Online visual detection method and device for flaws of high-speed moving cloth

Abstract

The invention provides an online visual detection method and device for flaws of high-speed moving cloth, and relates to the technical field of cloth flaw detection.The online visual detection method comprises the steps that firstly, a response diagram is obtained through time-sharing exposure, a normalized surface gradient modulus is calculated through a normalized difference ratio and logarithm compression, and color albedo interference is eliminated; curvature features are extracted by calculating the absolute value of a Hessian matrix and the square of a Laplacian operator and combined to generate a mixed curvature tensor, and tiny deformation is accurately captured; then manifold unsupervised reconstruction is carried out based on a Gaussian weighted model of a hollowed-out center pixel to obtain an ideal reconstruction tensor, and sample-free adaptive prediction is realized; difference is calculated, a local signal-to-noise ratio abnormal response index is obtained through local statistical window standard deviation normalization, and texture roughness interference is removed; and finally, positioning flaws, carrying out double-threshold judgment by calculating a flaw structure characteristic value and a flaw oil stain characteristic value, and outputting flaw coordinates and categories.

Description

Technical Field

[0001] This invention relates to the field of fabric defect detection technology, specifically to an online visual detection method and device for defects in high-speed moving fabrics. Background Technology

[0002] As a crucial pillar of traditional manufacturing, the textile printing and dyeing industry is gradually transforming and upgrading towards automation and intelligence. With the widespread adoption of modern high-speed looms, the operating speed of fabric production lines has significantly increased, placing stringent demands on real-time monitoring of weaving quality. Traditional fabric defect detection primarily relies on manual visual inspection. However, in high-speed motion scenarios, manual inspection is not only labor-intensive but also prone to missed or false detections due to visual fatigue. In recent years, machine vision inspection systems, which deeply integrate computer vision technology with optical imaging hardware, have gradually become a key means of replacing manual labor and ensuring the quality of finished fabrics. This technology aims to continuously scan and image high-speed moving fabrics using a line scan camera and automatically identify defects such as holes, knots, and oil stains using image processing algorithms. It is a core technological link in achieving "zero-defect" production in the textile industry.

[0003] In the prior art, CN104949990A discloses an online defect detection method for machine-made textiles, which employs a multi-point independent data acquisition followed by centralized unified judgment. The method first acquires real-time image status data of the target using acquisition terminals deployed on-site. Then, all front-end data is aggregated and transmitted to a back-end processing center via a network transmission protocol. Finally, the processing center uses a pre-set fixed logic algorithm to perform unified analysis on the received data to determine whether the target exhibits anomalies and trigger an alarm. This method primarily relies on standardized data flow paths and centralized judgment logic to maintain production supervision.

[0004] However, the aforementioned existing technologies have significant limitations when dealing with the inspection of complex high-end fabrics. First, the single-illumination imaging mode leads to photometric information coupling, making it impossible to physically distinguish between color changes and structural changes. This results in physical defects of the same color, such as cotton knots of the same color as the fabric, being missed due to indistinct grayscale differences, or normal dark texture patterns being misjudged as oil stains. Second, supervised learning-based algorithms suffer from severe sample dependence. In cold-start scenarios with diverse fabric types and unpredictable defect morphologies, effective detection is difficult without a large number of negative samples. Finally, existing algorithms are mostly based on single-dimensional alarm logic, only outputting a binary judgment of whether a defect exists, lacking the ability to decouple and classify the physical attributes of defects, making it difficult to meet the data requirements for targeted optimization of back-end production processes.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide an online visual inspection method and apparatus for defects in high-speed moving fabrics, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: An online visual inspection method for defects in high-speed moving fabrics, comprising the following steps: Using a line frequency synchronization control acquisition device for time-division exposure, side light response map and top light response map are acquired respectively. The color albedo component of the fabric surface is eliminated by normalized difference ratio including numerical gain coefficient. The dynamic range is compressed by combining logarithmic function, and the normalized surface gradient modulus characterizing the physical unevenness information of the fabric surface is calculated. Second-order differential analysis was performed on the normalized surface gradient modulus. The Gaussian curvature modulus was extracted by calculating the absolute value of the Hessian matrix, and the average curvature energy was extracted by calculating the square of the Laplacian operator. The two were then combined in a weighted linear manner to obtain the mixed curvature tensor. Based on the hybrid curvature tensor, and using the unsupervised reconstruction logic of the manifold, the Gaussian weighted spatial neighborhood prediction model of the hollowed-out center pixel is used to deduce the theoretical texture value of the current pixel using the texture features of the surrounding neighborhood, and the ideal reconstruction tensor is calculated. The absolute value of the difference between the hybrid curvature tensor and the ideal reconstruction tensor is calculated, and the difference is normalized using the feature standard deviation within the local statistical window to obtain the local signal-to-noise ratio anomaly response index after removing background texture roughness interference. The local signal-to-noise ratio anomaly response index is compared with a preset response index threshold. When the local signal-to-noise ratio anomaly response index exceeds the preset response index threshold, the fabric is determined to have defects and the defect coordinates are located. When a defect is determined to exist in the fabric, the defect's connected component is extracted using the defect coordinates. Based on this connected component, the defect's structural feature value and the defect's oil stain feature value are calculated. The defect's structural feature value is compared with a preset defect structure threshold, and the defect's oil stain feature value is compared with the defect's oil stain threshold. Based on the comparison results, the physical coordinate region and category of the defect are output.

[0008] Furthermore, by using a line scan camera in conjunction with a rotary encoder mounted on the conveying mechanism, the continuous fabric surface is divided into alternating optical-physical channels in the time dimension. The preset counter number is updated in real time by detecting the pulse signal output by the encoder. The initial value of the counter number is 0. When the fabric to be detected moves forward a preset distance, the encoder outputs a pulse signal to increment the counter number. When the counter number is odd, the line scan camera is triggered to expose, and the low-angle side light source is turned on and the top light source is turned off. The odd-numbered rows of the acquired fabric image are marked as structured light data. When the counter number is even, the line scan camera is triggered to expose, and the high-angle top light source is turned on and the top light source is turned off. The even-numbered rows of the acquired fabric image are marked as material light data. The structured light data is recombined to generate a single-pass grayscale image, which yields a side light response map that carries the shadow and highlight information of the fabric surface. The material light data is also recombined to generate a single-pass grayscale image, which yields a top light response map that carries the inherent reflectivity information of the fabric surface.

[0009] Furthermore, based on the side light response map and the top light response map, and based on the Lambert reflection model and the principle of photometric stereo vision, the fabric to be tested is desaturated, decoupled, and dynamically compressed to calculate the normalized surface gradient modulus that characterizes the physical unevenness of the fabric surface. Principle of normalized surface gradient modulus calculation:

[0010] Where x represents the x-coordinate, y represents the y-coordinate, and G... （x，y） Normalized surface gradient modulus at coordinates (x, y), IA （x，y） IB represents the grayscale value of the sidelight response map at coordinates (x, y). （x，y） θ represents the grayscale value of the top light response map at coordinates (x, y). A θ represents the incident angle of the side-lit light source. B denoted by , μ represents the incident angle of the top light source, k represents the preset numerical gain coefficient, k represents the preset luminous flux balance coefficient (obtained by whiteboard calibration), σ represents the preset thermal noise suppression factor (obtained by dark field calibration), and ε represents the preset regularization minimum value to prevent the denominator from being zero.

[0011] Furthermore, a second-order differential analysis is performed on the normalized surface gradient modulus. The Gaussian curvature modulus is extracted by calculating the absolute value of the Hessian matrix, the average curvature energy is extracted by calculating the square of the Laplacian operator, and the mixed curvature tensor for each coordinate is calculated. The principle of calculating the mixed curvature tensor for each coordinate:

[0012] Where, ψ (x,y) This represents the mixed curvature tensor at coordinates (x, y). Let represent the second partial derivative of the normalized surface gradient modulus at coordinates (x, y) along the horizontal axis. Let represent the second partial derivative of the normalized surface gradient modulus at coordinates (x, y) along the horizontal axis. The mixed partial derivative of the normalized surface gradient modulus at coordinates (x, y), where λ represents the preset shape weight coefficient.

[0013] Furthermore, based on the hybrid curvature tensor and the unsupervised reconstruction logic of the manifold, the Gaussian weighted spatial neighborhood prediction model of the hollowed-out center pixel is used to deduce the theoretical texture value of the current pixel using the texture features of the surrounding neighborhood, and the ideal reconstruction tensor is calculated. The principle of ideal reconstruction tensor calculation:

[0014] Where, ψ (x,y) Let W represent the ideal reconstruction tensor at coordinates (x, y). sum δ represents the weight normalization constant, i represents the increment of the horizontal axis, j represents the increment of the vertical axis, r represents the preset neighborhood window radius, τ represents the preset spatial correlation decay constant, and δ ij Let i represent the Kronecker function. When i=0 and j=0, the Kronecker function value is 1, otherwise the Kronecker function value is 0.

[0015] Furthermore, the absolute value of the difference between the hybrid curvature tensor and the ideal reconstruction tensor is calculated, and the difference is normalized using the feature standard deviation within the local statistical window to obtain the local signal-to-noise ratio anomaly response index after removing background texture roughness interference. Principle of calculating the local signal-to-noise ratio anomaly response index:

[0016] Among them, Ω （x，y） The local signal-to-noise ratio anomaly response index is represented at coordinates (x, y), D represents a local statistical window of a pre-defined rectangular region centered at coordinates (x, y), (p, q) represents the coordinate indices within the local statistical window, where p represents the x-coordinate index value, q represents the y-coordinate index value, and N represents the local statistical window. D ψ represents the total number of coordinate points within the local statistics window. DAVG η represents the average value of the mixed curvature tensor within the local statistical window, and η represents the preset base smoothing constant.

[0017] Furthermore, the local signal-to-noise ratio anomaly response index is compared with a preset response index threshold. If the local signal-to-noise ratio anomaly response index does not exceed the preset response index threshold, it is determined that the current fabric does not have a defect at coordinate (x, y). If the local signal-to-noise ratio anomaly response index exceeds the preset response index threshold, it is determined that the current fabric has a defect at coordinate (x, y), and the defect coordinates are output.

[0018] Furthermore, when it is determined that there is a defect in the fabric, the coordinates of the defect are obtained. The obtained defect coordinates are marked with connected components to obtain the physical coordinate range of the defect. For the mean of the local signal-to-noise ratio anomaly response index and the Weber contrast within the physical coordinate range of the defect, the defect structure feature value and the defect oil stain feature value are calculated respectively. Principle of Defect Structure Feature Value Calculation:

[0019] Among them, V S R represents the structural characteristic value of the defect, R represents the set of coordinates of the physical coordinate range of the defect, and N represents the structural characteristic value of the defect. R The total number of physical coordinate points representing the defect; Principle of calculating the characteristic value of oil stain defects:

[0020] Among them, V C The characteristic value of the defective oil stain, μ DE This represents the average grayscale value of all pixels within the physical coordinate range of the defect in the top light response map, in μ. BG This represents the average pixel grayscale value of a ring-shaped window one coordinate outside the physical coordinate range of the defect in the top light response map. The defect structure feature value is compared with the preset defect structure threshold. If the defect structure feature value does not exceed the preset defect structure threshold, the current defect is determined to be a structural defect. If the defect structure feature value exceeds the preset defect structure threshold, the current defect is determined to be a structural defect. The blemish oil stain feature value is compared with the blemish oil stain threshold. If the blemish oil stain feature value does not exceed the preset blemish oil stain threshold, the current blemish is determined to be a stain type blemish. If the blemish oil stain feature value exceeds the preset blemish oil stain threshold, the current blemish is determined to be a stain type blemish. Output the physical coordinates and category of the defect based on the comparison results.

[0021] A high-speed moving fabric defect online visual inspection device, the inspection device being used to implement the above-mentioned inspection method, comprising: Data acquisition and processing module: Used to use line frequency synchronization to control the acquisition device for time-division exposure, acquire side light response map and top light response map respectively, calculate and eliminate the color albedo component of the fabric surface by normalizing the difference ratio including numerical gain coefficient, and combine logarithmic function to compress dynamic range, calculate normalized surface gradient modulus characterizing the physical unevenness information of the fabric surface. Hybrid curvature tensor construction module: used to perform second-order differential analysis on normalized surface gradient modulus, extract Gaussian curvature modulus by calculating the absolute value of Hessian matrix, extract mean curvature energy by calculating the square of Laplacian operator, and calculate the weighted linear combination of the two to obtain hybrid curvature tensor. Manifold Unsupervised Reconstruction Module: Based on the hybrid curvature tensor and manifold unsupervised reconstruction logic, it uses a Gaussian weighted spatial neighborhood prediction model with the center pixel hollowed out, and uses the texture features of the surrounding neighborhood to deduce the theoretical texture value of the current pixel and calculate the ideal reconstruction tensor. Anomaly response index calculation module: used to calculate the absolute value of the difference between the hybrid curvature tensor and the ideal reconstruction tensor, and to normalize the difference using the feature standard deviation within the local statistical window, thereby calculating the local signal-to-noise ratio anomaly response index after removing background texture roughness interference. Defect Detection Module: Used to compare the local signal-to-noise ratio abnormal response index with the preset response index threshold. When the local signal-to-noise ratio abnormal response index exceeds the preset response index threshold, the fabric is determined to have defects and the defect coordinates are located. Defect Analysis Module: When a defect is determined to exist in the fabric, the module extracts the connected component of the defect using the defect coordinates, calculates the defect structure feature value and the defect oil stain feature value based on the connected component, compares the defect structure feature value with the preset defect structure threshold, compares the defect oil stain feature value with the defect oil stain threshold, and outputs the physical coordinate region and category of the defect based on the comparison results.

[0022] Compared with the prior art, the beneficial effects of the present invention are: This invention utilizes the time-division exposure mechanism of the acquisition device to obtain side-light response maps and top-light response maps respectively. It employs a normalized differential ratio calculation including a numerical gain coefficient to eliminate the color albedo component of the fabric surface. Furthermore, by combining this with a logarithmic function to compress the dynamic range, a normalized surface gradient modulus characterizing the physical unevenness of the fabric surface is calculated. This effectively shields the detection from interference by complex textures and monochromatic patterns at the imaging source. Further, second-order differential analysis is performed on the normalized surface gradient modulus. The absolute value of the Hessian matrix is ​​calculated to extract the Gaussian curvature modulus, and the square of the Laplacian operator is calculated to extract the average curvature energy. A weighted linear combination of these two values ​​yields a hybrid curvature tensor, which integrates the sensitivity of different curvature characteristics to point and linear physical defects, achieving precise capture of minute deformation features on the fabric surface.

[0023] This invention also utilizes a manifold unsupervised reconstruction logic and a Gaussian weighted spatial neighborhood prediction model with a hollowed-out center pixel. It uses the texture features of the surrounding neighborhood to deduce the theoretical texture value of the current pixel and calculate the ideal reconstruction tensor. By calculating the absolute value of the difference between the mixed curvature tensor and the ideal reconstruction tensor, and normalizing this difference using the feature standard deviation within a local statistical window, it calculates the local signal-to-noise ratio anomaly response index to remove background texture roughness interference, achieving adaptive detection even without sample training. After determining that the fabric has defects, it uses a hysteresis threshold segmentation algorithm to extract defect structural and color feature values. The defect structural feature values ​​are compared with a preset defect structural threshold to determine the structural type of the defect, and the defect color feature values ​​are compared with a defect color threshold to determine the color type of the defect. Based on the comparison results, it accurately outputs the physical coordinate region and category of the defect, solving the technical problem of traditional methods' difficulty in distinguishing between physical defects and color stains. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 This is a block diagram of the overall device module of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0026] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0027] Example: Please see Figure 1 The present invention provides a technical solution: An online visual inspection method for defects in high-speed moving fabrics, comprising the following steps: Step 1: Use the line frequency synchronization control acquisition device to perform time-division exposure, and acquire the side light response map and top light response map respectively. Calculate and eliminate the color albedo component of the fabric surface by normalizing the difference ratio containing the numerical gain coefficient, and combine the logarithmic function to compress the dynamic range, and calculate the normalized surface gradient modulus that characterizes the physical unevenness information of the fabric surface. In this embodiment, a line scan camera, in conjunction with a rotary encoder mounted on the conveying mechanism, is used to divide the continuous fabric surface into alternating optical-physical channels in the time dimension. The preset counter number is updated in real time by detecting the pulse signal output by the encoder. The initial value of the counter number is 0. When the fabric to be detected moves forward a preset distance, the encoder outputs a pulse signal to increment the counter number. When the counter number is odd, the line scan camera is triggered to expose, and the low-angle side light source is turned on and the top light source is turned off. The odd-numbered rows of the acquired fabric image are marked as structured light data. When the counter number is even, the line scan camera is triggered to expose, and the high-angle top light source is turned on and the top light source is turned off. The even-numbered rows of the acquired fabric image are marked as material light data. The structured light data is recombined to generate a single-pass grayscale image, which yields a side light response map that carries the shadow and highlight information of the fabric surface. The material light data is recombined to generate a single-pass grayscale image, which yields a top light response map that carries the inherent reflectivity information of the fabric surface. The main purpose of using a low-angle sidelight source is to utilize the shadowing effect of light to create strong shadows or highlights on the fabric surface's minute irregularities, such as knots and holes, under low-angle illumination, thereby highlighting physical structural features. The main purpose of using a high-angle toplight source is to obtain the fabric's albedo information, i.e., the fabric's color and printing pattern. In high-speed motion scenarios, if a single mixed exposure is used, the shadow information of physical defects will be submerged by the background color and texture, making it impossible to distinguish between black oil stains and black holes. This invention, through a time-division exposure mechanism with line-frequency synchronization, forcibly decouples structural and color information at the physical imaging source, providing an independent and pure data source for subsequent elimination of interference from backgrounds of the same color. This is the physical basis for achieving high-precision detection.

[0028] Based on the side light response map and top light response map, and based on the Lambert reflection model and the principle of photometric stereo vision, the fabric to be tested is desaturated, decoupled and dynamically compressed to calculate the normalized surface gradient modulus that characterizes the physical unevenness of the fabric surface. Principle of normalized surface gradient modulus calculation:

[0029] Where x represents the x-coordinate, y represents the y-coordinate, and G... （x，y） Normalized surface gradient modulus at coordinates (x, y), IA （x，y）IB represents the grayscale value of the sidelight response map at coordinates (x, y). （x，y） θ represents the grayscale value of the top light response map at coordinates (x, y). A θ represents the incident angle of the side-lit light source. B The angle of incidence of the top light source is represented by μ, which represents the preset numerical gain coefficient, set by relevant personnel according to the bit width design requirements of the target output data. k represents the preset luminous flux balance coefficient, obtained through whiteboard calibration. Before fabric inspection, a diffuse grayscale plate is placed on the conveyor belt, passing slowly through the camera's field of view. Side-light and top-light images are acquired according to the data acquisition process, and the average grayscale values ​​of the side-light and top-light images are calculated separately. The product of the average grayscale value of the side-light image and the cosine of the top light incident angle is calculated and recorded as the side-light group data. The product of the average grayscale value of the top-light image and the side-light incident angle is also calculated. The product of the cosine values ​​of the incident angle is recorded as the top light group data. The side light group data is divided by the top light group data to obtain the preset luminous flux balance coefficient. σ represents the preset thermal noise suppression factor, which is obtained through the dark field calibration method. Before fabric inspection, the camera lens cap is closed in an environment where all light sources are turned off. The same exposure time and gain as the actual inspection are set, and images are continuously acquired. The dark field standard deviation of the gray values ​​of all pixels in the continuously acquired images is calculated. Based on the typical industrial line scan camera settings, five times the calculated dark field standard deviation is used as the set thermal noise suppression factor. ε represents the preset regularization minimum value to prevent the denominator from being zero. By calculating the normalized surface gradient modulus, this study aims to construct a quantitative index that decouples color interference and deformation information from a photometric physical perspective. The magnitude of the normalized surface gradient modulus directly reflects the geometric tilt and physical unevenness of the fabric surface at the current coordinates. This means that the grayscale difference between the sidelight response map and the toplight response map is forcibly mapped to the surface height change rate, rather than the material's reflectivity. The normalized surface gradient modulus is positively correlated with the absolute value of the weighted difference between the sidelight response map and the toplight response map after correction for luminous flux balance coefficients. That is, as the physical unevenness of the fabric surface increases, the difference between the shadow produced by the sidelight response map and the inherent color produced by the toplight response map significantly increases, driving the normalized surface gradient modulus value to rise, thus accurately indicating the possibility of defects. Meanwhile, the normalized surface gradient modulus exhibits a non-linear normalization relationship with the total reflected energy, which is the sum of the side-light response map and the top-light response map. By dividing the light intensity difference by the total energy base, the influence of the fabric's color depth and printed patterns on the detection results is effectively eliminated, ensuring the consistency of detection standards across fabrics of different colors. Furthermore, to address the high dynamic range illumination changes that may occur with high-speed moving fabrics, the logarithmic function introduced in the formula exhibits a non-linear compression relationship with the light intensity difference. This suppresses signal overflow in highly reflective areas while preserving the sensitivity of small gradients in dark areas. The thermal noise suppression factor, negatively correlated with the background noise base, effectively shields the camera's dark current noise from interfering with the calculation of small gradients, acting as a threshold parameter for the logarithmic term. This calculation method, based on photometric stereo decoupling and dynamic range compression, is irreplaceable in high-speed online detection scenarios. It not only solves the technical challenge of distinguishing between physical defects of the same color and color textures in a single imaging mode but also significantly improves the system's detection stability under complex lighting and texture backgrounds through mathematical gain control and noise suppression.

[0030] Step 2: Perform second-order differential analysis on the normalized surface gradient modulus, extract the Gaussian curvature modulus by calculating the absolute value of the Hessian matrix, extract the average curvature energy by calculating the square of the Laplacian operator, and perform a weighted linear combination of the two to calculate the mixed curvature tensor. In this embodiment, a second-order differential analysis is performed on the normalized surface gradient modulus. The Gaussian curvature modulus is extracted by calculating the absolute value of the Hessian matrix, the average curvature energy is extracted by calculating the square of the Laplacian operator, and the mixed curvature tensor for each coordinate is calculated. The principle of calculating the mixed curvature tensor for each coordinate:

[0031] Where, ψ (x,y) This represents the mixed curvature tensor at coordinates (x, y). Let represent the second partial derivative of the normalized surface gradient modulus at coordinates (x, y) along the horizontal axis. Let represent the second partial derivative of the normalized surface gradient modulus at coordinates (x, y) along the horizontal axis. The mixed partial derivative of the normalized surface gradient modulus at coordinates (x, y), where λ represents the preset shape weight coefficient, which is preset by relevant staff based on the defect shape characteristics of the fabric to be inspected. By calculating the mixed curvature tensor, this study aims to establish a comprehensive geometric deformation index that can simultaneously characterize point-like abrupt changes and linear creases on the fabric surface. The magnitude of the mixed curvature tensor specifically reflects the sum of local curvature energy of the fabric surface at the current coordinates. This means that the second-order differential characteristic of the normalized surface gradient modulus is transformed into a scalar field measuring the degree of microscopic geometric topological anomalies on the surface. A higher mixed curvature tensor value indicates a greater probability of physical defects at that location. The absolute value of the determinant of the Hessian matrix formed by the mixed curvature tensor and the second-order partial derivatives of the normalized surface gradient modulus (i.e., the Gaussian curvature modulus) is positively correlated with the square of the Laplacian operator of the normalized surface gradient modulus (i.e., the mean curvature energy). The Gaussian curvature modulus term specifically responds to isotropic peak-valley point defects such as knots and holes, while the mean curvature energy term specifically responds to anisotropic ridge-valley linear defects such as creases and scratches. The Gaussian curvature modulus and mean curvature energy are weighted by a morphological weighting coefficient. Weighted harmonics enable the hybrid curvature tensor to comprehensively cover defect features of different shapes. Simultaneously, the hybrid curvature tensor exhibits a positive correlation with the shape weight coefficient, the value of which determines the system's sensitivity to linear defects. As the shape weight coefficient increases, the proportion of average curvature energy in the hybrid curvature tensor increases, enhancing the system's ability to capture minute scratches or shallow creases. This avoids the problems of missing linear defects due to using only Gaussian curvature modulus or being overly sensitive to noise due to using only average curvature energy. This hybrid curvature tensor construction method based on second-order differential analysis using normalized surface gradient modulus is significantly necessary for detecting fabric defects with varied surface morphologies. The hybrid curvature tensor utilizes differential geometry principles to achieve full compatibility with defect morphologies during the feature extraction stage, solving the technical challenge of simultaneously and accurately detecting both point and linear defects, and improving the algorithm's adaptability to complex defect morphologies.

[0032] Step 3: Based on the hybrid curvature tensor, using the manifold unsupervised reconstruction logic, the Gaussian weighted spatial neighborhood prediction model of the hollowed-out center pixel is used to deduce the theoretical texture value of the current pixel using the texture features of the surrounding neighborhood, and the ideal reconstruction tensor is calculated. In this embodiment, based on the hybrid curvature tensor and the unsupervised reconstruction logic of the manifold, the Gaussian weighted spatial neighborhood prediction model of the hollowed-out center pixel is used to deduce the theoretical texture value of the current pixel using the texture features of the surrounding neighborhood, and the ideal reconstruction tensor is calculated. The principle of ideal reconstruction tensor calculation:

[0033] Where, ψ (x,y) Let W represent the ideal reconstruction tensor at coordinates (x, y). sum δ represents the weight normalization constant, i represents the increment of the horizontal axis, j represents the increment of the vertical axis, r represents the preset neighborhood window radius, which is set by relevant personnel according to the reference surrounding range required to predict the current point, and is usually between 5 and 15. τ represents the preset spatial correlation decay constant, δ ij Let i represent the Kronecker function. When i=0 and j=0, the Kronecker function value is 1, otherwise the Kronecker function value is 0. The fundamental purpose of employing manifold unsupervised reconstruction logic to calculate the ideal reconstruction tensor is to mathematically construct a pure texture benchmark derived solely from the surrounding neighborhood environment, based on the local self-similarity of fabric textures. Specifically, the ideal reconstruction tensor represents the theoretical texture feature values ​​that the current pixel should possess under the assumption that the current region is flawless. The technical advantage of this derived value lies in its complete removal of the pixel's own true information, thus providing an idealized background reference for subsequent difference operations. In terms of computational mechanism, the generation of the ideal reconstruction tensor strictly relies on a Gaussian-weighted spatial neighborhood prediction model that hollows out the center pixel. The Kronecker function plays a crucial shielding role, forcibly resetting the weights of the current pixel's own mixture curvature tensor to zero. This ensures that the ideal reconstruction tensor is derived solely from the weighted sum of the mixture curvature tensors of surrounding pixels, thereby blocking the possibility of potential defective points participating in its own reconstruction at the source. The spatial correlation decay constant defines the spatial decay rate of the influence of neighborhood information on the center prediction point. From the perspective of numerical response patterns, the ideal reconstruction tensor is positively correlated with the mixing curvature tensor of each pixel in the neighborhood. That is, the more complex or stronger the neighborhood texture, the higher the predicted value of the ideal reconstruction tensor, thus maintaining the local continuity of the texture manifold. At the same time, the contribution weight of the neighborhood pixel to the ideal reconstruction tensor is non-linearly negatively correlated with the spatial distance from the pixel to the center point. The closer the neighborhood pixel is, the greater its weight, and the farther away it is, the weaker its correction effect becomes. This calculation method of using spatial correlation for blind spot prediction effectively solves the engineering problem of building a high-precision good product background model under the condition of lacking defective samples for training.

[0034] Step 4: Calculate the absolute value of the difference between the hybrid curvature tensor and the ideal reconstruction tensor, and normalize the difference using the feature standard deviation within the local statistical window to calculate the local signal-to-noise ratio anomaly response index after removing background texture roughness interference. In this embodiment, the absolute value of the difference between the hybrid curvature tensor and the ideal reconstructed tensor is calculated, and the difference is normalized using the feature standard deviation within the local statistical window to calculate the local signal-to-noise ratio anomaly response index after removing background texture roughness interference. Principle of calculating the local signal-to-noise ratio anomaly response index:

[0035] Among them, Ω （x，y） The local signal-to-noise ratio anomaly response index is represented at coordinates (x, y), D represents a local statistical window of a pre-defined rectangular region centered at coordinates (x, y), (p, q) represents the coordinate indices within the local statistical window, where p represents the x-coordinate index value, q represents the y-coordinate index value, and N represents the local statistical window. D ψ represents the total number of coordinate points within the local statistics window. DAVG η represents the average value of the mixed curvature tensor within the local statistical window, and η represents the preset base smoothing constant. The core objective of using the Local Signal-to-Noise Ratio Anomaly Response Index (LSRI) as the final defect determination criterion is to establish a normalized metric that can automatically and dynamically adjust detection sensitivity based on the roughness of the local fabric texture. Physically, the LSRI characterizes the statistical significance ratio of the texture abrupt change feature of the current pixel relative to the inherent fluctuation level of its local background, rather than simply the absolute error magnitude. This ratio-based design allows the algorithm to automatically reduce sensitivity in rough, complex texture areas to avoid false alarms, while maintaining high sensitivity in smooth texture areas to prevent missed detections. Numerically, the LSRI exhibits a positive correlation with the absolute value of the difference between the hybrid curvature tensor and the ideal reconstruction tensor. The difference, as the numerator, directly quantifies the physical significance of the current point's deviation from the good product benchmark. The larger the difference, the higher the index value, thus highlighting the defect. At the same time, the local signal-to-noise ratio anomaly response index and the feature standard deviation within the local statistical window are negatively correlated and inhibit each other. The feature standard deviation, as the denominator, objectively reflects the texture complexity or background noise intensity of the local area. The larger the standard deviation, the stronger the background interference. The numerator signal is normalized and attenuated through division, thereby effectively suppressing the false high response caused by rough texture. Combined with the base smoothing constant to prevent the denominator from being zero, this adaptive calculation logic that combines signal strength and background statistical characteristics ensures that the detection method still has extremely high robustness and accuracy when facing non-uniform woven textures.

[0036] Step 5: Compare the local signal-to-noise ratio anomaly response index with the preset response index threshold. When the local signal-to-noise ratio anomaly response index exceeds the preset response index threshold, it is determined that there is a defect in the fabric and the defect coordinates are located. In this embodiment, the local signal-to-noise ratio anomaly response index is compared with a preset response index threshold. If the local signal-to-noise ratio anomaly response index does not exceed the preset response index threshold, it is determined that the current fabric does not have a defect at coordinate (x, y). If the local signal-to-noise ratio anomaly response index exceeds the preset response index threshold, it is determined that the current fabric has a defect at coordinate (x, y), and the defect coordinates are output.

[0037] Step 6: When it is determined that there is a defect in the fabric, the defect connected region is extracted using the defect coordinates. Based on the connected region, the defect structure feature value and the defect oil stain feature value are calculated. The defect structure feature value is compared with the preset defect structure threshold, and the defect oil stain feature value is compared with the defect oil stain threshold. The physical coordinate area and category of the defect are output according to the comparison results.

[0038] In this embodiment, when it is determined that there is a defect in the fabric, the coordinates of the defect are obtained. The obtained defect coordinates are marked with connected components to obtain the physical coordinate range of the defect. For the mean of the local signal-to-noise ratio anomaly response index and the Weber contrast within the physical coordinate range of the defect, the defect structure feature value and the defect oil stain feature value are calculated respectively. Principle of Defect Structure Feature Value Calculation:

[0039] Among them, V S R represents the structural characteristic value of the defect, R represents the set of coordinates of the physical coordinate range of the defect, and N represents the structural characteristic value of the defect. R The total number of physical coordinate points representing the defect; Principle of calculating the characteristic value of oil stain defects:

[0040] Among them, V C The characteristic value of the defective oil stain, μ DE This represents the average grayscale value of all pixels within the physical coordinate range of the defect in the top light response map, in μ. BG This represents the average pixel grayscale value of a ring-shaped window one coordinate outside the physical coordinate range of the defect in the top light response map. The defect structure feature value is compared with the preset defect structure threshold. If the defect structure feature value does not exceed the preset defect structure threshold, the current defect is determined to be a structural defect. If the defect structure feature value exceeds the preset defect structure threshold, the current defect is determined to be a structural defect. The blemish oil stain feature value is compared with the blemish oil stain threshold. If the blemish oil stain feature value does not exceed the preset blemish oil stain threshold, the current blemish is determined to be a stain type blemish. If the blemish oil stain feature value exceeds the preset blemish oil stain threshold, the current blemish is determined to be a stain type blemish. Output the physical coordinates and category of the defect based on the comparison results; By utilizing dual-channel feature decoupling logic to calculate defect structure feature values ​​and defect oil stain feature values ​​separately, the core objective is to establish a classification and judgment mechanism capable of accurately identifying defect types based on multi-dimensional physical attributes. The numerical value of the defect structure feature value specifically reflects the average significance of the physical deformation signal within the defect's connected domain, meaning it quantifies the concavity and convexity energy density of that region in the normalized surface gradient modulus field. The technical effect is to accurately characterize the physical authenticity of weaving structure defects such as holes, knots, or broken yarns. In terms of numerical dependence, the defect structure feature value and the local signal-to-noise ratio anomaly response index within the defect's physical coordinate range show a positive cumulative correlation, i.e., the signal-to-noise ratio anomaly response index within the local region... The stronger the response, the higher the calculated defect structure feature value, and the greater the probability of identifying it as a structural defect. The value of the defect oil stain feature value specifically reflects the visual contrast intensity of the defect area under top lighting conditions. It means that the Weber contrast principle is used to quantify the difference in albedo between the defect subject and the surrounding background. The technical effect is that it is specifically sensitive to the color change characteristics of surface dirt defects such as oil stains, spots, or flyweed. This feature value is positively correlated with the absolute value of the difference between the average value of all pixel gray values ​​within the physical coordinate range of the defect in the top lighting response map and the average value of pixel gray values ​​in the annular window of the outer coordinate. That is, the more significant the color difference, the higher the defect oil stain feature value. This classification calculation method based on the orthogonal characteristics of structured light path and material light path information has significant rationality and necessity. In the back-end decision stage, it separates structural defects and oil stain defects again through independent threshold logic, solving the technical problem that it is difficult to distinguish the nature of defects by single-dimensional judgment. Thus, it realizes the refined diagnosis and classification output of fabric quality problems. In addition, this method utilizes the physical desaturation characteristics of normalized surface gradient modulus, enabling the system to automatically ignore normal printing and dyeing patterns and yarn-dyed textures on the fabric surface, avoiding the defect of traditional grayscale detection algorithms that easily misjudge dark patterns as oil stain defects.

[0041] Please see Figure 2 The present invention also provides an online visual inspection device for high-speed moving fabric defects, the inspection device being used to implement the above-mentioned inspection method, comprising: Data acquisition and processing module: Used to use line frequency synchronization to control the acquisition device for time-division exposure, acquire side light response map and top light response map respectively, calculate and eliminate the color albedo component of the fabric surface by normalizing the difference ratio including numerical gain coefficient, and combine logarithmic function to compress dynamic range, calculate normalized surface gradient modulus characterizing the physical unevenness information of the fabric surface. Hybrid curvature tensor construction module: used to perform second-order differential analysis on normalized surface gradient modulus, extract Gaussian curvature modulus by calculating the absolute value of Hessian matrix, extract mean curvature energy by calculating the square of Laplacian operator, and calculate the weighted linear combination of the two to obtain hybrid curvature tensor. Manifold Unsupervised Reconstruction Module: Based on the hybrid curvature tensor and manifold unsupervised reconstruction logic, it uses a Gaussian weighted spatial neighborhood prediction model with the center pixel hollowed out, and uses the texture features of the surrounding neighborhood to deduce the theoretical texture value of the current pixel and calculate the ideal reconstruction tensor. Anomaly response index calculation module: used to calculate the absolute value of the difference between the hybrid curvature tensor and the ideal reconstruction tensor, and to normalize the difference using the feature standard deviation within the local statistical window, thereby calculating the local signal-to-noise ratio anomaly response index after removing background texture roughness interference. Defect Detection Module: Used to compare the local signal-to-noise ratio abnormal response index with the preset response index threshold. When the local signal-to-noise ratio abnormal response index exceeds the preset response index threshold, the fabric is determined to have defects and the defect coordinates are located. Defect Analysis Module: When a defect is determined to exist in the fabric, the module extracts the connected component of the defect using the defect coordinates, calculates the defect structure feature value and the defect oil stain feature value based on the connected component, compares the defect structure feature value with the preset defect structure threshold, compares the defect oil stain feature value with the defect oil stain threshold, and outputs the physical coordinate region and category of the defect based on the comparison results.

[0042] 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.

[0043] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0044] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0045] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for online visual inspection of defects in high-speed moving fabrics, characterized in that, The specific steps include: Using a line frequency synchronization control acquisition device for time-division exposure, side light response map and top light response map are acquired respectively. The color albedo component of the fabric surface is eliminated by normalized difference ratio including numerical gain coefficient. The dynamic range is compressed by combining logarithmic function, and the normalized surface gradient modulus characterizing the physical unevenness information of the fabric surface is calculated. Second-order differential analysis was performed on the normalized surface gradient modulus. The Gaussian curvature modulus was extracted by calculating the absolute value of the Hessian matrix, and the average curvature energy was extracted by calculating the square of the Laplacian operator. The two were then combined in a weighted linear manner to obtain the mixed curvature tensor. Based on the hybrid curvature tensor, and using the unsupervised reconstruction logic of the manifold, the Gaussian weighted spatial neighborhood prediction model of the hollowed-out center pixel is used to deduce the theoretical texture value of the current pixel using the texture features of the surrounding neighborhood, and the ideal reconstruction tensor is calculated. The absolute value of the difference between the hybrid curvature tensor and the ideal reconstruction tensor is calculated, and the difference is normalized using the feature standard deviation within the local statistical window to obtain the local signal-to-noise ratio anomaly response index after removing background texture roughness interference. The local signal-to-noise ratio anomaly response index is compared with a preset response index threshold. When the local signal-to-noise ratio anomaly response index exceeds the preset response index threshold, the fabric is determined to have defects and the defect coordinates are located. When a defect is determined to exist in the fabric, the defect's connected component is extracted using the defect coordinates. Based on this connected component, the defect's structural feature value and the defect's oil stain feature value are calculated. The defect's structural feature value is compared with a preset defect structure threshold, and the defect's oil stain feature value is compared with the defect's oil stain threshold. Based on the comparison results, the physical coordinate region and category of the defect are output.

2. The online visual inspection method for defects in high-speed moving fabrics according to claim 1, characterized in that: Using a line scan camera, in conjunction with a rotary encoder mounted on the conveyor mechanism, the continuous fabric surface is divided into alternating optical-physical channels in the time dimension; The preset counter number is updated in real time by detecting the pulse signal output by the encoder. The initial value of the counter number is 0. When the fabric to be detected moves forward a preset distance, the encoder outputs a pulse signal to increment the counter number. When the counter number is odd, the line scan camera is triggered to expose, and the low-angle side light source is turned on and the top light source is turned off. The odd-numbered rows of the acquired fabric image are marked as structured light data. When the counter number is even, the line scan camera is triggered to expose, and the high-angle top light source is turned on and the top light source is turned off. The even-numbered rows of the acquired fabric image are marked as material light data. The structured light data is recombined to generate a single-pass grayscale image, which yields a side light response map that carries the shadow and highlight information of the fabric surface. The material light data is also recombined to generate a single-pass grayscale image, which yields a top light response map that carries the inherent reflectivity information of the fabric surface.

3. The online visual inspection method for defects in high-speed moving fabrics according to claim 2, characterized in that: Based on the side light response map and top light response map, and based on the Lambert reflection model and the principle of photometric stereo vision, the fabric to be tested is desaturated, decoupled and dynamically compressed to calculate the normalized surface gradient modulus that characterizes the physical unevenness of the fabric surface. Principle of normalized surface gradient modulus calculation: Where x represents the x-coordinate, y represents the y-coordinate, and G... （x，y） Normalized surface gradient modulus at coordinates (x, y), IA （x，y） IB represents the grayscale value of the sidelight response map at coordinates (x, y). （x，y） θ represents the grayscale value of the top light response map at coordinates (x, y). A θ represents the incident angle of the side-lit light source. B denoted by , μ represents the incident angle of the top light source, k represents the preset numerical gain coefficient, k represents the preset luminous flux balance coefficient (obtained by whiteboard calibration), σ represents the preset thermal noise suppression factor (obtained by dark field calibration), and ε represents the preset regularization minimum value to prevent the denominator from being zero.

4. The online visual inspection method for defects in high-speed moving fabric according to claim 3, characterized in that: Second-order differential analysis is performed on the normalized surface gradient modulus. The Gaussian curvature modulus is extracted by calculating the absolute value of the Hessian matrix, the average curvature energy is extracted by calculating the square of the Laplacian operator, and the mixed curvature tensor for each coordinate is calculated. The principle of calculating the mixed curvature tensor for each coordinate: Where, ψ (x,y) This represents the mixed curvature tensor at coordinates (x, y). Let represent the second partial derivative of the normalized surface gradient modulus at coordinates (x, y) along the horizontal axis. Let represent the second partial derivative of the normalized surface gradient modulus at coordinates (x, y) along the horizontal axis. The mixed partial derivative of the normalized surface gradient modulus at coordinates (x, y), where λ represents the preset shape weight coefficient.

5. The online visual inspection method for defects in high-speed moving fabrics according to claim 4, characterized in that: Based on the hybrid curvature tensor, and using the unsupervised reconstruction logic of the manifold, the Gaussian weighted spatial neighborhood prediction model of the hollowed-out center pixel is used to deduce the theoretical texture value of the current pixel using the texture features of the surrounding neighborhood, and the ideal reconstruction tensor is calculated. The principle of ideal reconstruction tensor calculation: Where, ψ (x,y) Let W represent the ideal reconstruction tensor at coordinates (x, y). sum δ represents the weight normalization constant, i represents the increment of the horizontal axis, j represents the increment of the vertical axis, r represents the preset neighborhood window radius, τ represents the preset spatial correlation decay constant, and δ ij Let i represent the Kronecker function. When i=0 and j=0, the Kronecker function value is 1, otherwise the Kronecker function value is 0.

6. The online visual inspection method for defects in high-speed moving fabric according to claim 5, characterized in that: The absolute value of the difference between the hybrid curvature tensor and the ideal reconstruction tensor is calculated, and the difference is normalized using the feature standard deviation within the local statistical window to obtain the local signal-to-noise ratio anomaly response index after removing background texture roughness interference. Principle of calculating the local signal-to-noise ratio anomaly response index: Among them, Ω （x，y） The local signal-to-noise ratio anomaly response index is represented at coordinates (x, y), D represents a local statistical window of a pre-defined rectangular region centered at coordinates (x, y), (p, q) represents the coordinate indices within the local statistical window, where p represents the x-coordinate index value, q represents the y-coordinate index value, and N represents the local statistical window. D ψ represents the total number of coordinate points within the local statistics window. DAVG η represents the average value of the mixed curvature tensor within the local statistical window, and η represents the preset base smoothing constant.

7. The online visual inspection method for defects in high-speed moving fabric according to claim 6, characterized in that: The local signal-to-noise ratio anomaly response index is compared with the preset response index threshold. If the local signal-to-noise ratio anomaly response index does not exceed the preset response index threshold, it is determined that there is no defect in the current fabric at coordinate (x, y). If the local signal-to-noise ratio anomaly response index exceeds the preset response index threshold, it is determined that there is a defect in the current fabric at coordinate (x, y), and the defect coordinates are output.

8. The online visual inspection method for defects in high-speed moving fabric according to claim 7, characterized in that: When it is determined that there is a defect in the fabric, the coordinates of the defect are obtained. The obtained defect coordinates are marked with connected components to obtain the physical coordinate range of the defect. For the mean of the local signal-to-noise ratio anomaly response index and the Weber contrast within the physical coordinate range of the defect, the defect structure feature value and the defect oil stain feature value are calculated respectively. Principle of Defect Structure Feature Value Calculation: Among them, V S R represents the structural characteristic value of the defect, R represents the set of coordinates of the physical coordinate range of the defect, and N represents the structural characteristic value of the defect. R The total number of physical coordinate points representing the defect; Principle of calculating the characteristic value of oil stain defects: Among them, V C The characteristic value of the defective oil stain, μ DE This represents the average grayscale value of all pixels within the physical coordinate range of the defect in the top light response map, in μ. BG This represents the average pixel grayscale value of a ring-shaped window one coordinate outside the physical coordinate range of the defect in the top light response map. The defect structure feature value is compared with the preset defect structure threshold. If the defect structure feature value does not exceed the preset defect structure threshold, the current defect is determined to be a structural defect. If the defect structure feature value exceeds the preset defect structure threshold, the current defect is determined to be a structural defect. The blemish oil stain feature value is compared with the blemish oil stain threshold. If the blemish oil stain feature value does not exceed the preset blemish oil stain threshold, the current blemish is determined to be a stain type blemish. If the blemish oil stain feature value exceeds the preset blemish oil stain threshold, the current blemish is determined to be a stain type blemish. Output the physical coordinates and category of the defect based on the comparison results.

9. A high-speed moving fabric defect online visual inspection device, characterized in that: The detection device is used to implement the detection method according to any one of claims 1-8, including: Data acquisition and processing module: Used to use line frequency synchronization to control the acquisition device for time-division exposure, acquire side light response map and top light response map respectively, calculate and eliminate the color albedo component of the fabric surface by normalizing the difference ratio including numerical gain coefficient, and combine logarithmic function to compress dynamic range, calculate normalized surface gradient modulus characterizing the physical unevenness information of the fabric surface. Hybrid curvature tensor construction module: used to perform second-order differential analysis on normalized surface gradient modulus, extract Gaussian curvature modulus by calculating the absolute value of Hessian matrix, extract mean curvature energy by calculating the square of Laplacian operator, and calculate the weighted linear combination of the two to obtain hybrid curvature tensor. Manifold Unsupervised Reconstruction Module: Based on the hybrid curvature tensor and manifold unsupervised reconstruction logic, it uses a Gaussian weighted spatial neighborhood prediction model with the center pixel hollowed out, and uses the texture features of the surrounding neighborhood to deduce the theoretical texture value of the current pixel and calculate the ideal reconstruction tensor. Anomaly response index calculation module: used to calculate the absolute value of the difference between the hybrid curvature tensor and the ideal reconstruction tensor, and to normalize the difference using the feature standard deviation within the local statistical window, thereby calculating the local signal-to-noise ratio anomaly response index after removing background texture roughness interference. Defect Detection Module: Used to compare the local signal-to-noise ratio abnormal response index with the preset response index threshold. When the local signal-to-noise ratio abnormal response index exceeds the preset response index threshold, the fabric is determined to have defects and the defect coordinates are located. Defect Analysis Module: When a defect is determined to exist in the fabric, the module extracts the connected component of the defect using the defect coordinates, calculates the defect structure feature value and the defect oil stain feature value based on the connected component, compares the defect structure feature value with the preset defect structure threshold, compares the defect oil stain feature value with the defect oil stain threshold, and outputs the physical coordinate region and category of the defect based on the comparison results.

Images