns on the textile surface. The Headwall Hyperspec V NIR-SWIR multispectral imaging system was used to image the markers. The spectral range covers 1-14 microns and includes 6 bands (1-2.52m, 2.5-52m, 5-7.52m, 7.5-102m, 10-122m, 12-142m). The spatial resolution was set to 10 microns / pixel and the exposure time was 50 milliseconds. Image processing was performed using MATLAB R2021b software, and multi-scale analysis was performed using the Haar wavelet transform to extract edge information with a feature scale of 50-100 microns. Morphological operations used a 3x3 pixel structural element for opening and closing operations to remove noise and connect disconnected edges. Finally, the RANSAC algorithm was used for feature point matching and affine transformation matrix calculation, with the number of iterations set to 1000, the inlier threshold to 2 pixels, and the minimum consistent set size to 0.1% of the total pixels of the image.

[0032] When making microscopic marks that do not affect fabric performance, the diffusion and penetration behavior of the marking liquid between fabric fibers are simulated according to the fabric's hygroscopic properties to obtain the optimal marking liquid formula and coating process parameters; based on the influence of the marking size on the air permeability and softness of the fabric, the Pareto optimal solution of the marking size and fabric performance is solved, and the marking size range that maximizes the fabric feel and texture while meeting the pattern positioning requirements is obtained.

[0033] The Lucas-Washburn equation is used to simulate the diffusion and penetration behavior of the marking liquid between fabric fibers, and the diffusion speed and penetration depth of the liquid between fibers are calculated according to the porosity, fiber diameter and surface tension of the fabric; after obtaining the viscosity and surface tension values ​​of the marking liquid, the orthogonal test method is used to design the marking liquid formula experiment, and the solvent ratio, waterproofing agent content and surfactant concentration are selected as influencing factors; the marking morphology is observed by atomic force microscopy and the marking components are analyzed by Fourier transform infrared spectroscopy to determine the optimal formula of the marking liquid; a relationship model between the marking size and the air permeability and softness of the fabric is established, and the air permeability and softness indicators of the fabric are measured by the KES-F system; the Pareto optimal solution set is solved by a multi-objective genetic algorithm, and the population size, number of iterations, crossover probability and mutation probability are set to obtain the optimal marking size range that meets the pattern positioning requirements and fabric performance; according to the optimal marking size range and formula, the coating process is simulated by ANSYS Fluent software, the fluid domain is divided by hexahedral grids, the number of grids, coating roller speed, pressure and temperature are set, and the coating process parameters are optimized to ensure the uniformity and adhesion of the marking.

[0034] Specifically, the Lucas-Washburn equation was used to simulate the diffusion and penetration behavior of the marking liquid between fabric fibers. Considering the porous structure of the fabric, the diffusion rate and penetration depth of the liquid between fibers were calculated according to the porosity, fiber diameter and surface tension of the fabric, and the optimal viscosity and surface tension values ​​of the marking liquid were obtained. The marking liquid formula experiment was designed by the orthogonal test method. The solvent ratio, waterproofing agent content and surfactant concentration were selected as influencing factors. The L9 (3^4) orthogonal table with three factors and three levels was set. The marking morphology was observed by atomic force microscopy and the marking components were analyzed by Fourier transform infrared spectroscopy to determine the optimal marking liquid formula. The relationship model between the marking size and the air permeability and softness of the fabric was established. The KES-F system was used to measure the air permeability and softness indicators of the fabric. The Pareto optimal solution set was solved by a multi-objective genetic algorithm. The population size was set to 200, the number of iterations was set to 1000, the crossover probability was set to 0.85, and the mutation probability was set to 0.05. The optimal marking size range that met the pattern positioning requirements and fabric performance was obtained. According to the obtained optimal marking size range and formula, the coating process was simulated by ANSYS Fluent software. The fluid domain was divided by hexahedral grids, and the number of grids was set to 1 million. The boundary conditions included coating roller speed, pressure and temperature. The coating process parameters were optimized to ensure the uniformity and adhesion of the marking, and the final marking liquid formula and coating process parameters were obtained. In practical applications, the Lucas-Washburn equation was first used to simulate the diffusion of the marking liquid in the fabric. The fabric porosity was set to 0.7, the fiber diameter was 20 microns, and the surface tension was 30 mN / m. Through numerical solution, the optimal marking liquid viscosity was 10 mPa·s and the surface tension was 25 mN / m. Subsequently, an orthogonal test was carried out, and the solvent ratio (water: ethanol) was set to 7:3, 6:4, and 5:5, the waterproofing agent content was 2%, 3%, and 4%, and the surfactant concentration was 0.1%, 0.2%, and 0.3%. The marking morphology was observed using a Bruker Dimension Icon atomic force microscope with a scanning range of 10 μm×10 μm and a resolution of 512×512 pixels. ThermoScientific Nicoleti S50 Fourier transform infrared spectrometer was used to analyze the marking components, with a scanning range of 400-4000 cm^-1 and a resolution of 4 cm^-1. Through orthogonal test analysis, the optimal formula was determined to be a solvent ratio of 6:4, a waterproofing agent content of 3%, and a surfactant concentration of 0.2%. The KES-F system was used to measure the fabric properties, with a permeability test pressure difference of 4 mm water column and a test area of ​​10 cm2; the softness test bending stiffness range was 0.01-5 gf·cm2 / cm. In the multi-objective genetic algorithm, the weights of permeability and softness were each 0.5, and the optimal marking size range was 60-80 microns.Finally, ANSYS Fluent 2021R2 was used to simulate the coating process. The fluid domain size was 10 mm × 5 mm × 2 mm, and the number of grids was 1,236,742. The coating roller speed was set to 0.5 m / s, the pressure was 0.2 MPa, and the temperature was 25 ° C. Through optimization, the final coating speed was 0.4 m / s, the pressure was 0.18 MPa, and the temperature was 28 ° C. The coefficient of variation of the marking uniformity was less than 5%, and the adhesion firmness reached level 4 (remained intact after washing 100 times.

[0035] Step S105, based on the waterproof mark position of the pattern, the movement trajectory of the waterproof fabric is tracked, the degree of image distortion is calculated according to the motion vector field, and the moving waterproof mark fabric image is corrected in real time to eliminate the influence of motion blur on waterproof mark recognition.

[0036] The waterproof mark image between consecutive frames is obtained, and the motion estimation method based on block matching is used to calculate the motion vector field of the waterproof mark between the consecutive frames. The block matching method sets the block size to a preset pixel value, the search range to a preset pixel range, and the matching criterion uses normalized cross-correlation. According to the motion vector field, an image distortion model based on Point Spread Function is constructed. The distortion model estimates the PSF parameters through Zernike moments, and uses Fourier transform to analyze the image spectrum characteristics to obtain the direction and degree of motion blur. The Lucy-Richardson deconvolution algorithm is used to deblur the waterproof mark image. The deconvolution algorithm sets the maximum number of iterations to a preset value and the convergence threshold to a preset threshold to obtain a preliminary corrected image. The phase correlation method is used to finely align the preliminary corrected image, calculate the phase difference between the reference image and the corrected image, obtain the displacement field through inverse Fourier transform, and use bicubic interpolation to resample the image. The processing flow is implemented on the GPU through the CUDA parallel computing framework, and the single frame processing time is controlled within the preset time.

[0037] Specifically, the motion estimation method based on block matching is used to calculate the motion vector field of the waterproof marker between consecutive frames. The block size is set to 16×16 pixels, the search range is 32×32 pixels, and the matching criterion is normalized cross-correlation to obtain the motion trajectory and speed information of the marker. According to the calculated motion vector field, an image distortion model based on PointSpreadFunction (PSF) is constructed. The PSF parameters are estimated by Zernike moments. The image spectrum characteristics are analyzed by Fourier transform to obtain the direction and degree of motion blur and determine the degree of image distortion. Based on the distortion model, the Lucy-Richardson deconvolution algorithm is used to deblur the moving waterproof marker image. The maximum number of iterations is set to 30 and the convergence threshold is set to 0.001 to obtain the preliminary corrected image. The phase correlation method is used to finely align the corrected image. The phase difference between the reference image and the corrected image is calculated. The displacement field is obtained by inverse Fourier transform. The image is resampled by bicubic interpolation. The least squares fitting is combined to achieve sub-pixel precise alignment. The alignment accuracy reaches 0.1 pixel. The entire processing flow is implemented on the GPU through the CUDA parallel computing framework, and the single frame processing time is controlled within 10 milliseconds, meeting the requirements of real-time correction. In practical applications, the motion estimation method based on block matching is first used to process continuous frame images with a resolution of 1920×1080. The image is divided into blocks of 16×16 pixels, and matching is performed within the search range of 32×32 pixels. The normalized mutual correlation coefficient is used as the matching criterion, and the threshold is set to 0.8. Through parallel computing, a processing speed of 100 frames per second is achieved on the NVIDIARTX3080GPU. After obtaining the motion vector field, the PSF is estimated using the Zernike moment, the coefficients of the first 20 orders of the Zernike polynomial are calculated, and the PSF kernel function is reconstructed. The Fourier transform uses 2DFFT to analyze the spectral energy distribution in the range of 0-π / 2 to determine the main motion direction and blur degree. In the Lucy-Richardson deconvolution algorithm, the initial estimate is set to the original image with Gaussian blur (σ=1.5), the maximum number of iterations is 30, and the convergence threshold is 0.001. In each iteration, the CUDA kernel function is used to accelerate the convolution operation, and the average processing time per frame is 5ms. In the phase correlation method, the image is processed in 32×32 blocks, the phase spectrum of each block is calculated, and the displacement field is obtained by inverse Fourier transform. The least squares method is used to fit the quadratic surface to achieve a sub-pixel registration accuracy of 0.1 pixel. The bicubic interpolation uses a 4×4 pixel neighborhood, and the weight coefficient is accelerated by table lookup.

[0038] Step S106, based on the correction of the fast-moving textile image, the textile texture is segmented, the waterproof mark is used as a seed point to expand around, the texture boundary is determined, and the large area pattern is divided into multiple sub-areas. At the same time, feature extraction and dislocation analysis are performed on each sub-area, and the analysis results of each sub-area are summarized to evaluate the dislocation recognition status of the overall textile pattern.

[0039] A textile image is received, and an adaptive histogram equalization method is used to enhance the image contrast, the method comprising setting a predetermined block size and a contrast limit parameter; based on the enhanced image, a Richardson-Lucy algorithm is used to perform motion blur correction, the algorithm setting a predetermined number of iterations; a bilateral filtering process is performed on the corrected image, the bilateral filtering comprising setting a predetermined spatial domain standard deviation and a value range standard deviation; an adaptive threshold segmentation method is used to process the filtered image to obtain a binary image, the method comprising setting a predetermined window size; if a waterproof mark is detected in the binary image, the waterproof mark is used as a seed point, and a regional growing algorithm is used to perform texture segmentation, the algorithm comprising: setting a predetermined local entropy threshold as a growth criterion, and judging whether the local entropy exceeds the preset threshold; if the local entropy exceeds the preset threshold, determining the corresponding area as a target texture area.

[0040] Specifically, the corrected textile image was preprocessed, and adaptive histogram equalization was used to enhance the image contrast. The block size was set to 8×8 pixels, and the contrast limit parameter was set to 0.01. The Richardson-Lucy algorithm was used for motion blur correction, and the number of iterations was set to 10. Bilateral filtering was used to remove noise and retain edge information. The standard deviation of the spatial domain was set to 3, and the standard deviation of the value domain was set to 0.1. The adaptive threshold segmentation method was used to obtain the binary image, and the window size was 51×51 pixels. The waterproof mark was used as the seed point, and the region growing algorithm was used for texture segmentation. The local entropy was used as the growth criterion, the threshold was set to 1.2 times the local entropy, the 8-neighborhood connectivity was used to determine the extension direction, and the maximum number of iterations was set to 10% of the total number of image pixels. The boundary extraction and sub-region division of the segmented texture area were performed, and the Canny edge detection algorithm was used to extract the texture boundary. The low threshold was 50 and the high threshold was 150. The SLIC superpixel algorithm was used for initial segmentation of large-area patterns. The number of superpixels was set to 0.1% of the image area, and the compactness factor was 10. The final sub-region division result is obtained by the regional similarity merging method, and the similarity threshold is set to 0.85. Feature extraction and dislocation analysis are performed for each sub-region. The local binary pattern (LBP) is used to extract texture features, and the feature similarity between sub-regions is calculated. The chi-square distance is used to measure the difference of the LBP histogram. The similarity threshold of 0.9 is set to determine whether the sub-region is dislocated. At the same time, the global Fourier descriptor is calculated to evaluate the overall dislocation status, and the first 20 low-frequency coefficients are taken as feature vectors. Finally, the analysis results of each sub-region are summarized, and the dislocation recognition status of the overall textile pattern is evaluated by combining local and global features. The entire processing flow is implemented on the GPU through CUDA parallel computing, and the single frame processing time is controlled within 20 milliseconds. In practical applications, the textile image with a resolution of 2048×1536 is first preprocessed. Adaptive histogram equalization uses 8×8 pixel blocks, and the contrast limit parameter is set to 0.01. Parallel processing is achieved through CUDA kernel function, which takes about 2ms. Richardson-Lucy algorithm is used for motion blur correction, assuming that the motion blur kernel size is 21×21 pixels, 10 iterations, parallel calculation on GPU, and takes about 5ms. Bilateral filtering uses a 5×5 pixel window, spatial domain σ=3, value range σ=0.1, and is accelerated by lookup table, taking about 1ms. Adaptive threshold segmentation uses a 51×51 pixel sliding window to calculate the local mean and standard deviation, and the threshold is set to the mean minus 0.2 times the standard deviation, which takes about 2ms. The region growing algorithm uses the detected waterproof mark center point as the seed, calculates the local entropy of the 9×9 neighborhood, and the growth threshold is set to 1.2 times the local entropy. The maximum iteration is 200,000 times (about 10% of the total number of pixels), which takes about 3ms.Canny edge detection uses 3×3 Sobel operator to calculate gradient, low threshold 50, high threshold 150, and obtains edge through non-maximum suppression and double threshold linking, which takes about 1ms. SLIC superpixel algorithm sets the initial number of superpixels to 3,000 (about 0.1% of image area), iterates 10 times, uses 5-dimensional feature vector (Lab* color space and xy coordinates), and takes about 4ms. Region merging is based on color histogram similarity, threshold 0.85, and uses union-find data structure to accelerate the merging process, which takes about 1ms. LBP feature extraction uses 3×3 neighborhood, 8-bit encoding, calculates 256-dimensional histogram, and normalizes the LBP features of each sub-region, which takes about 0.5ms. Fourier descriptor calculation uses 2DFFT, selects 20 low-frequency coefficients as global features, and takes about 0.5ms. The entire process is implemented on NVIDIA RTX3080 GPU, with a total time of about 20ms, meeting real-time processing requirements.

[0041] Obviously, those skilled in the art can make various changes and modifications to the embodiments of the present application without departing from the spirit and scope of the embodiments of the present application. Thus, if these modifications and variations of the embodiments of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for identifying misalignment of textile patterns, characterized in that: The method comprises: The textile images in a humid environment are preprocessed, and the pattern features are extracted. A reference model of the normal morphology of the pattern is established according to the pre-trained data set, and the pattern area and the background area are separated from the original image to obtain the pattern contour information. Perform geometric transformation analysis on the extracted pattern contour information, calculate the relative position relationship of each part of the pattern, establish a pattern moisture deformation model, and determine whether the pattern is dislocated or deformed by comparing the difference between the current pattern and the reference model; If the pattern is found to be misaligned or deformed, a mathematical model of the effect of moisture on the pattern is established based on the material characteristics of the textile and the production process parameters. By simulating the change rules of the pattern under different humidity conditions, the test results are dynamically corrected to eliminate the interference of moisture on the recognition accuracy. Based on the corrected pattern, special waterproof marks are added to the surface of wet textiles to make microscopic marks that do not affect the performance of the fabric. The positions of the waterproof marks are captured by infrared imaging equipment as auxiliary reference points for pattern positioning. Based on the waterproof mark position of the pattern, the movement trajectory of the waterproof fabric is tracked, the degree of image distortion is calculated according to the motion vector field, and the image of the waterproof mark fabric in motion is corrected in real time to eliminate the influence of motion blur on the waterproof mark recognition; Based on the correction of fast-moving textile images, the textile texture is segmented, and the waterproof mark is used as the seed point to expand to the surrounding area to determine the texture boundary. The large area pattern is divided into multiple sub-regions. At the same time, feature extraction and dislocation analysis are performed on each sub-region. The analysis results of each sub-region are summarized to evaluate the dislocation recognition status of the overall textile pattern.

2. The method according to claim 1, wherein: The method of preprocessing the textile image in a humid environment, extracting pattern features, establishing a reference model of the normal form of the pattern according to a pre-trained data set, separating the pattern area and the background area from the original image, and obtaining the pattern contour information includes: The original image of the textile in a humid environment is obtained, and the dark channel prior defogging algorithm is used to eliminate the blur effect of the original image to obtain the defogging image; The dehazed image is processed by adaptive histogram equalization and guided filtering, and the optimal threshold is calculated according to the OTSU algorithm for binarization to obtain the preprocessed binary image. Morphological operations are performed on the preprocessed binary image. Opening operations are used to remove noise areas, closing operations are used to fill holes inside the pattern, the pattern contour is extracted using the Canny edge detection algorithm, and the SIFT feature descriptor of the pattern is extracted using the scale-invariant feature transformation algorithm. According to a pre-trained normal morphology reference model, the SIFT feature is compared with the model feature, and the cosine similarity between the feature vectors is calculated; If the cosine similarity is higher than a preset threshold, it is determined to be a normal pattern, otherwise it is determined to be an abnormal area; Perform region growing segmentation on the normal pattern area, select the pixel point with the largest pattern edge gradient as the seed point, calculate the Euclidean distance and gradient direction difference of adjacent pixels, and include the pixel in the growth area if both are less than the preset threshold value, until no new pixel meets the growth condition or the preset maximum number of iterations is reached, and obtain the separated pattern area and background area.

3. The method according to claim 1, wherein: The method of performing geometric transformation analysis on the extracted pattern contour information, calculating the relative position relationship between the various parts of the pattern, establishing a pattern moisture deformation model, and determining whether the pattern is dislocated or deformed by comparing the difference between the current pattern and the reference model includes: Obtain pattern contour information, use Harris corner detection operator to identify feature points in the pattern, and filter out stable feature points based on preset corner response threshold and non-maximum suppression window size; Calculating the Euclidean distance and angle between the stable feature points, constructing a geometric structure description of the pattern, and obtaining the relative position relationship of each part of the pattern; For the stable feature points, an elastic mesh deformation model is established, the pattern is divided into meshes of a preset size, and the displacement of the internal points of the meshes is calculated using bicubic interpolation to obtain a nonlinear deformation field; Compare the geometric structure description of the current pattern with that of the reference model, use the accelerated robust feature algorithm to extract local feature descriptors, use the nearest neighbor ratio method to perform feature matching, and if the Euclidean distance difference between the matching point pairs exceeds the preset threshold, it is determined that the area is misaligned; The dislocated area is corrected by using a thin plate spline interpolation algorithm, a pair of matching points that have not been deformed is selected as a control point, a radial basis function is constructed, the interpolation coefficient is solved, the dislocated area is resampled, and the corrected pattern contour information is obtained.

4. The method according to claim 1, wherein: If the pattern is found to be misaligned or deformed, a mathematical model of the effect of moisture on the pattern is established according to the material characteristics of the textile and the production process parameters. By simulating the change law of the pattern under different humidity conditions, the detection result is dynamically corrected to eliminate the interference of moisture on the recognition accuracy, including: The material water absorption rate parameters and the ambient humidity value are obtained, and a multi-layer perceptron neural network is used to establish a nonlinear mapping relationship between the material water absorption rate, ambient humidity and pattern deformation degree; Constructing a mathematical model of the effect of moisture on the pattern according to the nonlinear mapping relationship; Use environmental humidity sensors to collect real-time humidity data, and build a time series prediction model based on the dynamic moisture absorption characteristics of the material; The time series prediction model uses a long short-term memory network to process the humidity data sequence and predict the change of moisture content inside the material over time; Inputting the predicted moisture content into the multi-layer perceptron neural network to calculate the theoretical deformation state of the pattern under the current humidity condition; Comparing the theoretical deformation state with the actually detected pattern misalignment or deformation, and calculating a deviation matrix between the two; Construct a state space model, where the state variables are the deviations of the pattern geometric features and the observed variables are the actual detection results; The Kalman filter algorithm is used for dynamic estimation and correction, and the process noise covariance matrix and the observation noise covariance matrix are estimated offline through historical data; Correcting the pattern geometric features using the filtered state estimate; Matching the corrected pattern geometric features with the original reference model, and calculating the affine transformation matrix using a mutual information-based image registration algorithm; The mutual information objective function is optimized using the gradient descent method, and the transformation parameters are iteratively updated until convergence or the maximum number of iterations is reached; It also includes: constructing a mathematical model of the influence of moisture on patterns according to different fabric material characteristics and production process parameters. The model is used to quantitatively describe the geometric distortion effect of changes in moisture content on the fabric surface pattern, so as to predict the distortion degree and distortion direction parameters. While identifying the fabric surface pattern, real-time geometric correction of the identification result is performed based on the predicted pattern distortion degree and distortion direction.

5. The method according to claim 4, wherein: According to the method, a mathematical model of the influence of moisture on the pattern is constructed for different fabric material characteristics and production process parameters. The model is used to quantitatively describe the geometric distortion effect of moisture content change on the fabric surface pattern, so as to predict the distortion degree and distortion direction parameters. When the fabric surface pattern is identified, the identification result is geometrically corrected in real time according to the predicted pattern distortion degree and distortion direction, including: Constructing a database of fabric material properties and production process parameters, and using a fully connected neural network to establish a mapping relationship between moisture content and pattern geometric distortion, wherein the fully connected neural network includes an input layer, a hidden layer, and an output layer; Using a capacitive sensor array to collect moisture content data on the surface of the fabric, the moisture content data is used as input to a long short-term memory network; Obtaining a moisture content prediction sequence according to the output of the long short-term memory network, and inputting the moisture content prediction sequence into the fully connected neural network; Obtaining the output of the fully connected neural network to obtain a sequence of predicted values ​​of the degree and direction of pattern distortion; The scale-invariant feature transformation algorithm is used to extract the actual pattern feature points and descriptors of the fabric surface, and the time-varying thin plate spline transformation matrix is ​​constructed according to the predicted value sequence of the pattern distortion degree and direction. If the time-varying thin plate spline transformation matrix is ​​constructed, a non-rigid correction is performed on the actual pattern feature points; The corrected pattern is compared with the standard pattern using a non-rigid registration algorithm based on mutual information; The transformation parameters are optimized by gradient descent method to obtain the pattern recognition results.

6. The method according to claim 1, wherein: The method of adding a special waterproof mark on the surface of the wet textile based on the corrected pattern to make a microscopic mark that does not affect the performance of the fabric, and capturing the position of the waterproof mark by an infrared imaging device as an auxiliary reference point for pattern positioning includes: Acquire geometric information of a microscopic waterproof mark, wherein the geometric information is generated by AutoCAD software, and the size of the microscopic waterproof mark is controlled within the range of 50-100 microns and the shape is an irregular hexagon; According to the geometric information, a mark is prepared by using a nano-scale waterproof material, and the waterproof material solution is atomized into droplets with a diameter of 20-30 microns by using a piezoelectric inkjet technology; The droplets are positioned and sprayed on the surface of the wet textile by a piezoelectric inkjet print head with a resolution of 1200 dpi, and the spraying frequency is set to 10 kHz to obtain a microscopic waterproof mark; The microscopic waterproof mark is imaged using a multispectral imaging system, and the position information and geometric features of the waterproof mark are extracted through wavelet transformation and morphological operations; Determine whether the position information and geometric features match the original pattern, and if so, use the RANSAC algorithm to select matching point pairs; Calculating an affine transformation matrix according to the matching point pairs, and establishing an enhanced pattern positioning model including the waterproof mark; It also includes: when making microscopic marks that do not affect fabric performance, based on the fabric's hygroscopic properties, simulating the diffusion and penetration behavior of the marking liquid between fabric fibers, and obtaining the optimal marking liquid formula and coating process parameters; according to the influence of the marking size on the air permeability and softness of the fabric, solving the Pareto optimal solution of the marking size and fabric performance, and obtaining the marking size range that maximizes the fabric feel and texture while meeting the pattern positioning requirements.

7. The method according to claim 6, wherein: When making the microscopic mark that does not affect the fabric performance, the diffusion and penetration behavior of the marking liquid between the fabric fibers are simulated according to the hygroscopic properties of the fabric to obtain the best marking liquid formula and coating process parameters; According to the influence of the marking size on the air permeability and softness of the fabric, the Pareto optimal solution of the marking size and fabric performance is solved, and the marking size range that can meet the pattern positioning requirements while maximally maintaining the fabric feel and texture is obtained, including: The Lucas-Washburn equation is used to simulate the diffusion and penetration behavior of the marking liquid between fabric fibers, and the diffusion speed and penetration depth of the liquid between fibers are calculated according to the porosity, fiber diameter and surface tension of the fabric; after obtaining the viscosity and surface tension values ​​of the marking liquid, the orthogonal test method is used to design the marking liquid formula experiment, and the solvent ratio, waterproofing agent content and surfactant concentration are selected as influencing factors; the marking morphology is observed by atomic force microscopy and the marking components are analyzed by Fourier transform infrared spectroscopy to determine the optimal formula of the marking liquid; a relationship model between the marking size and the air permeability and softness of the fabric is established, and the air permeability and softness indicators of the fabric are measured by the KES-F system; the Pareto optimal solution set is solved by a multi-objective genetic algorithm, and the population size, number of iterations, crossover probability and mutation probability are set to obtain the optimal marking size range that meets the pattern positioning requirements and fabric performance; according to the optimal marking size range and formula, the coating process is simulated by ANSYS Fluent software, the fluid domain is divided by hexahedral grids, the number of grids, coating roller speed, pressure and temperature are set, and the coating process parameters are optimized to ensure the uniformity and adhesion of the marking.

8. The method according to claim 1, wherein: The waterproof mark position based on the pattern tracks the motion trajectory of the waterproof fabric, calculates the degree of image distortion according to the motion vector field, performs real-time correction on the moving waterproof mark fabric image, and eliminates the influence of motion blur on waterproof mark recognition, including: Acquire waterproof mark images between consecutive frames, and calculate the waterproof mark motion vector field between the consecutive frames by using a motion estimation method based on block matching, wherein the block matching method sets the block size to a preset pixel value, the search range to a preset pixel range, and the matching criterion adopts normalized cross-correlation; According to the motion vector field, an image distortion model based on PointSpreadFunction is constructed, wherein the distortion model estimates PSF parameters through Zernike moments and obtains the direction and degree of motion blur by analyzing the image spectrum characteristics using Fourier transform; The Lucy-Richardson deconvolution algorithm is used to deblur the waterproof mark image, wherein the deconvolution algorithm sets the maximum number of iterations to a preset value and the convergence threshold to a preset threshold to obtain a preliminarily corrected image; The image after preliminary correction is finely registered by using a phase correlation method, the phase difference between the reference image and the correction image is calculated, the displacement field is obtained by inverse Fourier transform, and the image is resampled by using bicubic interpolation; The processing flow is implemented on the GPU through the CUDA parallel computing framework, and the single frame processing time is controlled within a preset time.

9. The method according to claim 1, wherein: The correction based on the fast-moving textile image is to segment the textile texture, expand it to the surrounding area with the waterproof mark as the seed point, determine the texture boundary, and divide the large area pattern into multiple sub-areas; at the same time, feature extraction and dislocation analysis are performed on each sub-area, and the analysis results of each sub-area are summarized to evaluate the dislocation recognition status of the overall textile pattern, including: Receiving a textile image, and enhancing the image contrast using an adaptive histogram equalization method, the method comprising setting a predetermined block size and a contrast limit parameter; According to the enhanced image, using the Richardson-Lucy algorithm to perform motion blur correction, the algorithm is set with a predetermined number of iterations; Performing bilateral filtering on the corrected image, wherein the bilateral filtering includes setting a predetermined spatial domain standard deviation and a value range standard deviation; Processing the filtered image using an adaptive threshold segmentation method to obtain a binary image, the method comprising setting a predetermined window size; If a waterproof mark is detected in the binary image, the waterproof mark is used as a seed point and a region growing algorithm is used to perform texture segmentation, the algorithm comprising: setting a predetermined local entropy threshold as a growth criterion and determining whether the local entropy exceeds the preset threshold; If the local entropy exceeds a preset threshold, the corresponding area is determined to be a target texture area.

Images