CN-121978124-A - Online visual detection method and device for high-speed movement fabric flaws

CN121978124ACN 121978124 ACN121978124 ACN 121978124ACN-121978124-A

Abstract

The invention provides a high-speed movement cloth flaw online visual detection method and device, and relates to the technical field of cloth flaw detection, wherein a response chart is firstly obtained by time-sharing exposure, and normalized surface gradient modulus is calculated through normalized differential ratio and logarithmic compression, so that color albedo interference is eliminated; the method comprises the steps of calculating the absolute value of a Hessen matrix and the square of a Laplace operator to extract curvature characteristics, combining the obtained curvature characteristics to generate a mixed curvature tensor, accurately capturing micro deformation, carrying out manifold unsupervised reconstruction based on a Gaussian weighted model of a hollowed central pixel to obtain an ideal reconstruction tensor, realizing sample-free self-adaptive prediction, further calculating differences, normalizing by using a local statistical window standard deviation to obtain a local signal-to-noise ratio abnormal response index, removing texture roughness interference, finally, positioning flaws, carrying out double-threshold judgment by calculating flaw structure characteristic values and flaw oil stain characteristic values, and outputting flaw coordinates and categories.

Inventors

CAO LIXIA
QIAN JUN
YANG XIFENG
AN XIAOLONG
ZHANG DEBAO
HE MING
KUANG YAWEI

Assignees

常熟市宝沣特种纤维有限公司
苏州工学院

Dates

Publication Date: 20260505
Application Date: 20260107

Claims (9)

1. The online visual detection method for the defects of the cloth moving at a high speed is characterized by comprising the following specific steps of. Carrying out time-sharing exposure by utilizing a line frequency synchronous control acquisition device, respectively acquiring an edge light response graph and a top light response graph, eliminating color albedo components of the cloth surface through normalized differential ratio calculation comprising numerical gain coefficients, and calculating normalized surface gradient modulus representing physical concave-convex information of the cloth surface by combining a logarithmic function compression dynamic range; performing second-order differential analysis on the normalized surface gradient modulus, calculating the square of the Laplacian operator to extract average curvature energy by calculating the absolute value of the Heisen matrix to extract Gaussian curvature modulus, performing weighted linear combination on the two to obtain mixed curvature tensor by calculation; based on manifold unsupervised reconstruction logic, according to the mixed curvature tensor, utilizing a Gaussian weighted spatial neighborhood prediction model of hollowed-out central pixels, and utilizing texture features of surrounding neighborhood to deduce a theoretical texture value of a current pixel point, and calculating an ideal reconstruction tensor; Calculating the absolute value of the difference between the mixed curvature tensor and the ideal reconstruction tensor, carrying out normalization processing on the difference by utilizing the characteristic standard deviation in the local statistical window, and calculating to obtain a local signal-to-noise ratio abnormal response index for removing the background texture roughness interference; Comparing the local signal-to-noise ratio abnormal response index with a preset response index threshold, and judging that the cloth has flaws and positioning flaw coordinates when the local signal-to-noise ratio abnormal response index exceeds the preset response index threshold; When judging that the cloth has flaws, extracting flaw connected domains by utilizing flaw coordinates, calculating flaw structure characteristic values and flaw oil stain characteristic values based on the connected domains, comparing the flaw structure characteristic values with preset flaw structure threshold values, comparing the flaw oil stain characteristic values with the flaw oil stain threshold values, and outputting physical coordinate areas and categories of flaws according to comparison results.
2. The method for online visual inspection of high-speed moving cloth defects according to claim 1, wherein the continuous cloth surface is segmented into alternate optical physical channels in a time dimension by using a linear camera in combination with a rotary encoder arranged on a conveying mechanism; The method comprises the steps of updating a preset counter number in real time by detecting a pulse signal output by an encoder, wherein the initial value of the counter number is 0, outputting a pulse signal once by the encoder after a cloth to be detected advances for a preset distance, enabling the counter number to be increased by one, triggering a linear array camera to expose and synchronously starting a low-angle side light source and closing a top light source when the counter number is odd, marking the odd-numbered data of an acquired cloth image as structured light data, triggering the linear array camera to expose and synchronously starting a high-angle top light source and closing the top light source when the counter is even, and marking the acquired even-numbered image of the cloth as material light data; And recombining the structured light data to generate a single-pass gray level image to obtain a side light response image carrying surface shadow and highlight information of the cloth, and recombining the material light data to generate a single-pass gray level image to obtain a top light response image carrying inherent reflectivity information of the surface of the cloth.
3. The online visual detection method for the high-speed movement fabric defects is characterized in that the fabric to be detected is subjected to decoupling and dynamic compression based on a lambertian reflection model and a luminosity stereoscopic vision principle according to a side light response diagram and a top light response diagram, and normalized surface gradient modulus representing physical concave-convex information of the fabric surface is obtained through calculation; Normalized surface gradient modulus calculation principle: Wherein x represents the abscissa, y represents the ordinate, G （x,y） represents the normalized surface gradient modulus at the coordinate (x, y), IA （x,y） represents the gray value of the side light response map at the coordinate (x, y), IB （x,y） represents the gray value of the top light response map at the coordinate (x, y), θ A represents the incident angle of the side light source, θ B represents the incident angle of the top light source, μ represents a preset numerical gain coefficient, k represents a preset luminous flux balance coefficient, obtained by a whiteboard calibration method, σ represents a preset thermal noise suppression factor, obtained by a dark field calibration method, and ε represents a preset regularized minimum value preventing denominator from being zero.
4. The method for online visual inspection of high-speed motion cloth defects according to claim 3, wherein the normalized surface gradient modulus is subjected to second order differential analysis, the absolute value of the hessian matrix is calculated to extract gaussian curvature modulus, the square of the laplace operator is calculated to extract average curvature energy, and the mixed curvature tensor of each coordinate is calculated; Principle of mixed curvature tensor calculation for each coordinate: wherein, ψ (x,y) denotes the mixed curvature tensor at coordinates (x, y), Representing the second order partial derivative of the normalized surface gradient modulus at coordinates (x, y) in the abscissa direction, Representing the second order partial derivative of the normalized surface gradient modulus at coordinates (x, y) in the abscissa direction, The mixed partial derivative of the normalized surface gradient modulus at coordinates (x, y), λ, represents a preset morphological weighting coefficient.
5. The method for online visual detection of high-speed motion cloth flaws according to claim 4, wherein the method is characterized in that based on manifold unsupervised reconstruction logic, an ideal reconstruction tensor is calculated by utilizing Gaussian weighted spatial neighborhood prediction model of hollowed-out central pixels and utilizing texture features of surrounding neighborhood to deduce theoretical texture value of current pixel points; ideal reconstruction tensor calculation principle: Wherein, psi is (x,y) Representing an ideal reconstruction tensor at coordinates (x, y), W sum represents a weight normalization constant, i represents an abscissa increment, j represents an ordinate increment, r represents a preset domain window radius, τ represents a preset spatial correlation decay constant, δ ij represents a kronecker function, which takes 1 when i=0 and j=0, otherwise takes 0.
6. The method for online visual inspection of high-speed motion cloth defects according to claim 5, wherein the method is characterized in that absolute values of differences between a mixed curvature tensor and an ideal reconstruction tensor are calculated, the differences are normalized by utilizing characteristic standard deviations in a local statistical window, and a local signal-to-noise ratio abnormal response index for removing background texture roughness interference is calculated; Calculating a local signal-to-noise ratio abnormal response index principle: Wherein Ω （x,y） denotes a local signal-to-noise ratio abnormality response index at coordinates (x, y), D denotes a local statistical window of a rectangular area preset with coordinates (x, y) as a center, (p, q) denotes a coordinate index within the local statistical window, p denotes an abscissa index value, q denotes an ordinate index value, N D denotes a total number of coordinate points within the local statistical window, ψ DAVG denotes a mixed curvature tensor average value within the local statistical window, η denotes a preset base smoothing constant.
7. The method for online visual inspection of high-speed motion cloth defects according to claim 6, wherein the local signal-to-noise ratio abnormal response index is compared with a preset response index threshold, if the local signal-to-noise ratio abnormal response index does not exceed the preset response index threshold, it is judged that the current cloth has no defects at coordinates (x, y), and if the local signal-to-noise ratio abnormal response index exceeds the preset response index threshold, it is judged that the current cloth has defects at coordinates (x, y), and the defect coordinates are output.
8. The method for online visual detection of the high-speed movement fabric defects is characterized in that when the fabric defects are judged, defect coordinates are obtained, the obtained defect coordinates are subjected to connected domain marking to obtain a physical coordinate range of the defects, and a defect structure characteristic value and a defect greasy dirt characteristic value are respectively calculated aiming at a local signal-to-noise ratio abnormal response index mean value and a Weber contrast in the physical coordinate range of the defects; the defect structure characteristic value calculating principle: Wherein V S represents a flaw structural characteristic value, R represents a coordinate set of a physical coordinate range of a flaw, and N R represents the total number of coordinate points of the physical coordinate range of the flaw; The principle of calculating the characteristic value of the flaw oil stain is as follows: Wherein V C represents a flaw oil stain characteristic value, mu DE represents an average value of gray values of all pixels in a physical coordinate range of a flaw in a top light response diagram, mu BG represents an average value of gray values of pixels in an annular window with one coordinate at the periphery of the physical coordinate range of the flaw in the top light response diagram; Comparing the flaw structure characteristic value with a preset flaw structure threshold value, if the flaw structure characteristic value does not exceed the preset flaw structure threshold value, judging that the current flaw is not a structure type flaw, and if the flaw structure characteristic value exceeds the preset flaw structure threshold value, judging that the current flaw is a structure type flaw; comparing the flaw oil stain characteristic value with a flaw oil stain threshold value, if the flaw oil stain characteristic value does not exceed the preset flaw oil stain threshold value, judging that the current flaw is not a stain type flaw, and if the flaw oil stain characteristic value exceeds the preset flaw oil stain threshold value, judging that the current flaw is a stain type flaw; and outputting the physical coordinate area and the category of the flaw according to the comparison result.
9. An online visual detection device for detecting flaws of cloth moving at high speed is characterized in that the detection device is used for realizing the detection method according to any one of claims 1-8, and comprises the following steps: The data acquisition and processing module is used for carrying out time-sharing exposure by utilizing the line frequency synchronous control acquisition device, respectively acquiring a side light response graph and a top light response graph, calculating and eliminating color albedo components of the cloth surface through normalized differential ratio comprising numerical gain coefficients, and calculating normalized surface gradient modulus representing physical concave-convex information of the cloth surface by combining logarithmic function compression dynamic range; The mixed curvature tensor construction module is used for carrying out second-order differential analysis on the normalized surface gradient modulus, calculating the square of the Laplacian operator to extract the average curvature energy by calculating the absolute value of the Heisen matrix to extract the Gaussian curvature modulus, carrying out weighted linear combination on the two to obtain the mixed curvature tensor by calculation; the manifold non-supervision reconstruction module is used for calculating an ideal reconstruction tensor by utilizing a Gaussian weighted spatial neighborhood prediction model of the hollowed-out central pixel and utilizing texture characteristics of surrounding neighborhood to deduce a theoretical texture value of a current pixel point based on manifold non-supervision reconstruction logic according to the mixed curvature tensor; the abnormal response index calculation module is used for calculating the absolute value of the difference value between the mixed curvature tensor and the ideal reconstruction tensor, carrying out normalization processing on the difference value by utilizing the characteristic standard deviation in the local statistical window, and calculating to obtain the local signal-to-noise ratio abnormal response index for removing the background texture roughness interference; The flaw judging module is used for comparing the local signal-to-noise ratio abnormal response index with a preset response index threshold value, judging that flaws exist in the cloth when the local signal-to-noise ratio abnormal response index exceeds the preset response index threshold value, and positioning flaw coordinates; And the flaw analysis module is used for extracting a flaw connected domain by utilizing flaw coordinates when judging that flaws exist in the cloth, calculating flaw structure characteristic values and flaw oil stain characteristic values based on the connected domain, comparing the flaw structure characteristic values with preset flaw structure threshold values, comparing the flaw oil stain characteristic values with the flaw oil stain threshold values, and outputting physical coordinate areas and categories of flaws according to comparison results.

Description

Online visual detection method and device for high-speed movement fabric flaws Technical Field The invention relates to the technical field of cloth flaw detection, in particular to an online visual detection method and device for high-speed movement cloth flaws. Background The textile printing industry is gradually changing and upgrading to an automatic and intelligent direction as an important support of the traditional manufacturing industry. With the popularization of modern high-speed looms, the running speed of a cloth production line is greatly improved, and the real-time monitoring of the weaving quality is strictly required. Traditional cloth flaw detection mainly relies on manual naked eye inspection, but under high-speed motion scene, manual detection is not only labor intensity big, and extremely easily leads to leaking to examine or false detection because of visual fatigue. In recent years, a machine vision detection system which integrates a computer vision technology with optical imaging hardware in depth gradually becomes a key means for replacing manpower and guaranteeing the quality of finished cloth. The technology aims at continuously scanning and imaging the cloth moving at high speed through a linear array camera, and automatically identifying defects such as holes, neps, greasy dirt and the like by utilizing an image processing algorithm, and is a core technical link for realizing zero defect production in the textile industry. In the prior art, publication number CN104949990A discloses a flaw online detection method suitable for machine-made textiles, and a processing flow of multipoint independent acquisition-centralized unified judgment is adopted. The method comprises the steps of firstly acquiring image state data of a target in real time through an acquisition terminal deployed on site, and then converging and transmitting all front-end data to a processing center of a rear end through a network transmission protocol. And finally, performing unified analysis on the received data by utilizing a fixed logic algorithm preset by the processing center so as to judge whether the target is abnormal or not and trigger an alarm. This approach relies primarily on standardized data flow paths and centralized decision logic to maintain production supervision. However, the above prior art has significant technical limitations in facing complex high end material detection. Firstly, the single illumination imaging mode leads to luminosity information coupling, color change and structure change can not be distinguished physically, and physical flaws with the same color are caused, for example, cotton knots with the same color as cloth are missed due to unobvious gray scale difference, or normal dark texture patterns are misjudged as greasy dirt flaws. Secondly, the algorithm based on supervised learning has serious sample dependence, and is difficult to effectively detect in the absence of a large number of negative samples in the face of a cold start scene with various cloth types and unpredictable flaw morphology. Finally, the existing algorithm is mostly based on single-dimension alarm logic, can only output binary judgment of whether flaws exist or not, lacks decoupling classification capability for physical properties of the flaws, and is difficult to meet data requirements of targeted optimization of a back-end production process. The above information disclosed in the background section is only for enhancement of understanding of the background of the disclosure and therefore it may include information that does not form the prior art that is already known to a person of ordinary skill in the art. Disclosure of Invention The invention aims to provide an online visual detection method and device for flaws of high-speed moving cloth, which are used for solving the problems in the background technology. In order to achieve the above purpose, the present invention provides the following technical solutions: an online visual detection method for flaws of cloth moving at a high speed comprises the following specific steps: Carrying out time-sharing exposure by utilizing a line frequency synchronous control acquisition device, respectively acquiring an edge light response graph and a top light response graph, eliminating color albedo components of the cloth surface through normalized differential ratio calculation comprising numerical gain coefficients, and calculating normalized surface gradient modulus representing physical concave-convex information of the cloth surface by combining a logarithmic function compression dynamic range; performing second-order differential analysis on the normalized surface gradient modulus, calculating the square of the Laplacian operator to extract average curvature energy by calculating the absolute value of the Heisen matrix to extract Gaussian curvature modulus, performing weighted linear combination on the two to obtain mixed curvature tensor by calculatio