A multi-channel image data driven fabric weave analysis system

Abstract

The application discloses a multi-channel image data driven fabric fabric analysis system, and relates to the technical field of intelligent detection of the textile industry, comprising an image sampling unit, a data processing unit and a component prediction unit; the application is characterized in that a sampling detection belt is arranged on the detected fabric, a multi-channel sensor group comprising visible light, near-infrared and polarization imaging sensors is arranged at the detection belt, multi-dimensional image information of the fabric is collected and preprocessed, an image feature vector group is extracted, the classification and proportion of components in the sampling area are analyzed, and a texture direction reasoning mechanism is combined to realize high-precision prediction of the component distribution and proportion of the whole fabric image; the precision of component recognition is effectively improved, the robustness is significantly reduced, the data processing burden and required computing resources are reduced, the system response speed is improved, and engineering application feasibility is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection technology in the textile industry, and in particular to a fabric analysis system driven by multi-channel image data. Background Technology

[0002] Fabric analysis refers to the process of systematically identifying, measuring, and evaluating the structural, compositional, performance, and quality characteristics of fabrics (including fabrics, yarns, and fibers) through a series of physical, chemical, instrumental, and sensory testing methods.

[0003] Multi-channel image data-driven fabric analysis is a system that integrates multiple optical imaging technologies (such as visible light, near-infrared spectroscopy, and polarized light imaging) to acquire multi-dimensional information about fabrics and performs collaborative analysis through artificial intelligence algorithms. Its core function is to overcome the limitations of single sensors and achieve component identification (such as the proportion of blended materials), defect detection (holes, stains), structural analysis (fiber orientation, weaving density), and color quantification (dye uniformity, color fastness).

[0004] Current fabric analysis technology is typically used in the fields of intelligent sorting of waste textiles, quality inspection of fabric production, and optimization of silk and cotton printing processes. Especially when selecting processing solutions for recycled fabrics, contamination is often encountered. When dealing with recycled materials with complex compositions (blended, contaminated), its adaptability is too poor, and the quality deteriorates significantly (strength, color). This leads to misjudgment of composition due to differences in optical reflection in mixed fibers and multi-layered woven fabrics (e.g., dark fabrics absorb near-infrared signals). Analyzing its composition, impurity content, and recyclability is extremely difficult. Considering the inspection cost and testing time, detailed analysis of each piece of old clothing is more costly and time-consuming than the recycling market expects.

[0005] To address the aforementioned technical deficiencies, a solution is proposed. Summary of the Invention

[0006] The purpose of this invention is to solve the problem that the cost of analyzing the composition of recycled materials with complex components (blended or dyed) is too high, making it difficult to apply in practice.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a multi-channel image data-driven fabric analysis system, comprising an image sampling unit, a data processing unit, and a composition prediction unit.

[0008] The image sampling unit is used to set up a sensor sampling detection belt on the fabric to be detected, and set up a sensor group at the detection belt to collect multi-dimensional information of the fabric. After preprocessing the multi-dimensional information, a multi-channel dataset is generated and sent to the data processing unit.

[0009] Furthermore, the image sampling unit generates a multi-channel dataset, including the following steps:

[0010] The sensor group includes a visible light sensor, a near-infrared channel one sensor, a near-infrared channel two sensor, and a polarized light imaging sensor.

[0011] The sampling detection band includes horizontal sampling bands and vertical sampling bands, and each type of sampling band includes several sampling points; and the setting of the sampling bands on the fabric to be tested depends on the analysis results of the data collected by the visible light sensor;

[0012] The sampling band is set up as follows: based on the visible light image, the longest straight line on the fabric surface is extracted as the horizontal main axis, and sampling points are set at equal intervals along this axis to form a horizontal sampling band; the vertical sampling band is set up in the same way; the extension range of the sampling band is determined based on the optimal detection distance of each channel sensor.

[0013] Generating a multi-channel dataset includes the following steps:

[0014] The visible light sensors, with a spectral range of 400-700nm, are arranged in a ring around the fabric to be tested. The distance between each visible light sensor and the fabric is set to the optimal distance to ensure the accuracy of data acquisition.

[0015] At each of the sampling points, a set of near-infrared channel one sensor, near-infrared channel two sensor and polarized light imaging sensor are provided;

[0016] The near-infrared channel one sensor has a spectral range of 900-1100nm; the near-infrared channel two sensor has a spectral range of 1450-1550nm; the polarized light imaging uses specific polarization angles, including 0°, 45°, 90°, and 135°; and the distance between each sensor and the fabric is set to the optimal distance to ensure data acquisition accuracy.

[0017] Define ki folders (i=1, 2, 3, ..., n, where n is the total number of fabrics to be detected and i is the fabric number to be detected). Each ki folder contains four types of folders: a, b, c, and d. Define the set of all ki folders as a multi-channel dataset.

[0018] Among them, folder A contains image acquisition data of the corresponding fabric in folder ki by the visible light sensor, denoted as data A;

[0019] Type B files include image acquisition data of the corresponding fabric in the ki folder from a near-infrared channel sensor, denoted as Data B;

[0020] Class C files include image acquisition data of the corresponding fabric in the ki folder from the near-infrared channel dual sensors, denoted as data C;

[0021] Type d files include image acquisition data of the corresponding fabric in the ki folder from the polarization sensor, denoted as data D;

[0022] The optimal distance is obtained through the following steps: Under fixed light source conditions, the distance between the sensor and the fabric is gradually adjusted, and image quality indicators (e.g., reflectance coefficient, texture sharpness, noise) are recorded at each distance. A comprehensive optimal distance is selected based on different spectral channels (visible light, near-infrared, polarized light). After determining the distance, multiple data acquisitions are performed to ensure that the quality of each acquired image is stable and that there is no obvious distortion or falsification. Fine-tuning is performed based on the acquired data, for example, adjusting the sensor distance based on changes in the reflectance spectrum, thereby achieving more refined image data.

[0023] The data processing unit is used to receive and process multi-channel datasets, construct an image feature analysis model based on the convolutional neural network (CNN) algorithm to extract four-dimensional data feature vector groups. The input of the image feature analysis model is the multi-channel dataset, and the output of the image feature analysis model is four types of feature vector groups. The four types of feature vector groups are integrated to generate a multi-channel vector dataset and sent to the component prediction unit.

[0024] Furthermore, the data processing unit preprocesses the multi-channel dataset, including the following steps:

[0025] Data A, data B, data C, and data D are preprocessed respectively, including: normalization, noise reduction, data augmentation, and channel independence processing;

[0026] The data enhancements include rotation, translation, scaling, and brightness adjustment;

[0027] The channel preprocessing includes:

[0028] Band differential processing is performed on near-infrared channel one and near-infrared channel two to suppress dye absorption interference;

[0029] Calculate Stokes vectors I, Q, U, and V from polarization image data and generate images of polarization degree and polarization angle;

[0030] Perform white balance and color constancy correction on visible light images;

[0031] Furthermore, an image feature analysis model is constructed to extract four types of feature vector groups, generating a multi-channel vector dataset, including the following steps:

[0032] Based on the image data in folders a, b, c, and d within folder ki of each fabric in the preprocessed multi-channel dataset, an image feature extraction model is trained.

[0033] Define the input format of the image feature extraction model as F0 = {Fa, Fb, Fc, Fd}; define the output format of the image feature extraction model as F1 = {k, y, e, p}, where k, y, e, and p are all vector groups.

[0034] Wherein, k is a set of visible parameter vectors, including color intensity, chroma, saturation, brightness, and spatial distribution features of texture;

[0035] The y is a set of parameter vectors, including reflectance, spectral reflectance curve, hygroscopicity, and fiber composition characteristics;

[0036] The term e is a set of two types of parameter vectors, including reflection intensity and dye absorption characteristics;

[0037] p is a set of polarization difference parameter vectors, including degree of polarization, polarization angle, and surface texture features;

[0038] Among them, the visible parameter vector group k, the first type parameter vector group y, the second type parameter vector group e, and the polarization difference parameter vector group p together form the feature vector group Fki, where i is the number of the fabric corresponding to this group of parameters.

[0039] Generate a multi-channel vector dataset, including the preprocessed ki group folder and the corresponding feature vector group;

[0040] Composition prediction unit: It is used to receive multi-channel vector datasets and sequentially pass through the composition mapping model construction module and the full-image composition distribution prediction module to complete the prediction of the overall composition category and proportion of the fabric.

[0041] The component mapping model construction module, based on a multi-task learning network, integrates image features from different channels in the multi-channel vector dataset, extracts component categories and their corresponding proportions within the sampling interval, generates sampling analysis data, and sends it to the full-image component distribution prediction module.

[0042] Furthermore, the component mapping model construction module generates sampling analysis data, including the following steps:

[0043] Based on a multi-task learning network structure, the input is a multi-channel vector dataset corresponding to the sampling region. The fused representation vector is generated through a channel-independent encoder and a feature fusion module. These vectors are then input into the component category recognition branch and the proportional regression branch, respectively. The output is the principal component category of the region and its proportion, forming the sampling analysis data.

[0044] The full-image component distribution prediction module, based on the sampling analysis data, uses visible light images and polarization images to calculate the fiber texture direction vector at the sampling point, and predicts the similarity between its main components and the sampling area in the direction of the vector, thereby completing the overall prediction of the fabric.

[0045] Furthermore, the process by which the component prediction unit completes the prediction of the overall image component distribution includes the following steps:

[0046] Based on the sampled analysis data, the main texture direction vector Xo is calculated using visible light images and polarization images, where o is the sampling point number; a sliding window sequence is constructed along this vector direction, and the texture features of each window region are extracted;

[0047] Subsequently, the similarity between the texture features of these regions and the features of the sampled regions is calculated, and regions similar to the features of the sampled regions are selected to form a set of similar regions J. A mask Mj is generated based on the set J, where j is the number of the similar regions. The main orientation vector Xo of the non-sampled regions and its mask Mj information are predicted by the orientation field interpolation method, thereby inferring the composition categories and their proportional relationships of the entire image.

[0048] Based on this, the feature vectors of neighboring sampling points are combined to perform weighted component prediction on the non-sampling area. Finally, the prediction results are evaluated for error to ensure the accuracy and reliability of the prediction.

[0049] The formula for calculating the similarity is as follows:

[0050] Calculate directional consistency using the cosine similarity formula:

[0051] ,

[0052] Where θ is the angle between the vectors corresponding to the texture features and the sampling region features;

[0053] Similarity calculation formula:

[0054] ;

[0055] in, T represents the similarity between the sliding window and the monitoring area for the sampling point. h Let h be the texture vector of the sliding window, and h represent the h-th sliding window. The weighting of fabric composition and orientation. The Mahalanobis distance;

[0056] If the similarity is higher than the preset value Xs, the regions are considered to have the same orientation. All regions that continuously meet this condition are marked on the image, and a set D of continuous regions with the same orientation is constructed. The preset value Xs is derived from a large number of experiments.

[0057] Label the regions within set D, including component categories and proportions, and generate a mask Mj;

[0058] Based on the orientation field interpolation function, an orientation propagation prediction model is established to predict the main orientation vector Xo and its mask Mj in the non-sampling area, thereby obtaining the orientation field of the whole image, and then completing the prediction of the composition category and proportion relationship of the fabric as a whole.

[0059] For non-sampled regions, the predicted components are weighted using the vector set Fki of the nearest neighboring sampled points, and the error is evaluated through cross-validation.

[0060] Furthermore, obtaining the main directional vector Xo includes the following steps:

[0061] The vector representing the average direction of the main fiber texture within the sampling area is defined as vector Xo, where o is the sampling point number. Vector Xo is used to divide the sampling area into local texture regions using data A and data D. The structure tensor of each texture region is calculated to obtain the local dominant direction vector Yi, where i is the texture region number. All vectors Yi are combined and weighted to obtain vector Xo within the sampling area.

[0062] Furthermore, the selection of the step size of the sliding window includes the following steps:

[0063] Based on the sampling analysis data, the fibers of all component categories within the sampling area are first identified, and the component category with the shortest average length of a single fiber is selected. The maximum fiber length within that component category is used as a benchmark, and the step size of the sliding window is set to 5 times that maximum value.

[0064] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0065] This multi-channel image data-driven fabric analysis system collects and preprocesses multi-dimensional image information of the fabric by setting up a sampling detection zone on the fabric and arranging a multi-channel sensor group including visible light, near-infrared and polarization imaging sensors at the detection zone. It extracts image feature vector groups, analyzes the composition categories and proportions in the sampling area, and combines the texture direction inference mechanism to achieve high-precision prediction of the composition distribution and proportion of the entire fabric image.

[0066] Traditional methods struggle to accurately identify the composition distribution of blended and composite fabrics, especially when fiber orientation is complex or color interference is significant, leading to errors. This method integrates visible light, near-infrared, and polarization channel data and introduces a texture orientation inference mechanism to achieve spatial continuity modeling of regional components and full-image component prediction. This effectively improves the accuracy and robustness of component identification, significantly reduces data processing burden and required computing resources, and enhances system response speed and engineering application feasibility. Attached Figure Description

[0067] Figure 1 The schematic diagram of the principle framework of the present invention is shown. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Example:

[0070] like Figure 1 As shown, this application provides a multi-channel image data-driven fabric analysis system, including an image sampling unit, a data processing unit, a composition prediction unit, and a recyclability assessment and feedback unit.

[0071] The image sampling unit is used to set up a sensor sampling detection belt on the fabric to be detected, and set up a sensor group at the detection belt to collect multi-dimensional information of the fabric. After preprocessing the multi-dimensional information, a multi-channel dataset is generated and sent to the data processing unit.

[0072] Generating a multi-channel dataset includes the following steps:

[0073] The sensor group includes a visible light sensor, a near-infrared channel one sensor, a near-infrared channel two sensor, and a polarized light imaging sensor.

[0074] The sampling detection band includes horizontal sampling bands and vertical sampling bands, and each type of sampling band includes several sampling points; and the setting of the sampling bands on the fabric to be tested depends on the analysis results of the data collected by the visible light sensor;

[0075] The sampling band is set up as follows: based on the visible light image, the longest straight line on the fabric surface is extracted as the horizontal main axis, and sampling points are set at equal intervals along this axis to form a horizontal sampling band; the vertical sampling band is set up in the same way; the extension range of the sampling band is determined based on the optimal detection distance of each channel sensor.

[0076] Generating a multi-channel dataset includes the following steps:

[0077] The visible light sensors, with a spectral range of 400-700nm, are arranged in a ring around the fabric to be tested. The distance between each visible light sensor and the fabric is set to the optimal distance to ensure the accuracy of data acquisition.

[0078] At each of the sampling points, a set of near-infrared channel one sensor, near-infrared channel two sensor and polarized light imaging sensor are provided;

[0079] The near-infrared channel one sensor has a spectral range of 900-1100nm; the near-infrared channel two sensor has a spectral range of 1450-1550nm; the polarized light imaging uses specific polarization angles, including 0°, 45°, 90°, and 135°; and the distance between each sensor and the fabric is set to the optimal distance to ensure data acquisition accuracy.

[0080] Define ki folders (i=1, 2, 3, ..., n, where n is the total number of fabrics to be detected and i is the fabric number to be detected). Each ki folder contains four types of folders: a, b, c, and d. Define the set of all ki folders as a multi-channel dataset.

[0081] Among them, folder A contains image acquisition data of the corresponding fabric in folder ki by the visible light sensor, denoted as data A;

[0082] Type B files include image acquisition data of the corresponding fabric in the ki folder from a near-infrared channel sensor, denoted as Data B;

[0083] Class C files include image acquisition data of the corresponding fabric in the ki folder from the near-infrared channel dual sensors, denoted as data C;

[0084] Type d files include image acquisition data of the corresponding fabric in the ki folder from the polarization sensor, denoted as data D;

[0085] The optimal distance is obtained through the following steps: Under fixed light source conditions, the distance between the sensor and the fabric is gradually adjusted, and image quality indicators (e.g., reflectance coefficient, texture clarity, noise) are recorded at each distance. A comprehensive optimal distance is selected based on different spectral channels (visible light, near-infrared, polarized light). After determining the distance, multiple data acquisitions are performed to ensure that the quality of each acquired image is stable and that there is no obvious distortion or falsification. Fine-tuning is performed based on the acquired data, for example, adjusting the sensor distance according to changes in the reflectance spectrum, thereby achieving more refined image data.

[0086] The data processing unit is used to receive and process multi-channel datasets, construct an image feature analysis model based on the convolutional neural network (CNN) algorithm to extract four-dimensional data feature vector groups. The input of the image feature analysis model is the multi-channel dataset, and the output of the image feature analysis model is four types of feature vector groups. The four types of feature vector groups are integrated to generate a multi-channel vector dataset and sent to the component prediction unit.

[0087] Preprocessing of multichannel datasets includes the following steps:

[0088] Data A, data B, data C, and data D are preprocessed respectively, including: normalization, noise reduction, data augmentation, and channel independence processing;

[0089] The data enhancements include rotation, translation, scaling, and brightness adjustment;

[0090] The channel preprocessing includes:

[0091] Band differential processing is performed on near-infrared channel one and near-infrared channel two to suppress dye absorption interference;

[0092] Calculate Stokes vectors I, Q, U, and V from polarization image data and generate images of polarization degree and polarization angle;

[0093] Perform white balance and color constancy correction on visible light images;

[0094] Constructing an image feature analysis model, extracting four types of feature vector groups, and generating a multi-channel vector dataset includes the following steps:

[0095] Based on the image data in folders a, b, c, and d within folder ki of each fabric in the preprocessed multi-channel dataset, an image feature extraction model is trained.

[0096] Define the input format of the image feature extraction model as F0 = {Fa, Fb, Fc, Fd}; define the output format of the image feature extraction model as F1 = {k, y, e, p}, where k, y, e, and p are all vector groups.

[0097] Wherein, k is a set of visible parameter vectors, including color intensity, chroma, saturation, brightness, and spatial distribution features of texture;

[0098] The y is a set of parameter vectors, including reflectance, spectral reflectance curve, hygroscopicity, and fiber composition characteristics;

[0099] The term e is a set of two types of parameter vectors, including reflection intensity and dye absorption characteristics;

[0100] p is a set of polarization difference parameter vectors, including degree of polarization, polarization angle, and surface texture features;

[0101] Among them, the visible parameter vector group k, the first type parameter vector group y, the second type parameter vector group e, and the polarization difference parameter vector group p together form the feature vector group Fki, where i is the number of the fabric corresponding to this group of parameters.

[0102] Generate a multi-channel vector dataset, including the preprocessed ki group folder and the corresponding feature vector group;

[0103] Composition prediction unit: It is used to receive multi-channel vector datasets and sequentially pass through the composition mapping model construction module and the full-image composition distribution prediction module to complete the prediction of the overall composition category and proportion of the fabric.

[0104] The component mapping model construction module, based on a multi-task learning network, integrates image features from different channels in the multi-channel vector dataset, extracts component categories and their corresponding proportions within the sampling interval, generates sampling analysis data, and sends it to the full-image component distribution prediction module.

[0105] The component mapping model construction module generates sampling analysis data, including the following steps:

[0106] Based on a multi-task learning network structure, the input is a multi-channel vector dataset corresponding to the sampling region. The fused representation vector is generated through a channel-independent encoder and a feature fusion module. These vectors are then input into the component category recognition branch and the proportional regression branch, respectively. The output is the principal component category of the region and its proportion, forming the sampling analysis data.

[0107] The full-image component distribution prediction module calculates the main texture direction vector Xo based on the sampled analysis data using visible light images and polarization images, where o is the sampling point number; it constructs a sliding window sequence along this vector direction and extracts the texture features of each window region;

[0108] Subsequently, the similarity between the texture features of these regions and the features of the sampled regions is calculated, and regions similar to the features of the sampled regions are selected to form a set of similar regions J. A mask Mj is generated based on the set J, where j is the number of the similar regions. The main orientation vector Xo of the non-sampled regions and its mask Mj information are predicted by the orientation field interpolation method, thereby inferring the composition categories and their proportional relationships of the entire image.

[0109] By fusing visible light, near-infrared and polarization channel data and introducing a texture orientation inference mechanism, spatial continuity modeling of regional components and full-image component prediction are achieved, effectively improving the accuracy and robustness of component recognition, significantly reducing the data processing burden and required computing resources, and improving system response speed and engineering application feasibility.

[0110] Based on this, the feature vectors of neighboring sampling points are combined to perform weighted component prediction on the non-sampling area. Finally, the prediction results are evaluated for error to ensure the accuracy and reliability of the prediction.

[0111] The formula for calculating the similarity is as follows:

[0112] Calculate directional consistency using the cosine similarity formula:

[0113] ,

[0114] Where θ is the angle between the vectors corresponding to the texture features and the sampling region features;

[0115] Similarity calculation formula:

[0116] ;

[0117] in, T represents the similarity between the sliding window and the monitoring area for the sampling point. h Let h be the texture vector of the sliding window, and h represent the h-th sliding window. The weighting of fabric composition and orientation. The Mahalanobis distance;

[0118] If the similarity is higher than the preset value Xs, the regions are considered to have the same orientation. All regions that continuously meet this condition are marked on the image, and a set D of continuous regions with the same orientation is constructed. The preset value Xs is derived from a large number of experiments.

[0119] Example: Xo = (1, 2, 3), Tj = (2, 4, 6), then sim = 1.0; =0.2, =5, then =0.333;

[0120] Label the regions within set D, including component categories and proportions, and generate a mask Mj;

[0121] Based on the orientation field interpolation function, an orientation propagation prediction model is established to predict the main orientation vector Xo and its mask Mj in the non-sampling area, thereby obtaining the orientation field of the whole image, and then completing the prediction of the composition category and proportion relationship of the fabric as a whole.

[0122] For non-sampled regions, the predicted components are weighted using the vector set Fki of the nearest neighboring sampled points, and the error is evaluated through cross-validation.

[0123] Obtaining the main directional vector Xo includes the following steps:

[0124] The vector representing the average direction of the main fiber texture within the sampling area is defined as vector Xo, where o is the sampling point number. Vector Xo is used to divide the sampling area into local texture regions using data A and data D. The structure tensor of each texture region is calculated to obtain the local dominant direction vector Yi, where i is the texture region number. All vectors Yi are combined and weighted to obtain vector Xo within the sampling area.

[0125] The step size selection of the sliding window includes the following steps:

[0126] Based on the sampling analysis data, the fibers of all component categories within the sampling area are first identified, and the component category with the shortest average length of a single fiber is selected. The maximum fiber length within that component category is used as a benchmark, and the step size of the sliding window is set to 5 times that maximum value.

[0127] The size of the interval and threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data; as long as it does not affect the ratio between the parameter and the quantized value.

[0128] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0129] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fabric analysis system driven by multi-channel image data, characterized in that, It includes an image sampling unit, a data processing unit, and a component prediction unit; The image sampling unit is used to set up a sensor sampling detection belt on the fabric to be detected, and set up a sensor group at the detection belt to collect multi-dimensional information of the fabric. After preprocessing the multi-dimensional information, a multi-channel dataset is generated and sent to the data processing unit. The data processing unit is used to receive and process multi-channel datasets, construct an image feature analysis model based on the convolutional neural network (CNN) algorithm to extract four-dimensional data feature vector groups. The input of the image feature analysis model is the multi-channel dataset, and the output of the image feature analysis model is four types of feature vector groups. The four types of feature vector groups are integrated to generate a multi-channel vector dataset and sent to the component prediction unit. Composition prediction unit: It is used to receive multi-channel vector datasets and sequentially pass through the composition mapping model construction module and the full-image composition distribution prediction module to complete the prediction of the overall composition category and proportion of the fabric. The process by which the component prediction unit completes the prediction of the component distribution of the entire map includes the following steps: Based on the sampled analysis data, the main texture direction vector Xo is calculated using visible light images and polarization images, where o is the sampling point number; a sliding window sequence is constructed along this vector direction, and the texture features of each window region are extracted; Subsequently, the similarity between the texture features of these regions and the features of the sampled regions is calculated, and regions similar to the features of the sampled regions are selected to form a set of similar regions J. A mask Mj is generated based on the set J, where j is the number of the similar regions. The main texture orientation vector Xo of the non-sampled regions and its mask Mj information are predicted by the orientation field interpolation method, thereby inferring the composition categories and their proportional relationships of the entire image. Based on this, the feature vectors of neighboring sampling points are combined to perform weighted component prediction on the non-sampling area. Finally, the prediction results are evaluated for error to ensure the accuracy and reliability of the prediction. Calculate directional consistency using the cosine similarity formula: ， Where θ is the angle between the vectors corresponding to the texture features and the sampling region features; Similarity calculation formula: ； in, T represents the similarity between the sliding window and the monitoring area of ​​the sampling point. h Let h be the texture vector of the sliding window, and h represent the h-th sliding window. The weighting of fabric composition and orientation. The Mahalanobis distance; The component mapping model construction module, based on a multi-task learning network, integrates image features from different channels in the multi-channel vector dataset, extracts component categories and their corresponding proportions within the sampling interval, generates sampling analysis data, and sends it to the full-image component distribution prediction module. The full-image component distribution prediction module, based on sampling analysis data, uses visible light images and polarization images to calculate the fiber texture direction vector at the sampling point, and predicts the similarity between the main components and the sampling area in the direction of the vector, thereby completing the overall prediction of the fabric.

2. The fabric analysis system driven by multi-channel image data according to claim 1, characterized in that, The image sampling unit generates a multi-channel dataset, including the following steps: The sensor group includes a visible light sensor, a near-infrared channel one sensor, a near-infrared channel two sensor, and a polarized light imaging sensor. The sampling detection band includes horizontal sampling bands and vertical sampling bands, and each type of sampling band includes several sampling points; and the setting of the sampling bands on the fabric to be tested depends on the analysis results of the data collected by the visible light sensor; The sampling band is set up as follows: based on the visible light image, the longest straight line on the fabric surface is extracted as the horizontal main axis, and sampling points are set at equal intervals along this axis to form a horizontal sampling band; the vertical sampling band is set up in the same way; the extension range of the sampling band is determined based on the optimal detection distance of each channel sensor. Generating a multi-channel dataset includes the following steps: The visible light sensors, with a spectral range of 400-700nm, are arranged in a ring around the fabric to be tested. The distance between each visible light sensor and the fabric is set to the optimal distance to ensure the accuracy of data acquisition. At each of the sampling points, a set of near-infrared channel one sensor, near-infrared channel two sensor and polarized light imaging sensor are provided; The near-infrared channel one sensor has a spectral range of 900-1100nm; the near-infrared channel two sensor has a spectral range of 1450-1550nm; the polarized light imaging uses specific polarization angles, including 0°, 45°, 90°, and 135°; and the distance between each sensor and the fabric is set to the optimal distance to ensure data acquisition accuracy. Define folders ki = i = 1, 2, 3, ..., n, where n is the total number of fabrics to be detected and i is the fabric number to be detected. Each ki folder contains four types of folders: a, b, c, and d. Define the set of all ki folders as a multi-channel dataset. Among them, folder A contains image acquisition data of the corresponding fabric in folder ki by the visible light sensor, denoted as data A; Type B files include image acquisition data of the corresponding fabric in the ki folder from a near-infrared channel sensor, denoted as Data B; Class C files include image acquisition data of the corresponding fabric in the ki folder from the near-infrared channel dual sensors, denoted as data C; The d-type files contain image acquisition data of the corresponding fabric in the ki folder from the polarization light sensor, denoted as data D.

3. The fabric analysis system driven by multi-channel image data according to claim 2, characterized in that, The data processing unit preprocesses the multi-channel dataset, including the following steps: Data A, data B, data C, and data D are preprocessed respectively, including: normalization, noise reduction, data augmentation, and channel independence processing; The data enhancements include rotation, translation, scaling, and brightness adjustment; The preprocessing of the multi-channel dataset includes: Band differential processing is performed on near-infrared channel one and near-infrared channel two to suppress dye absorption interference; Calculate Stokes vectors I, Q, U, and V from polarization image data and generate images of polarization degree and polarization angle; Perform white balance and color constancy correction on visible light images.

4. The fabric analysis system driven by multi-channel image data according to claim 3, characterized in that, Constructing an image feature analysis model, extracting four types of feature vector groups, and generating a multi-channel vector dataset includes the following steps: Based on the image data in folders a, b, c, and d within folder ki of each fabric in the preprocessed multi-channel dataset, an image feature extraction model is trained. Define the input format of the image feature extraction model as F0={Fa, Fb, Fc, Fd}; define the output format of the image feature extraction model as F1={k, y, e, p}, where k, y, e, and p are all vector groups; Wherein, k is a set of visible parameter vectors, including color intensity, chroma, saturation, brightness, and spatial distribution features of texture; The y is a set of parameter vectors, including reflectance, spectral reflectance curve, hygroscopicity, and fiber composition characteristics; The term e is a set of two types of parameter vectors, including reflection intensity and dye absorption characteristics; p is a set of polarization difference parameter vectors, including degree of polarization, polarization angle, and surface texture features; Among them, the visible parameter vector group k, the first type parameter vector group y, the second type parameter vector group e, and the polarization difference parameter vector group p together form the feature vector group Fki, where i is the fabric number to be detected corresponding to this group of parameters. Generate a multi-channel vector dataset, including the preprocessed ki group folder and the corresponding feature vector group.

5. The fabric analysis system driven by multi-channel image data according to claim 4, characterized in that, The component mapping model construction module generates sampling analysis data, including the following steps: Based on a multi-task learning network structure, the input is a multi-channel vector dataset corresponding to the sampling region. The fused representation vector is generated through a channel-independent encoder and a feature fusion module. These vectors are then input into the component category recognition branch and the proportional regression branch, respectively. The output is the principal component category of the region and its proportion, forming the sampling analysis data.

6. The fabric analysis system driven by multi-channel image data according to claim 5, characterized in that, Obtaining the main texture orientation vector Xo includes the following steps: The vector representing the average direction of the main fiber texture within the sampling area is defined as the main texture direction vector Xo, where o is the sampling point number. The main texture direction vector Xo is used to divide the sampling area into local texture regions using data A and data D. The structure tensor of each texture region is calculated to obtain the local dominant direction vector Yr, where r is the texture region number. All local dominant direction vectors Yr are combined and weighted to obtain the main texture direction vector Xo within the sampling area.

7. The fabric analysis system driven by multi-channel image data according to claim 6, characterized in that, The step size selection of the sliding window includes the following steps: Based on the sampling analysis data, the fibers of all component categories within the sampling area are first identified, and the component category with the shortest average length of a single fiber is selected. The maximum fiber length within that component category is used as a benchmark, and the step size of the sliding window is set to 5 times that maximum value.

Images