Textile defect detection device and textile defect detection method

Abstract

The invention relates to the technical field of textile detection. The invention discloses a textile defect detection device and a textile defect detection method.The textile defect detection device comprises a gradient light absorption assembly below a textile, the gradient light absorption assembly is composed of a light absorption bottom layer and a reflective porous adjusting surface layer, and the porosity of the surface layer is distributed in a gradient mode in the detection direction so as to amplify local changes of penetrating light. And a pulse polarized light source group is arranged above and emits polarized light to form reflected light and penetrating light. And a hyperspectrum-light intensity fusion detection device is arranged above the side, is separated from the light source and is used for collecting spectrum and light intensity distribution. A partition dynamic tension mechanism is arranged on one side of the textile and fixes and controls the distance between the partition dynamic tension mechanism and the gradient light absorption assembly. And a multi-modal data processing unit is arranged on the side close to the detection device, is connected with the detection device, is used for pre-storing a dynamic reference library and is used for identifying flaws and quantifying edge gradients. According to the method, the textile defect detection accuracy and adaptability are improved through multi-dimensional innovation. The gradient light absorption assembly amplifies the light intensity change of the flaws, and the recognition precision of the tiny flaws is improved to be within 0.3 mm.

Description

Technical Field

[0001] This invention relates to the field of textile testing technology, and in particular to a textile defect detection device and a textile defect detection method. Background Technology

[0002] In the current field of textile defect detection, traditional manual inspection still accounts for a certain proportion. However, it relies on the experience and visual acuity of the inspectors, which has significant limitations. Prolonged operation can easily lead to worker fatigue, resulting in increased rates of missed and false detections, especially for tiny defects less than 1mm in diameter or deep fiber defects. At the same time, manual inspection is inefficient, requiring multiple inspectors per production line, and the inspection speed cannot keep up with the high-speed production pace of modern textile industry, severely restricting the improvement of production efficiency and the stability of quality control.

[0003] While existing optical inspection technologies have replaced manual inspection to some extent, they still face numerous technical bottlenecks. Most devices employ a single wavelength light source or a fixed polarization direction, making it difficult to simultaneously address the detection needs of different types of defects in textiles, such as surface stains and deep yarn unevenness, resulting in limited ability to identify complex defects. Furthermore, traditional optical devices lack sufficient amplification of defect signals. When textiles contain low-contrast defects, the intensity changes of transmitted or reflected light are subtle, making it difficult for the detection device to capture effective signals, leading to significant missed detections. Simultaneously, differences in textile material and fluctuations in surface smoothness can easily cause optical signal interference, further reducing detection accuracy.

[0004] Existing detection methods generally lack dynamic adaptation mechanisms, making it difficult to cope with diverse detection scenarios. Some devices fail to consider the stability of the distance between textiles and detection components, or lack dynamic benchmark models for different materials, resulting in detection results being significantly affected by environmental factors and fabric characteristics. Furthermore, most technologies rely solely on single-spectrum or light intensity signals for analysis, lacking multi-dimensional data fusion, making it difficult to accurately distinguish and classify defect types. This hinders the provision of quantitative evidence for production quality control and restricts the development of refined quality management in the textile industry. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention discloses a textile defect detection device capable of detecting minute defects in textiles, improving detection accuracy, and increasing detection speed.

[0006] This invention discloses a textile defect detection device, which includes:

[0007] The gradient light-absorbing component is located under the textile. The gradient light-absorbing component consists of a bottom layer with light-absorbing effect and a porous adjustment surface layer with reflective effect. The porous adjustment surface layer has several pores, and the porosity of the porous adjustment surface layer is gradient distributed along the detection direction. It is used to amplify the local changes in the intensity of transmitted light through the difference in porosity.

[0008] A pulsed polarization light source group placed above the textile can emit polarized light. Part of the polarized light is reflected by the textile to form reflected light, and the other part passes through the textile to form transmitted light.

[0009] The hyperspectral-intensity fusion detection device is located above the textile and on opposite sides of the pulsed polarization light source. It is used to collect the spectrum and intensity distribution of the light reflected from the textile surface after the transmitted light is reflected by the porous adjustable surface layer.

[0010] A zoned dynamic tension mechanism is installed on one side of the textile to fix the textile and control the distance between the textile and the gradient light-absorbing component.

[0011] The multimodal data processing unit is located on the side of the textile close to the hyperspectral-intensity fusion detection device and connected to the hyperspectral-intensity fusion detection device. The multimodal data processing unit pre-stores a dynamic benchmark library based on flawless textile samples, which is used for comparative analysis and identification of defective areas of textiles and quantification of their edge gradients.

[0012] When textiles have defects, the porosity difference in different areas of the surface layer is adjusted by the porous structure, which causes the amount of light absorbed by the underlying layer to be different, thus resulting in different changes in the intensity of reflected light on the surface layer.

[0013] Furthermore, the bottom layer of the gradient light absorption component is a high-absorbency material layer with an absorbency of ≥95%;

[0014] The porous surface layer has pore sizes in the micrometer range, and the porosity gradient distribution along the detection direction ranges from 10% to 50%.

[0015] The bottom layer uses a nano carbon black coating or black ceramic material, and the porous adjustment surface layer uses an electrochromic material or an optical grade polymer. The thickness of the porous adjustment surface layer is 50-200μm.

[0016] Furthermore, the pulsed polarization light source group can emit polarized light of at least two wavelengths, has pulse modulation function and polarization direction switching function, and the light source intensity can be dynamically adjusted;

[0017] The pulsed polarization light source emits polarized light including 500-550nm green light and 600-650nm red light, with a pulse frequency of 10-50Hz and the polarization direction can be switched alternately from 0° to 90°.

[0018] Furthermore, the hyperspectral-light intensity fusion detection device includes a hyperspectral camera and a light intensity sensor array, which are set at an angle of 30°-60° to the surface of the textile to simultaneously collect spectral and light intensity distribution data of reflected light and transmitted light;

[0019] The hyperspectral camera has a spectral range of 400-1000nm, the light intensity sensor array has a detection accuracy of ≤0.1lux, and both have a sampling frequency of ≥100fps.

[0020] Furthermore, the zoned dynamic tension mechanism consists of clamping units driven independently by multiple servo motors, which can adjust the tension of the textile in zones, so that the distance between the textile and the gradient light-absorbing component is controlled at 1-5mm and the flatness deviation is <100μm.

[0021] The multimodal data processing unit fuses spectral features and light intensity deviations through a convolutional neural network, achieving a gradient quantization accuracy of ≤0.3mm for defect edges.

[0022] This invention discloses a method for detecting textile defects, which can be used in any of the textile defect detection devices mentioned above. S1: Construct a dynamic benchmark library, fix a defect-free textile sample through a partitioned dynamic tension mechanism, irradiate and collect reflection spectrum and light intensity distribution data using a pulsed polarization light source group, and establish a benchmark feature model.

[0023] S2: Real-time spectrum and light intensity acquisition. Repeat the irradiation and acquisition operation of S1 on the textile to be tested to obtain real-time detection data.

[0024] S3: Defect feature analysis, which compares real-time data with dynamic benchmark library through multimodal data processing unit, and combines light intensity changes amplified by gradient light absorption component to identify suspected defect areas;

[0025] S4: Multi-dimensional verification involves repeatedly testing the suspected area switching detection parameters to verify the consistency of defect features;

[0026] S5: Output defect information, quantify the edge gradient of the defect area and output it in grades.

[0027] Furthermore, when constructing the dynamic benchmark library in S1, it is necessary to collect at least 30 sets of flawless sample data under different materials and environmental conditions, and extract common benchmark features through principal component analysis;

[0028] During real-time spectrum and light intensity acquisition in S2, the pulsed polarization light source group alternately emits polarized light of different wavelengths, and the hyperspectral-light intensity fusion detection device synchronously records the multi-band spectrum and light intensity distribution matrix.

[0029] The polarized light in S2 includes green and red light of different wavelengths. Green light is used to enhance the spectral contrast of surface defects, while red light is used to amplify the light intensity differences of deep fiber defects.

[0030] Furthermore, in S3, the defect feature analysis uses wavelet transform to denoise the light intensity data, extracts the contour features of the defect area through the edge detection algorithm, and distinguishes the defect type by combining spectral angle matching.

[0031] Furthermore, the multi-dimensional verification in S4 includes switching the wavelength and polarization direction of the pulse polarization light source group and adjusting the surface reflectivity of the gradient light absorption component. If the defect feature deviation of multiple tests is ≤5%, it is judged as a real defect.

[0032] The defect classification in S5 is based on light intensity deviation amplitude, spectral shift and edge gradient, and is divided into three levels: slight, medium and severe, with each level corresponding to a specific parameter threshold range.

[0033] Furthermore, the textiles are pretreated before testing, including surface cleaning and humidity control, to keep the moisture regain of the textiles at 8%-15% and reduce environmental interference.

[0034] The multimodal data processing unit uses a convolutional neural network to fuse features of real-time data, and the model training sample size is ≥1000 groups.

[0035] The beneficial effects of this invention are:

[0036] The textile defect detection device and method of this invention significantly improve the accuracy and adaptability of textile defect detection through multi-dimensional technological innovation. The porosity gradient design of the gradient light absorption component overcomes the limitation of weak defect signals in traditional detection methods. Its underlying high-absorbency material and porous, adjustable surface work synergistically to amplify the differences in light intensity in the defect area, improving the identification accuracy of minute defects to within 0.3 mm and solving the problem of missed detection of low-contrast defects in traditional methods. The multi-wavelength and polarization direction switching design of the pulsed polarization light source group, combined with a hyperspectral-intensity fusion detection device, achieves simultaneous capture of surface and deep defects. Green light enhances the spectral contrast of surface stains, while red light amplifies the light intensity differences of uneven yarn in the deep layers, improving accuracy compared to single-wavelength detection. The zoned dynamic tension mechanism controls the spacing and flatness through independent servo motors, adapting to a variety of fabrics from lightweight silk to heavy denim, avoiding detection deviations caused by material differences and broadening the application range of the device.

[0037] This invention further optimizes the textile inspection process with its high efficiency and reliability. The multimodal data processing unit integrates spectral and light intensity data, combined with a dynamic benchmark library built based on more than 30 flawless samples, and achieves feature matching through a convolutional neural network. The defect identification accuracy is ≥95%, reducing the false negative rate by 80% compared to manual inspection. The multi-dimensional verification and tiered output in the inspection method provide quantitative evidence for quality control. All components work collaboratively, with a sampling frequency ≥100fps, meeting the real-time inspection needs of high-speed production lines and improving efficiency by more than 10 times compared to manual fabric inspection. Simultaneously, preprocessing reduces environmental interference, ensuring long-term operational stability. In summary, this device not only overcomes the accuracy bottleneck of traditional inspection technologies but also considers the high efficiency and compatibility of industrial production, providing an integrated solution for textile quality inspection. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the process of a textile defect detection device according to an embodiment of this application.

[0039] Figure 2 This is a schematic diagram of a textile defect detection device according to an embodiment of this application.

[0040] Figure 3 This is another structural schematic diagram of a textile defect detection device according to an embodiment of this application.

[0041] Figure 4 This is a schematic diagram of a gradient light absorption component according to an embodiment of this application.

[0042] In the figure: gradient light absorption component 11, bottom layer 111, porous adjustment surface layer 112, pulsed polarization light source group 12, hyperspectral-light intensity fusion detection device 13, hyperspectral camera 131, light intensity sensor array 132, partitioned dynamic tension mechanism 14, multimodal data processing unit 15. Detailed Implementation

[0043] To enable those skilled in the art to better understand the present invention, the technical solutions in the specific embodiments of the present invention will be clearly and completely described below.

[0044] This invention discloses a textile defect detection device, which includes: a gradient light-absorbing component disposed below the textile, the gradient light-absorbing component being composed of a bottom layer with light-absorbing effect and a porous adjustment surface layer with reflective effect, the porous adjustment surface layer having a plurality of pores, the porosity of the porous adjustment surface layer being gradient distributed along the detection direction, used to amplify local changes in transmitted light intensity through porosity differences.

[0045] A pulsed polarization light source group placed above the textile can emit polarized light. Part of the polarized light is reflected by the textile to form reflected light, and the other part passes through the textile to form transmitted light.

[0046] The hyperspectral-intensity fusion detection device is located above the textile and on opposite sides of the pulsed polarization light source. It is used to collect the spectrum and intensity distribution of the light reflected from the textile surface after the transmitted light is reflected by the porous adjustable surface layer.

[0047] A zoned dynamic tension mechanism is installed on one side of the textile to fix the textile and control the distance between the textile and the gradient light-absorbing component.

[0048] The multimodal data processing unit is located on the side of the textile near the hyperspectral-intensity fusion detection device and connected to the device. The multimodal data processing unit pre-stores a dynamic benchmark library based on flawless textile samples for comparative analysis and identification of defective areas of the textile and quantification of their edge gradients. The edge gradient is the spatial transition width from the edge of the defect to the normal area, in mm, with a quantification accuracy ≤0.3 mm. The intensity gradient is the rate of change of intensity per unit distance at the edge of the defect, in lux / mm.

[0049] When textiles have defects, the porosity difference in different areas of the surface layer is adjusted by the porous structure, which causes the amount of light absorbed by the underlying layer to be different, thus resulting in different changes in the intensity of reflected light on the surface layer.

[0050] The gradient absorption component serves as the detection basis, with its bottom layer made of a highly absorbent material that efficiently absorbs light penetrating the textile. The porous adjustment surface has a porous structure with a gradient porosity distribution along the detection direction. When the textile is flawless, light transmission is uniform, and the difference in reflected light intensity between different porosity regions of the porous adjustment surface is small. When the textile has defects, more light passes through the defective areas. This extra light is more easily absorbed by the bottom layer in areas with high porosity, resulting in a significantly greater change in reflected light intensity in these areas compared to areas with low porosity, thus amplifying the defect signal. The pulsed polarized light source emits polarized light that obliquely illuminates the textile. Part of the light is reflected from the textile surface, while the other part passes through the textile as transmitted light. The hyperspectral-intensity fusion detection device is positioned on either side of the light source, synchronously acquiring the spectral and intensity data of the reflected and transmitted light after reflection through the porous adjustment surface at a specific angle. This ensures that the changes in light signals caused by both surface features and internal defects of the textile are captured simultaneously. The zoned dynamic tension mechanism secures the textile using independently driven clamping units, controlling the distance between the textile and the gradient light-absorbing components and maintaining the textile's flatness. This avoids light signal interference caused by textile wrinkles or spacing fluctuations, providing a benchmark for stable detection. The multimodal data processing unit compares the real-time acquired light signals with a pre-stored dynamic benchmark library of flawless samples, identifying defective areas by analyzing the magnitude of light intensity changes and spectral feature differences. Simultaneously, this unit calculates the rate of light signal change from the normal area to the defective area at the defect edge, quantifying the edge gradient and achieving a precise description of the defect morphology.

[0051] The porosity gradient design of the gradient light absorption component overcomes the limitation of weak defect signals in traditional detection methods. By amplifying the differences in light intensity variations across different regions, it improves the accuracy of identifying minute defects and solves the problem of missed detection of low-contrast defects using traditional methods. The multimodal data processing unit integrates spectral and light intensity data, combined with a dynamic benchmark library, to achieve multi-dimensional comparison. It can not only distinguish defect types but also classify defects through edge gradient quantization, significantly improving the accuracy of detection results compared to single-signal analysis. The zoned dynamic tension mechanism can adjust tension and spacing according to the textile material, adapting to a variety of fabrics from lightweight silk to heavy denim, avoiding detection bias caused by differences in fabric characteristics. All components of the textile defect detection device work collaboratively, automating the entire process from light source illumination to data output. The detection speed can reach over 30 frames per second, meeting the real-time detection needs of high-speed production lines, improving efficiency by more than 10 times compared to manual fabric inspection, while reducing the missed detection rate by 80%.

[0052] Light intensity only reflects changes in light strength and is suitable for detecting fluctuations in light signals caused by differences in light transmittance, but it cannot distinguish the type of defect or the material characteristics. Spectral data, on the other hand, provides much richer information. Furthermore, the base color and dye composition of textiles can affect the baseline value of the light intensity signal, making misjudgment easy when relying solely on light intensity. Spectral features can eliminate interference from the base color, allowing for precise location of defects by comparing the spectral morphology of normal and defective areas. The fusion of hyperspectral and light intensity data essentially combines intensity changes with wavelength characteristics, capturing the spatial location of defects through light intensity and differentiating defect types through the spectrum, ultimately improving detection accuracy.

[0053] The detection process centers on the modulation and analysis of optical signals. When polarized light emitted from a pulsed polarization source illuminates the textile, it splits into two signals: reflected light and transmitted light. The reflected light carries the surface characteristics of the textile, while the transmitted light is modulated by a gradient absorption component. When defects exist in the textile, the amount of transmitted light passing through the defective area increases. This extra light is more easily absorbed by the underlying layer in areas with high porosity in the porous surface layer, resulting in a significantly greater variation in reflected light intensity in these areas compared to other areas, thus amplifying the defect signal. After capturing the two optical signals using a hyperspectral-intensity fusion detection device, a multimodal data processing unit compares them with a dynamic benchmark library. Through comprehensive analysis of intensity deviation, spectral shift, and edge gradient, the defective area is accurately identified and its characteristics quantified. This process, through gradient amplification, multi-signal fusion, and dynamic calibration, achieves efficient and accurate detection of textile defects.

[0054] A flowchart of a textile defect detection device is shown below. Figure 1 As shown in the image.

[0055] In the process of textile defect detection, the average light intensity of defect-free samples of the same material and under the same environmental conditions in the dynamic benchmark library should be used as the benchmark value, and a clear range for judging light intensity fluctuation should be set: when using a pulsed polarized light source to emit 500-550nm green light to detect surface defects such as stains or surface fiber damage, if the real-time light intensity of a certain area of ​​the textile under test fluctuates by more than ±8% compared to the benchmark value; or when using 600-650nm red light to detect deep defects such as uneven yarn or deep fiber breakage, if the real-time light intensity fluctuates by more than ±12% compared to the benchmark value, then that area is marked as a suspected defect area. Subsequently, a multi-dimensional verification process is required, switching the polarization direction of the light source to alternate between 0° and 90° and adjusting the surface reflectivity of the gradient light absorption component. If the deviation of the light intensity fluctuation amplitude in that area is ≤5% in multiple tests, and combined with the spectral shift characteristics captured by the hyperspectral camera, it can be finally determined that there is a real defect in that area; if the light intensity fluctuation amplitude is always within the range of ±8% or ±12% of the corresponding wavelength, then the textile is determined to be defect-free and meets the quality standards.

[0056] A schematic diagram of the textile defect detection device is shown below. Figure 2 As shown in the diagram, the textile defect detection device includes a gradient light absorption component 11, a pulsed polarized light source group 12, a hyperspectral-intensity fusion detection device 13, a zoned dynamic tension mechanism 14, and a multimodal data processing unit 15. The pulsed polarized light source group 12 and the hyperspectral-intensity fusion detection device 13 can be adjusted in position and angle as needed.

[0057] Example 1: Defect detection in cotton fabric. In actual detection, there are slight fluctuations in the light source, sensor accuracy, and slight unevenness on the surface of the textile. The measured value has an error of within ±1.5% compared with the theoretical value.

[0058] Test subject: A piece of cotton fabric to be tested, which may have minor holes, local stains and other defects.

[0059] The gradient light absorption component is placed directly below the fabric. The bottom layer is a nano carbon black coating with a light absorption rate of 96% and almost no reflection. The porous adjustment surface is an optical-grade polyethylene porous plate with a thickness of 100μm. Small holes with a diameter of 5μm are evenly distributed on the plate. The porosity is gradient distributed from left to right along the detection direction, with 10% on the left and 50% on the right. The solid part of the porous adjustment surface can reflect 40% of the incident light.

[0060] The pulsed polarization light source group is installed on the upper left side of the fabric. It can emit 530nm green polarized light with a polarization direction of 0° and a pulse frequency of 30Hz. The light shines obliquely onto the fabric surface.

[0061] The hyperspectral-light intensity fusion detection device is installed on the upper right side of the fabric, on opposite sides of the light source, at a 45° angle to the fabric surface. It includes a hyperspectral camera with a detection spectral range of 400-800nm ​​and a light intensity sensor array with an accuracy of 0.05 lux, used to simultaneously capture the color characteristics and brightness changes of reflected light.

[0062] The partitioned dynamic tension mechanism clamps the two edges of the fabric and adjusts the tension through four servo motors to keep the fabric flat and maintain a stable distance of 2mm between it and the gradient light-absorbing component.

[0063] The multimodal data processing unit has a built-in dynamic benchmark library based on 100 flawless cotton fabric samples, which includes spectral curves and light intensity distribution models of normal areas. It connects to the detection device and analyzes the data in real time.

[0064] Step 1: The process of inspecting flawless cotton fabric

[0065] The 530nm green light emitted by the pulsed polarized light source shines on the flawless cotton fabric. Most of the light is reflected by the cotton fibers, forming surface reflection light. Only 15% of the light penetrates the cotton fabric and reaches the gradient light-absorbing component below.

[0066] After the transmitted light reaches the porous adjustment surface layer, it falls on the left side of the sparsely porous region with a porosity of 10%, where 90% is solid and 10% is micropores. 90% of the transmitted light is reflected by the solid layer of the porous adjustment surface layer, and 10% passes through the pores and is absorbed by the underlying layer. Therefore, the relative value of the reflected light intensity is 15% × 90% = 13.5%.

[0067] The light falls in the dense pore region on the right with a porosity of 50%. In this dense pore region, 50% of the porous adjustment surface layer is a solid part without pores and 50% is a small pore. 50% of the transmitted light is reflected by the porous adjustment surface solid layer and 50% passes through the pores and is absorbed by the bottom layer. At this time, the relative value of the reflected light intensity is 15% × 50% = 7.5%.

[0068] The test results showed that the light intensity measured by the light intensity sensor array was stable at 15 lux on the left and 8 lux on the right; the spectral curve collected by the hyperspectral camera was a perfect match with the green light spectrum of flawless cotton fabric in the dynamic reference library.

[0069] The multimodal data processing unit determines that there are no defects and updates the real-time fluctuation range of the dynamic benchmark library.

[0070] Step 2: Inspection process for cotton fabric containing tiny holes

[0071] There is a tiny hole with a diameter of 0.5mm in the middle of the fabric, which is difficult to detect with the naked eye.

[0072] When the light source shines on the perforated area, due to the absence of fiber obstruction, the light transmission rate increases from 15% to 60%, with an additional 45% penetrating before reaching the gradient light-absorbing component.

[0073] The light intensity falls within the sparsely porous region with a porosity of 10% on the left. At this point, the relative value of the reflected light intensity is 60% × 90% = 54%, which is 40.5% higher than the 13.5% of the normal region.

[0074] The light falls in the dense pore area with 50% porosity on the right side. At this time, the relative value of the reflected light intensity is 60% × 50% = 30%, which is 22.5% higher than the 7.5% of the normal area.

[0075] At this time, in terms of the corresponding absolute illumination intensity, the light intensity sensor array measured a sudden increase to 75 lux in the left region corresponding to the hole, a change of 60 lux from the normal 15 lux, while the light intensity in the right region increased to 40 lux, a change of 32 lux from the normal 8 lux. The hyperspectral camera found that although the main peak of the spectral curve of the hole region was still 530 nm, the peak light intensity was significantly increased.

[0076] The multimodal data processing unit compared real-time data with a dynamic benchmark library and found that the light intensity variation far exceeded the ±2 lux threshold, and the 60 lux variation on the left was much greater than the 32 lux variation on the right. This gradient difference amplification directly located the hole. Simultaneously, the rate of light intensity change from the normal area to the hole edge was calculated; for every 0.1 mm distance, the light intensity increased by 5 lux, the light intensity gradient was quantized to 50 lux / mm, and the edge gradient was quantized to 0.1 mm, thus identifying it as a genuine defect.

[0077] The light intensity sensor array possesses high spatial resolution, enabling it to capture the light intensity distribution over minute areas on the surface of textiles. When a hole is detected, the sensor records the continuous change in light intensity from the normal area to the hole area. At the edge of the hole, the light intensity increases by 5 lux at intervals of 0.1 mm. The multimodal data processing unit calculates the edge gradient by dividing the light intensity change by the distance: 5 lux ÷ 0.1 mm = 50 lux / mm, thus quantifying the rate of change in light intensity from the normal to the defective area.

[0078] The porosity of the gradient light-absorbing component exhibits a gradient distribution along the detection direction. This spatial distribution characteristic creates a fixed pattern in the reflection or absorption ratio of transmitted light at different locations. When a hole exists, the amount of transmitted light in the corresponding area increases significantly, and this increased transmitted light covers a certain spatial range of the gradient light-absorbing component. In other words, the location of the hole on the fabric corresponds to a continuous area of ​​the gradient light-absorbing component below. The light intensity sensor array is an array structure with spatial resolution, capable of acquiring the light intensity distribution at different locations on the textile surface in real time. In the gradient light-absorbing component area corresponding to the hole, the change in reflected light intensity in the low-porosity area on the left is much greater than that in the high-porosity area on the right. This more significant gradient difference on the left forms a local anomaly region in the spatial data of the sensor array, representing the location where the light intensity change is most drastic and conforms to the porosity gradient pattern. Simultaneously, the increase in the light intensity peak in the hole area in the spectral data acquired by the hyperspectral camera corresponds perfectly in spatial location to the anomaly region of the sensor array, further verifying that this area is where the hole is located. The multimodal data processing unit can accurately locate the hole by matching the spatial distribution of light intensity gradient differences with the location of spectral anomalies.

[0079] In one implementation, the bottom layer of the gradient absorption component is a high-absorbency material layer with an absorbency ≥95%. The pore size of the porous adjustment surface layer is in the micrometer range, and the porosity gradient distribution along the detection direction ranges from 10% to 50%. The bottom layer uses a nano-carbon black coating or black ceramic material, while the porous adjustment surface layer uses an electrochromic material or an optical-grade polymer, with a thickness of 50-200 μm.

[0080] The bottom layer of the gradient light absorption component uses a high-absorbency material with an absorption rate of 95% or higher. This material efficiently absorbs light passing through the pores of the porous adjustment surface, providing a stable low-background basis for optical signal comparison. The bottom layer material can be a nano-carbon black coating or black ceramic material; both materials possess excellent light absorption performance, and the choice can be made based on the specific application requirements. The pore size of the porous adjustment surface is in the micrometer range. This micrometer-scale design ensures effective light transmission through the pores while also allowing for optical signal modulation through pore distribution. Its porosity is gradient-distributed along the detection direction in the range of 10%-50%, resulting in a regular difference in the absorption and reflection ratios of transmitted light in different areas, creating conditions for amplifying defect signals. The porous adjustment surface uses either an electrochromic material or an optical-grade polymer; both materials have their advantages. Electrochromic materials can have their reflection characteristics altered through external control, while optical-grade polymers possess stable physicochemical properties. The thickness of the porous adjustment layer is controlled between 50-200 μm. This thickness range ensures structural stability while achieving an ideal balance between light reflection and transmission within the porous adjustment layer. The bottom layer and the porous adjustment layer work together. The strong light absorption properties of the bottom layer ensure that the light signal at the pores is effectively absorbed, while the porosity gradient distribution of the porous adjustment layer causes significant differences in the transmitted light variation in the defect area at different locations, jointly achieving precise amplification of the defect signal.

[0081] The bottom layer, made of a highly absorbent material with an absorption rate ≥95%, minimizes interference from reflected light, making the difference in light signals between normal and defective areas more prominent. This enhances the contrast in defect identification, allowing even minute defects to be clearly captured. The micron-level pore size design of the porous adjustment surface adapts to the dimensional characteristics of textile fibers, avoiding light signal distortion caused by excessively large or small pore sizes. This ensures that the ratio of light transmission to reflection meets detection requirements, providing a foundation for accurate detection. A porosity gradient distribution of 10%-50% generates different signal responses for defects of varying degrees, giving the textile defect detection device excellent detection sensitivity for defects of various types and sizes, expanding the detection range. The selection of electrochromic materials or optical-grade polymers, along with thickness control of 50-200μm, gives the porous adjustment surface excellent optical performance and structural stability, allowing it to adapt to long-term industrial testing environments, reducing maintenance costs, and extending component lifespan.

[0082] The detection principle of the gradient light-absorbing component is based on the synergistic effect of the bottom layer and the porous adjustment surface layer. When light shines on the textile and penetrates the gradient light-absorbing component, the highly absorbent material of the bottom layer absorbs the light passing through the pores of the porous adjustment surface layer, while the solid part of the porous adjustment surface layer reflects some of the light. Since the porosity of the porous adjustment surface layer exhibits a gradient distribution of 10%-50% along the detection direction, the amount of light transmitted to the defective area increases when defects are present in the textile. In areas with high porosity, more transmitted light passes through the pores and is absorbed by the bottom layer, resulting in a greater change in reflected light intensity in these areas compared to areas with low porosity, thus amplifying the defect signal. This signal amplification achieved through the porosity gradient distribution, combined with the high light-absorbing performance of the bottom layer, allows the detection device to more easily capture changes in the light signal caused by defects, providing a clear and reliable basis for subsequent data analysis and defect identification.

[0083] In one implementation, the pulsed polarization light source group can emit polarized light of at least two wavelengths, has pulse modulation and polarization direction switching functions, and the light source intensity can be dynamically adjusted. The polarized light emitted by the pulsed polarization light source group includes 500-550nm green light and 600-650nm red light, with a pulse frequency of 10-50Hz and a polarization direction that can be alternately switched from 0° to 90°.

[0084] The pulsed polarization light source assembly incorporates a multi-wavelength emission module, capable of emitting at least two wavelengths of polarized light, including 500-550nm green light and 600-650nm red light. The differences in the characteristics of different wavelengths meet diverse detection needs. The pulse modulation module converts continuous light into pulsed light, with the pulse frequency adjustable within the 10-50Hz range, resulting in a periodic change in the light signal output by the light source, facilitating synchronous signal acquisition by the detection device. The polarization direction switching function is implemented by a polarization controller, which can control the polarized light to alternate between 0° and 90°, enhancing the ability to identify different types of defects by changing the direction of light vibration. Light source intensity adjustment is achieved through a power control unit, dynamically adjusting the output light intensity according to changes in textile material, color, and the detection environment, ensuring the stability and applicability of the light signal during detection. The emission of at least two wavelengths of polarized light, especially 500-550nm green light and 600-650nm red light, enhances the contrast of the light signal for surface defects in textiles, while red light highlights deep fiber defects, improving the detection capability for defects of different depths and types. The polarization direction alternates between 0° and 90°, adapting to different fiber orientations in textiles and allowing for differentiated light-fiber interactions. This results in clearer exposure of uneven fiber arrangement defects, increasing the comprehensiveness of the detection. The light source intensity is dynamically adjustable, providing suitable light intensity for different textiles, such as light and dark colors, thin and thick fabrics. This avoids signal saturation due to excessive light intensity or signal blurring due to insufficient light intensity, ensuring detection accuracy.

[0085] As one implementation method, the hyperspectral-intensity fusion detection device includes a hyperspectral camera and an intensity sensor array, positioned at an angle of 30°-60° to the surface of the textile, simultaneously acquiring spectral and intensity distribution data of reflected and transmitted light. The hyperspectral camera has a spectral range of 400-1000 nm, the intensity sensor array has a detection accuracy of ≤0.1 lux, and both have a sampling frequency of ≥100 fps.

[0086] The hyperspectral-intensity fusion detection device integrates a hyperspectral camera and an intensity sensor array, which work together to collect spectral data and intensity distribution data, respectively. The device is positioned at a 30°-60° angle to the textile surface. This angle avoids interference from direct light while efficiently capturing signals from reflected and transmitted light after reflection by the gradient absorption components, ensuring the representativeness of the acquired light signals. The hyperspectral camera covers a spectral range of 400-1000nm, capturing the characteristics of reflected and transmitted light at different wavelengths, providing spectral basis for distinguishing different types of defects. The intensity sensor array achieves a detection accuracy of ≤0.1 lux, capable of sensing minute changes in intensity. Simultaneously, both the hyperspectral camera and the intensity sensor array have sampling frequencies ≥100fps, ensuring real-time synchronous data acquisition and meeting rapid detection requirements.

[0087] The hyperspectral-intensity fusion detection device is aimed at the textile at a specific angle. When a pulsed polarized light source illuminates the textile, some light is reflected from the textile surface, while some penetrates the textile and is reflected by a gradient absorption component. The hyperspectral camera in the device captures the spectral information of both types of light. By analyzing the spectral characteristics at different wavelengths, it distinguishes between normal and defective areas of the textile. For example, stains can cause enhanced absorption of specific wavelengths of light, and fiber defects can cause changes in spectral reflectance. The intensity sensor array simultaneously collects the intensity distribution data of both types of light. Because defective areas cause different intensity variations compared to normal areas, and these variations are amplified by the gradient absorption component, the sensors can accurately capture these intensity differences. The spectral and intensity data are simultaneously transmitted to a multimodal data processing unit, where they are compared and analyzed against a dynamic benchmark library, ultimately enabling the identification and location of defective areas in the textile.

[0088] In one implementation, the zoned dynamic tension mechanism consists of clamping units driven independently by multiple servo motors, which can adjust the tension of the textile in zones, controlling the distance between the textile and the gradient light-absorbing component to 1-5mm, with a flatness deviation of <100μm. The multimodal data processing unit fuses spectral features and light intensity deviations through a convolutional neural network, achieving a gradient quantization accuracy of ≤0.3mm for defect edges.

[0089] The zoned dynamic tension mechanism consists of multiple clamping units, each driven by an independent servo motor. Through precise control of the servo motors, each clamping unit can apply different tensions to different areas of the textile, achieving zoned adjustment. This adjustment can stably control the distance between the textile and the gradient light-absorbing component within 1-5 mm, while ensuring the textile's flatness deviation is less than 100 μm, providing reliable spatial conditions for stable propagation and acquisition of light signals.

[0090] When the zoned dynamic tension mechanism is in operation, each servo motor drives its corresponding clamping unit to apply appropriate tension to different areas based on the material, thickness, and other characteristics of the textile. Through this zoned adjustment, the textile is flattened and maintains a stable distance of 1-5mm from the gradient light-absorbing components, avoiding problems such as light signal refraction and scattering caused by unevenness or improper spacing of the textile, ensuring a stable path and intensity of the light signal during propagation. After receiving the spectral features and intensity deviations transmitted from the hyperspectral-intensity fusion detection device, the multimodal data processing unit uses a convolutional neural network to process this data layer by layer. By learning the feature patterns of flawless samples, the network can quickly identify abnormal information that differs from normal features; these abnormalities correspond to potential defective areas.

[0091] This invention discloses a method for detecting defects in textiles, comprising:

[0092] S1: Construct a dynamic benchmark library, fix flawless textile samples through a partitioned dynamic tension mechanism, irradiate and collect reflection spectrum and light intensity distribution data using a pulsed polarization light source group, and establish a benchmark feature model;

[0093] S2: Real-time spectrum and light intensity acquisition. Repeat the irradiation and acquisition operation of S1 on the textile to be tested to obtain real-time detection data.

[0094] S3: Defect feature analysis, which compares real-time data with dynamic benchmark library through multimodal data processing unit, and combines light intensity changes amplified by gradient light absorption component to identify suspected defect areas;

[0095] S4: Multi-dimensional verification involves repeatedly testing the suspected area switching detection parameters to verify the consistency of defect features;

[0096] S5: Output defect information, quantify the edge gradient of the defect area and output it in grades.

[0097] When constructing the dynamic benchmark library, the S1 uses a partitioned dynamic tension mechanism to fix the flawless textile samples, keeping them flat and maintaining a stable distance from the gradient light absorption components. A pulsed polarization light source group illuminates the samples according to set parameters, and a hyperspectral-intensity fusion detection device simultaneously acquires reflectance spectrum and intensity distribution data. A multimodal data processing unit analyzes this data, extracts common features, and establishes a benchmark feature model as a reference standard for subsequent testing.

[0098] In the S2 real-time spectrum and light intensity acquisition stage, the textile under test is fixed in the same way as in S1, with the same light source parameters and acquisition method. The zoned dynamic tension mechanism ensures that the textile under test is in the same detection state as the sample. After irradiation by the pulsed polarization light source group, the hyperspectral-light intensity fusion detection device acquires real-time spectral and light intensity data, providing a basis for subsequent comparison.

[0099] In the S3 defect feature analysis, the multimodal data processing unit compares the real-time detection data with the benchmark feature models in the dynamic benchmark library, calculating the differences between the two in terms of spectral characteristics, light intensity distribution, etc. Simultaneously, by combining the light intensity changes amplified by the gradient absorption component, areas with significant deviations from normal features are screened out and marked as suspected defect areas.

[0100] During S4 multi-dimensional verification, for suspected defect areas, the parameters of the pulsed polarization light source group, such as wavelength, polarization direction, and pulse frequency, or the reflectivity of the gradient absorption component, are adjusted, and spectral and light intensity acquisition is repeated. The multimodal data processing unit analyzes the multiple detection results; if the defect characteristics remain consistent, the area is confirmed as a true defect area. In the S5 defect information output stage, the multimodal data processing unit quantifies the edge gradient of the defect area and classifies the defect according to a preset grading standard, finally outputting detection results containing information such as defect location, type, and grade.

[0101] The textile defect detection method uses a dynamic benchmark library as a reference and achieves defect detection through comparative analysis. In S1, the spectral and light intensity data of the defect-free sample under standardized conditions are recorded and constructed into a benchmark model, reflecting the light signal characteristics of normal textiles. S2 collects real-time data of the textile under test, including light signal changes caused by possible defects. In S3, the multimodal data processing unit compares the real-time data with the benchmark model. Because the gradient absorption component amplifies the light intensity changes in the defect area, the difference between the light signal in the defect area and the normal area becomes more significant, thus accurately identifying suspected defect areas. S4 repeats the detection by switching detection parameters to further verify whether the characteristics of the suspected area are stable and eliminates misjudgments caused by random factors. S5 quantifies and classifies the defects based on features such as edge gradients, and the final output comprehensively reflects the defect status of the textile, achieving accurate detection and evaluation of textile defects.

[0102] As one implementation method, when constructing the dynamic benchmark library in S1, at least 30 sets of flawless sample data under different materials and environmental conditions need to be collected, and common benchmark features are extracted through principal component analysis. In S2, during real-time spectral and intensity acquisition, a pulsed polarized light source group alternately emits polarized light of different wavelengths, and a hyperspectral-intensity fusion detection device simultaneously records multi-band spectra and intensity distribution matrices. The polarized light of different wavelengths in S2 includes green and red light; green light is used to enhance the spectral contrast of porous surface defects, and red light is used to amplify the intensity differences of deep fiber defects.

[0103] When constructing the dynamic benchmark library, S1 selects at least 30 sets of flawless textile samples, covering different materials such as cotton, linen, silk, and synthetic fibers. Simultaneously, different environmental conditions are simulated, such as adjusting parameters like temperature and humidity. These samples are fixed using a partitioned dynamic tension mechanism, and under uniform detection settings, illuminated by a pulsed polarization light source. A hyperspectral-intensity fusion detection device collects the reflectance spectrum and intensity distribution data of each sample set. The multimodal data processing unit performs principal component analysis on this data to extract common spectral and intensity characteristics of different samples, thereby constructing a common benchmark feature model.

[0104] In the S2 real-time spectrum and intensity acquisition stage, the pulsed polarized light source group alternately emits polarized light of different wavelengths, namely green and red light, according to a set program. Simultaneously, the hyperspectral-intensity fusion detection device works to record multi-band spectral data and intensity distribution matrix reflected from the textile, forming a complete real-time detection data set. Specifically, the green light emission targets the textile surface. When green light illuminates surface defects, it interacts with the defects, making the spectral contrast of the surface defects more obvious, facilitating the hyperspectral camera to capture their features. Red light, on the other hand, penetrates the textile surface more easily, reaching deeper fibers. When fiber defects exist in the deeper layers, the intensity of the red light changes accordingly, amplifying the intensity differences of the deep fiber defects, allowing the intensity sensor array to accurately detect them.

[0105] As one implementation method, in S3, the defect feature analysis uses wavelet transform to denoise the light intensity data, extracts the contour features of the defect area through the edge detection algorithm, and distinguishes the defect type by combining spectral angle matching.

[0106] When performing defect feature analysis, S3 first applies wavelet transform to the collected light intensity data. Wavelet transform decomposes the light intensity data into different frequency components. By retaining the effective signal frequencies related to defects and filtering out irrelevant high-frequency or low-frequency noise such as environmental interference and equipment noise, the denoised light intensity data is obtained. Subsequently, an edge detection algorithm is used to process the denoised light intensity data. This algorithm identifies regions of abrupt changes in light intensity, determines the boundary between defective and normal regions, and then extracts the contour features of the defective region, such as shape and size. Simultaneously, a spectral angle matching method is used to compare the spectral features of the defective region with those of known defect types. Spectral angle matching calculates the angle between different spectra; the smaller the angle, the more similar the spectral features, thus distinguishing different types of defects such as stains, fiber breaks, and uneven yarn. Throughout the process, wavelet transform, edge detection algorithm, and spectral angle matching work together to gradually complete a comprehensive analysis of defect features, from denoising and contour extraction to type differentiation.

[0107] Wavelet transform denoising is based on the difference between noise and effective signals in the frequency domain. Light intensity changes caused by defects have a specific frequency range, while noise is usually distributed in other frequency ranges. Wavelet transform decomposes the light intensity data into different frequency channels, eliminating components from the noise channels and retaining components from the effective signal channels, thus achieving denoising and making the defect signal more prominent. The core of edge detection algorithms is identifying abrupt changes in light intensity. In textiles, there is a significant difference in light intensity between defective and normal areas, manifesting as abrupt changes in light intensity at the boundaries. Edge detection algorithms calculate the gradient change in light intensity, finding points where the gradient value exceeds a threshold. The curve formed by connecting these points is the outline of the defective area, thus extracting the defect outline. Spectral angle matching distinguishes defect types based on the unique spectral characteristics of different types of defects. Each defect, due to differences in composition and structure, has different light reflection and absorption characteristics, resulting in a specific shape for its spectral curve. By calculating the angle between the spectrum of the defect to be detected and the spectrum of known defect types, the smaller the angle, the greater the probability that they belong to the same type, thus achieving accurate differentiation of defect types. These three methods work together to first purify the data, then extract the contours, and finally determine the type, gradually and thoroughly analyzing the characteristics of defects and providing a scientific and accurate basis for identifying suspected defect areas.

[0108] As one implementation method, the multi-dimensional verification in S4 includes switching the wavelength and polarization direction of the pulsed polarization light source group and adjusting the surface reflectivity of the gradient light absorption component. If the defect feature deviation of multiple tests is ≤5%, it is judged as a real defect. The defect classification in S5 is based on the light intensity deviation amplitude, spectral shift, and edge gradient, and is divided into three levels: slight, moderate, and severe, with each level corresponding to a specific parameter threshold range.

[0109] During multi-dimensional verification using S4, for suspected defect areas, the wavelength type of the pulsed polarization light source group is first switched, alternating between green and red light. Simultaneously, the polarization direction is changed, alternating between 0° and 90°. Furthermore, the surface reflectivity of the gradient absorption component is adjusted by changing the voltage of the electrochromic material or selecting optical-grade polymers with different reflectivities, allowing the surface reflectivity to vary within a set range. After each parameter adjustment, the hyperspectral-intensity fusion detection device re-acquires data, and the multimodal data processing unit compares the defect characteristics from multiple detections. If the deviation is ≤5%, it is determined to be a genuine defect.

[0110] When S5 performs defect grading, the multimodal data processing unit first calculates the light intensity deviation amplitude of the defective area, i.e., the proportion of the light intensity difference between the defective area and the normal area in the dynamic benchmark library. Then, it analyzes the spectral shift, i.e., the degree of wavelength shift between the spectral curve of the defective area and the benchmark spectral curve. Simultaneously, combined with the quantified edge gradient, and based on preset parameter threshold ranges, defects are divided into three levels: minor, moderate, and severe. Each level corresponds to a clearly defined light intensity deviation amplitude, spectral shift, and edge gradient range, ensuring a consistent grading standard.

[0111] The principle of S4 multi-dimensional verification is based on the stability of the optical characteristics of genuine defects, while the characteristics of fake defects fluctuate significantly with changes in detection parameters. When switching polarization types, green light is sensitive to surface defects, while red light is sensitive to deep defects. If a suspected area exhibits similar characteristics under both types of light, it is likely a genuine defect. Changing the polarization direction affects the interaction between light and fibers. The intensity and spectral variation patterns of genuine defects are consistent across different polarization directions, while fake defects do not exhibit this pattern. When adjusting the surface reflectivity of the gradient absorption component, the intensity variation of genuine defects changes proportionally to the reflectivity and the trend is stable, while the intensity variation of fake defects is irregular. When the defect feature deviation from multiple detections is ≤5%, it indicates that the characteristics are not affected by parameter adjustments and conform to the optical characteristics of genuine defects; therefore, it can be determined as a genuine defect.

[0112] As one implementation method, the textiles are pretreated before testing, including surface cleaning and humidity adjustment, to maintain the moisture regain rate of the textiles at 8%-15% and reduce environmental interference. The multimodal data processing unit uses a convolutional neural network to fuse features from real-time data, with a model training sample size of ≥1000 groups.

[0113] Before testing, textiles undergo pretreatment, starting with surface cleaning. Dust, fiber debris, and other impurities are removed using dust removal equipment to prevent these from being misidentified as defects. Humidity is then regulated using temperature and humidity control equipment to maintain the textile's moisture regain at 8%-15%. Moisture regain is monitored in real-time using a humidity sensor, ensuring it reaches the preset range before proceeding to the testing stage. The multimodal data processing unit employs convolutional neural networks for feature fusion. At least 1000 sets of textile sample data containing different types and degrees of defects are collected, encompassing spectral characteristics and light intensity distribution, among other multidimensional information. This sample data is then input into the convolutional neural network for training. Through multi-layer neural network computation, the model gradually learns the characteristic patterns of different defects, ultimately achieving a defect identification accuracy of ≥95%.

[0114] Example 2: Detection of stains on the surface of silk fabric

[0115] Test subject: Lightweight silk fabric, approximately 0.05mm thick, with surface stains such as oil and water stains with a diameter of 1-2mm, which are difficult to see with the naked eye.

[0116] The bottom layer of the gradient absorption component is a nano carbon black coating with an absorption rate of 96%. The porous adjustment surface layer is optical grade polyethylene with a thickness of 50 μm, a pore diameter of 3 μm, and a porosity gradient of 10%-30% along the detection direction.

[0117] The pulsed polarized light source group alternately emits 530nm green light and 630nm red light, with a pulse frequency of 30Hz, switching polarization direction from 0° to 90°, and the light intensity is set to the low level.

[0118] The hyperspectral-intensity fusion detection device is positioned at a 45° angle to the fabric, featuring a hyperspectral camera with a detection accuracy of 400-800nm, an intensity sensor array with an accuracy of 0.05lux, and a sampling frequency of 100fps.

[0119] The zoned dynamic tension mechanism controls the distance between the fabric and the gradient light-absorbing components to 1mm, with a flatness deviation of <50μm.

[0120] Thirty groups of flawless silk samples from different batches were collected, and a benchmark model was established at 25℃ and 50% humidity. The reflectance spectral characteristics of the 530nm green light band were extracted.

[0121] When the light source emits 530nm green light, the light reflectivity of the stained area decreases due to the reduced grease / moisture content, which is 15%-20% lower than that of the normal area, thus increasing the amount of light transmitted.

[0122] The high porosity region of the gradient light absorption component absorbs twice as much transmitted light as the normal region, and the light intensity sensor measured a light intensity deviation of 5 lux in this region.

[0123] After the multimodal unit is denoised by wavelet transform, the edge detection algorithm extracts the stain outline. Spectral angle matching shows that the stain area has a spectral angle of >15° in the 530nm band, and is determined to be a surface stain.

[0124] Repeat the test by switching the polarization direction to 90°. The stain feature deviation is less than 3%, confirming its authenticity. The light intensity gradient quantization is 30 lux / mm, and the edge gradient is 0.2mm, which is judged as a minor defect according to the standard.

[0125] The detection results show a 100% accuracy rate in stain recognition, a lower false negative rate compared to traditional optical detection, and a detection speed of 30 frames per second.

[0126] Example 3: Detection of uneven yarn distribution in the deep layers of denim

[0127] The tested material was a thick denim fabric with a thickness of 1.2mm. It showed uneven yarn density in some areas, with deep fiber stacking / sparseness, and no obvious traces on the surface.

[0128] The bottom layer of the gradient light absorption component is black ceramic with an absorbency of 95%, and the porous adjustment surface layer is an electrochromic material with a thickness of 200μm, a pore size of 10μm, and a porosity gradient of 30%-50%.

[0129] The pulsed polarized light source group mainly uses 630nm red light, with a pulse frequency of 50Hz, a polarization direction of 0°, and a light intensity set to the highest level.

[0130] The hyperspectral-intensity fusion detection device is positioned at a 60° angle to the fabric. The detection accuracy of the hyperspectral camera is 600-1000nm, the accuracy of the intensity sensor array is 0.1lux, and the sampling frequency is 100fps.

[0131] The zoned dynamic tension mechanism controls the distance between the fabric and components to 5mm, with the tension adjusted to a medium-high level.

[0132] Thirty flawless denim samples of different shades were collected, and a baseline model was established at 28℃ and 60% humidity to extract the transmitted light intensity distribution characteristics in the 630nm red light band.

[0133] When red light penetrates denim, the amount of light transmitted increases by 30%-50% in areas with sparse yarns. In areas with 50% porosity in the porous surface layer, the amount of transmitted light absorbed by the bottom layer is 40% higher than in normal areas, resulting in a 2-3 lux decrease in reflected light intensity.

[0134] The multimodal unit integrates spectral data and light intensity distribution, and identifies uneven yarn regions through a convolutional neural network, with edge gradient quantization to 0.2 mm.

[0135] Switch the polarization direction to 90°, repeat the test and the deviation is <2%, confirming authenticity; based on the magnitude of the light intensity deviation, it is judged as a medium defect.

[0136] Successfully identifies deep yarn unevenness with a detection accuracy of 0.3mm, and is compatible with denim production lines with a speed of 10m / min.

[0137] Example 4: Detection of fiber breakage in wool fabric

[0138] The test subject was a coarse wool fabric with a thickness of 1mm. It contained tiny clumps formed by local fiber breakage. The clumps were 0.5-1mm in diameter and had a loose internal structure.

[0139] Device parameter settings

[0140] The bottom layer of the gradient light absorption component is black ceramic with an absorbency of 95%, and the porous adjustment surface layer is an electrochromic material with a thickness of 100μm, a pore diameter of 5μm, and a porosity gradient of 20%-50%.

[0141] The pulsed polarized light source group mainly uses 650nm red light, with a pulse frequency of 20Hz, a polarization direction of 0°, and a light intensity set to medium.

[0142] The hyperspectral-intensity fusion detection device is positioned at a 30° angle to the fabric. The hyperspectral camera has a detection accuracy of 500-1000nm and a sampling frequency of 100fps.

[0143] The zoned dynamic tension mechanism controls the distance between the fabric and components to 3mm, with a flatness deviation of <80μm.

[0144] Testing process and results

[0145] Thirty flawless wool samples were collected, and their spectral characteristics in the 650nm red light band were extracted.

[0146] Due to the loose structure, the red light penetration in the fiber fractured and clumped area increased by 25%. In the porous adjustment surface layer with a porosity of 30%-50%, the change in reflected light intensity was 35% higher than that in the normal area, and the light intensity sensor captured a local light intensity surge of 1.5-2 lux.

[0147] Spectral angle matching showed that the spectral angle of the agglomerate region in the 650nm band was >10°. The edge detection algorithm extracted the agglomerate contour, and the edge gradient quantization was 0.25mm.

[0148] Adjust the light source intensity by ±20%, and the repeated detection deviation is <4%, confirming authenticity; based on the edge gradient and spectral offset, it is determined to be a minor defect.

[0149] It achieves a 98% recognition rate for tiny lumps, reduces the false negative rate by 80% compared to manual detection, and has a detection speed of 30 frames per second.

[0150] Common conclusions of the examples

[0151] The gradient light absorption component amplifies the signal by increasing the light intensity variation in the defective area by 2-4 times through the porosity gradient design, thus solving the problem of weak signal from tiny defects.

[0152] Advantages of multimodal data fusion: Combining spectral and light intensity data improves the accuracy of detection by 20%-30% compared to single signal detection.

[0153] With its zoned dynamic tension mechanism and adjustable light source parameters, it can be adapted to a variety of fabrics, from lightweight silk to heavy wool, with a detection accuracy of less than 0.3mm.

[0154] With a detection speed of ≥100fps, it meets the real-time detection needs of industrial production lines and improves efficiency by more than 10 times compared to manual detection.

[0155] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A textile defect detection device, characterized in that, include: The gradient light-absorbing component is located under the textile. The gradient light-absorbing component consists of a bottom layer with light-absorbing effect and a porous adjustment surface layer with reflective effect. The porous adjustment surface layer has several pores, and the porosity of the porous adjustment surface layer is gradient distributed along the detection direction. It is used to amplify the local changes in the intensity of transmitted light through the difference in porosity. A pulsed polarization light source group placed above the textile can emit polarized light. Part of the polarized light is reflected by the textile to form reflected light, and the other part passes through the textile to form transmitted light. The hyperspectral-intensity fusion detection device is located above the textile and on opposite sides of the pulsed polarization light source. It is used to collect the spectrum and intensity distribution of the light reflected from the textile surface after the transmitted light is reflected by the porous adjustable surface layer. A zoned dynamic tension mechanism is installed on one side of the textile to fix the textile and control the distance between the textile and the gradient light-absorbing component. The multimodal data processing unit is located on the side of the textile close to the hyperspectral-intensity fusion detection device and connected to the hyperspectral-intensity fusion detection device. The multimodal data processing unit pre-stores a dynamic benchmark library based on flawless textile samples, which is used for comparative analysis and identification of defective areas of textiles and quantification of their edge gradients. When textiles have defects, the porosity difference in different areas of the surface layer is adjusted by the porous structure, which causes the amount of light absorbed by the underlying layer to be different, thus resulting in different changes in the intensity of reflected light on the surface layer.

2. The textile defect detection device according to claim 1, characterized in that: The bottom layer of the gradient light absorption component is a high-absorbency material layer with an absorbency of ≥95%; The porous surface layer has pore sizes in the micrometer range, and the porosity gradient distribution along the detection direction ranges from 10% to 50%. The bottom layer uses a nano carbon black coating or black ceramic material, and the porous adjustment surface layer uses an electrochromic material or an optical grade polymer. The thickness of the porous adjustment surface layer is 50-200μm.

3. The textile defect detection device according to claim 1, characterized in that: The pulsed polarization light source group can emit polarized light of at least two wavelengths, and has pulse modulation and polarization direction switching functions. The light source intensity can be dynamically adjusted. The pulsed polarization light source emits polarized light including 500-550nm green light and 600-650nm red light, with a pulse frequency of 10-50Hz and the polarization direction can be switched alternately from 0° to 90°.

4. The textile defect detection device according to claim 1, characterized in that: The hyperspectral-light intensity fusion detection device includes a hyperspectral camera and a light intensity sensor array, which are set at an angle of 30°-60° to the surface of the textile to simultaneously collect spectral and light intensity distribution data of reflected and transmitted light. The hyperspectral camera has a spectral range of 400-1000nm, the light intensity sensor array has a detection accuracy of ≤0.1lux, and both have a sampling frequency of ≥100fps.

5. The textile defect detection device according to claim 1, characterized in that: The zoned dynamic tension mechanism consists of clamping units driven independently by multiple servo motors. It can adjust the tension of textiles in zones, so that the distance between the textiles and the gradient light-absorbing components is controlled at 2-5mm and the flatness deviation is <100μm. The multimodal data processing unit fuses spectral features and light intensity deviations through a convolutional neural network, achieving a gradient quantization accuracy of ≤0.3mm for defect edges.

6. A method for detecting defects in textiles, used in any one of the textile defect detection devices according to claims 1-5, characterized in that, include: S1: Construct a dynamic benchmark library, fix flawless textile samples through a partitioned dynamic tension mechanism, irradiate and collect reflection spectrum and light intensity distribution data using a pulsed polarization light source group, and establish a benchmark feature model; S2: Real-time spectrum and light intensity acquisition. Repeat the irradiation and acquisition operation of S1 on the textile to be tested to obtain real-time detection data. S3: Defect feature analysis, which compares real-time data with dynamic benchmark library through multimodal data processing unit, and combines light intensity changes amplified by gradient light absorption component to identify suspected defect areas; S4: Multi-dimensional verification involves repeatedly testing the suspected area switching detection parameters to verify the consistency of defect features; S5: Output defect information, quantify the edge gradient of the defect area and output it in grades.

7. The method for detecting defects in textiles according to claim 6, characterized in that: When building a dynamic benchmark library in S1, at least 30 sets of flawless sample data with different materials and environmental conditions need to be collected, and common benchmark features are extracted through principal component analysis. During real-time spectrum and light intensity acquisition in S2, the pulsed polarization light source group alternately emits polarized light of different wavelengths, and the hyperspectral-light intensity fusion detection device synchronously records the multi-band spectrum and light intensity distribution matrix. The polarized light in S2 includes green and red light of different wavelengths. Green light is used to enhance the spectral contrast of surface defects, while red light is used to amplify the light intensity differences of deep fiber defects.

8. The method for detecting defects in textiles according to claim 6, characterized in that: In S3, the defect feature analysis uses wavelet transform to denoise the light intensity data, extracts the contour features of the defect area through the edge detection algorithm, and distinguishes the defect type by combining spectral angle matching.

9. The method for detecting defects in textiles according to claim 6, characterized in that: The multi-dimensional verification in S4 includes switching the wavelength and polarization direction of the pulse polarization light source group and adjusting the surface reflectivity of the gradient light absorption component. If the defect feature deviation of multiple tests is ≤5%, it is judged as a real defect. The defect classification in S5 is based on light intensity deviation amplitude, spectral shift and edge gradient, and is divided into three levels: slight, medium and severe, with each level corresponding to a specific parameter threshold range.

10. The method for detecting defects in textiles according to claim 6, characterized in that: Before testing, textiles are pretreated, including surface cleaning and humidity control, to keep the moisture regain rate of textiles at 8%-15% and reduce environmental interference. The multimodal data processing unit uses a convolutional neural network to fuse features of real-time data, and the model training sample size is ≥1000 groups.

Images