Intelligent identification method and system for defects in production of webbing
By combining multi-scale feature analysis and clustering algorithms with the differences between local and global features, the problems of scale adaptability and high false detection rate of traditional webbing defect detection algorithms are solved, and efficient and accurate identification of webbing defects is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHONGSHAN YIJIN TEXTILE CO LTD
- Filing Date
- 2026-03-04
- Publication Date
- 2026-06-05
AI Technical Summary
Traditional webbing defect detection algorithms have poor adaptability to scale changes, are prone to missing small defects or falsely detecting large defects, and lack collaborative analysis of texture and structure, resulting in a high false detection rate.
A multi-scale feature analysis method is adopted, which combines LBP feature difference, frequency characteristic difference, and local and global similarity analysis. By fusing gray-scale changes, texture features, and periodic structural destruction factors, the degree of webbing defects is obtained, and clustering algorithms are used to identify defective regions.
It improves the accuracy and robustness of webbing defect detection, takes into account both large and small defects, and significantly enhances detection performance.
Smart Images
Figure CN122156138A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of webbing defect detection technology, specifically to an intelligent method and system for identifying defects in webbing manufacturing. Background Technology
[0002] With the continuous development of the webbing manufacturing industry, webbing, as an important material widely used in clothing, home furnishings, and industrial fields, directly determines the performance and quality of the final product. In the actual production process of webbing, due to the combined effects of raw materials, equipment precision, process parameters, and environmental factors, the surface of webbing is prone to various defects such as holes, stains, broken warp threads, skipped yarns, and color differences. These defects not only affect the appearance of the webbing but also reduce its strength, durability, and other key properties. Therefore, timely and accurate intelligent identification and classification of defects in the webbing manufacturing process is of paramount importance for improving webbing quality and ensuring production efficiency.
[0003] Traditional webbing defect detection algorithms generally use single-scale feature analysis. Although it has low computational complexity and fast processing speed, and can quickly extract significant defect features at a specific scale, such as edge breaks or texture anomalies, single-scale analysis has poor adaptability to scale changes. Small defects (such as a single broken yarn) may be missed due to weak feature response, while large defects (such as large-area stains) may be incompletely segmented due to insignificant local features. At the same time, single-scale feature analysis lacks the collaborative analysis of texture and structure. Webbing has periodic texture and directional structure, and traditional methods are prone to misjudging normal texture changes as defects, resulting in a high false detection rate. Summary of the Invention
[0004] To address the aforementioned technical problems, the present invention aims to provide an intelligent method and system for identifying defects in webbing manufacturing. The specific technical solution adopted is as follows: In a first aspect, one embodiment of the present invention provides an intelligent defect identification method and system for webbing manufacturing, the method comprising: Acquire a surface image of the webbing and convert it to grayscale to obtain the target image; establish a window centered on a pixel in the target image as the first window for that pixel; obtain the suspected defect level of the pixel based on the grayscale changes and texture features of the pixels within the first window; The detail texture anomaly factor of a pixel is obtained based on the difference in LBP features, frequency characteristics, and similarity between the first window of a pixel and the target image, as well as the similarity between the first window and surrounding first windows. The first window of a pixel is enlarged to obtain the second window of the pixel. The periodic structure disruption factor of a pixel is obtained based on the difference in spectrum, gradient direction, and projection curves in the horizontal and vertical directions between the second window of a pixel and the target image. The adjustment factor of a pixel is obtained by multiplying the detail texture anomaly factor and the periodic structure disruption factor and taking the square root. The defect level of a pixel is obtained by multiplying the suspected defect level of the pixel by the adjustment factor of the pixel and normalizing it; all pixels in the target image are clustered using the coordinates and defect levels of each pixel to obtain different clusters; and the defect regions in the target image are obtained according to the defect levels of the pixels in each cluster.
[0005] Preferably, the degree of suspected defect of a pixel is obtained based on the grayscale variation of the pixel within the first window and the texture features, including: The gray-level change feature is obtained by summing the normalized value of the maximum gray-level gradient of each pixel within the first window of a pixel with the normalized value of the standard deviation of the gray-level gradient of each pixel. The normalized value of the absolute difference between the correlation value of the gray-level co-occurrence matrix of the pixel within the first window and the correlation value of the gray-level co-occurrence matrix of the target image is obtained and denoted as the correlation difference. The normalized value of the absolute difference between the entropy value of the gray-level co-occurrence matrix of the pixel within the first window and the entropy value of the gray-level co-occurrence matrix of the target image is obtained and denoted as the entropy difference. The gray-level change feature is multiplied by the sum of the correlation difference and the entropy difference to obtain the suspected defect degree of the pixel.
[0006] Preferably, the detail texture anomaly factor of a pixel is obtained based on the LBP feature difference, frequency characteristic difference, and similarity between the first window of a pixel and the target image, as well as the similarity between the first window and surrounding first windows, including: The LBP algorithm is used to obtain the LBP histogram of the first window of a pixel and the LBP histogram of the target image at different scales. The chi-square distance at each scale is obtained based on the LBP histogram of the first window of the pixel and the LBP histogram of the target image at each scale. Gabor filters of different directions and scales are used to process the pixels within the first window of the pixel and the pixels in the target image, respectively, to obtain the mean window response amplitude and the mean global response amplitude of each filter. Pixels within the first window of the pixel are grouped into a central image block. Pixels within the eight neighborhoods of the central image block are grouped into separate image blocks, denoted as the neighborhood image blocks of the central image block. The mean normalized cross-correlation of the central image block and its neighborhood image blocks is obtained, denoted as the average cross-correlation of the pixel. The detail texture anomaly factor of the pixel is obtained using the chi-square distance at each scale, the mean window response amplitude and the mean global response amplitude of each filter, and the average cross-correlation of the pixel.
[0007] Preferably, the detail texture anomaly factor of the pixel is obtained using the chi-square distance at each scale, the mean window response amplitude of each filter, the mean global response amplitude, and the average cross-correlation of the pixel, including: The normalized value of the mean absolute value of the difference between the mean window response amplitude of each filter and the mean global response amplitude is obtained, and multiplied by the normalized value of the mean chi-square distance of each scale to obtain the local texture anomaly feature of the pixel; the normalized value of the difference between the first preset value and the absolute value of the average cross-correlation of the pixel is obtained, and added to the local texture anomaly feature of the pixel to obtain the detail texture anomaly factor of the pixel.
[0008] Preferably, the periodic structure disruption factor of a pixel is obtained based on the spectral difference, gradient direction difference, and differences in the projection curves in the horizontal and vertical directions between the second window of a pixel and the target image, including: A two-dimensional Fourier transform is performed on the grayscale values of pixels within a second window for each pixel. The zero frequency is shifted to the center of the spectrum, and other frequency components are symmetrically distributed according to high and low frequencies to obtain the spectrum. A predetermined number of frequency points with the largest amplitude in the spectrum are arranged in ascending order of amplitude to form the frequency point sequence of the second window. Similarly, the same analysis is performed on the target image to obtain its frequency point sequence. Two frequency points with the same position in their respective sequences are grouped into a frequency point pair. The Euclidean distance between the positions of the two frequency points in a frequency point pair in the spectrum is obtained and multiplied by the absolute value of the difference in amplitude between the two frequency points in the pair to obtain the distance and amplitude feature difference of the frequency point pair. A grayscale gradient direction histogram of the second window is constructed based on the grayscale gradient direction angle of each pixel, and the frequency in the grayscale gradient direction angle histogram is obtained. The largest gray-level gradient direction angle is used as the main gray-level gradient direction angle of the second window. Similarly, the main gray-level gradient direction angle of the target image is obtained. The absolute value of the difference between the main gray-level gradient direction angles of the second window and the target image is obtained and normalized to obtain the gradient direction angle difference. The similarity between the horizontal projection curve of the pixel in the second window and the horizontal projection curve of the pixel in the target image is obtained and recorded as the horizontal projection similarity. The similarity between the vertical projection curve of the pixel in the second window and the vertical projection curve of the pixel in the target image is obtained and recorded as the vertical projection similarity. The mean values of the horizontal projection similarity and the vertical projection similarity are normalized to obtain the comprehensive similarity. The normalized value of the mean value of the distance and amplitude feature differences of all frequency point pairs, the gradient direction angle difference, and the comprehensive similarity are multiplied to obtain the periodic structure destruction factor of the pixel.
[0009] Preferably, the defect region in the target image is obtained based on the defect degree of pixels in various clusters, including: The average defect level of all pixels in the cluster with the largest number of pixels is obtained and denoted as the background defect level. The average defect level of all pixels in a cluster is obtained and denoted as the average defect level of that cluster. Clusters with an average defect level greater than the background defect level are denoted as candidate clusters. Candidate clusters with an average defect level greater than a preset defect level threshold are selected as defect clusters. A region is formed in the target image using the pixels in a defect cluster and denoted as the defect region.
[0010] Secondly, the present invention also provides an intelligent identification system for defects in ribbon manufacturing, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of an intelligent identification method for defects in ribbon manufacturing.
[0011] The embodiments of the present invention have at least the following beneficial effects: This application acquires a surface image of the webbing and converts it to grayscale as a target image. Then, it uses the eight-neighborhood of a pixel in the target image as the first window of that pixel. Based on the grayscale changes and texture features of the pixel within the first window, the suspected defect level of the pixel is obtained. Then, based on the LBP feature difference, frequency characteristic difference, and similarity between the first window and surrounding first windows of the pixel, the detail texture anomaly factor of the pixel is obtained. Based on the spectral difference, gradient direction difference, and difference in projection curves in the horizontal and vertical directions between the second window and the target image of the pixel, the periodic structure destruction factor of the pixel is obtained. These two parameters are obtained from both microscopic and macroscopic features. Then, the detail texture anomaly factor and the periodic structure destruction factor of the pixel are fused to obtain the adjustment factor of the pixel, which is used to adjust the suspected defect level of the pixel to obtain the defect level of the pixel. Finally, the coordinates and defect levels of each pixel in the target image are used to cluster all pixels to obtain different clusters. Based on the defect level of the pixels in each cluster, the defect region in the target image is obtained. This application obtains the defect level of each pixel through coarse positioning and fine analysis from two angles, thereby taking into account both large and small defect detection and significantly improving detection performance. Attached Figure Description
[0012] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0013] Figure 1 A flowchart illustrating an intelligent defect identification method for webbing manufacturing provided in this embodiment of the invention. Detailed Implementation
[0014] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of an intelligent defect identification method and system for webbing manufacturing based on the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.
[0015] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0016] The following description, in conjunction with the accompanying drawings, details the specific scheme of the intelligent defect identification method and system for webbing manufacturing provided by this invention.
[0017] Example 1: The main application scenario of the present invention is as follows: This application acquires surface images of the fabricated webbing, analyzes the surface images accordingly, obtains the defective areas therein, and then locates the defective areas on the webbing.
[0018] Please see Figure 1 The diagram illustrates a flowchart of an intelligent defect identification method for webbing manufacturing according to an embodiment of the present invention. The method includes the following steps: Step S1: Obtain the surface image of the webbing and convert it to grayscale as the target image; establish a window centered on a pixel in the target image as the first window of that pixel; obtain the suspected defect level of the pixel based on the grayscale change of the pixel within the first window and the texture features.
[0019] In this application, image acquisition of the webbing surface is a crucial foundational step. First, high-definition industrial cameras must be strategically positioned on the webbing production line to ensure their position and angle fully cover the webbing width, acquiring complete and clear images. The camera's shooting frequency is adjusted according to production speed and accuracy requirements to ensure timely acquisition of images for each segment of webbing. Simultaneously, suitable light sources, such as uniform ring lights, are provided to eliminate shadows and reflections, improving image contrast and clarity. Finally, the surface images of the webbing are acquired and stored in real-time. Furthermore, a single webbing surface image is converted to grayscale and used as the target image for subsequent analysis. A dedicated image database is established to provide rich data support for subsequent defect identification and analysis, ensuring a reliable data source for the entire intelligent identification process.
[0020] In the target image, normal and defective regions often exhibit significant differences in gradient and texture. Local gradients reflect changes in image edges and details. Gradient changes in normal webbing regions are relatively uniform, while defective regions, such as broken or stained areas, can cause abnormal local gradients, such as sudden increases or drastic changes in gradient magnitude. Meanwhile, texture consistency is a key indicator for determining whether a region is normal. Normal webbing textures are regular, and local and global textures should be similar, while defective regions show significant differences from the overall texture. By fusing gradient outliers and texture consistency outliers, gradient and texture information can be combined to preliminarily determine the degree to which each pixel is suspected of being a defect, providing a foundation for subsequent detailed analysis.
[0021] First, a window is created centered on a pixel in the target image as the first window for that pixel. The size of the first window is n*n (where n=9, a moderate window that can capture local gradient anomalies while also obtaining local texture consistency).
[0022] Then, based on the grayscale changes and texture features of a pixel within the first window, the degree of suspected defect of that pixel is obtained.
[0023] Specifically, the gray-level change feature is obtained by summing the normalized value of the maximum gray-level gradient of each pixel within the first window of a pixel with the normalized value of the standard deviation of the gray-level gradient of each pixel; the normalized value of the absolute difference between the correlation value of the gray-level co-occurrence matrix of the pixel within the first window and the correlation value of the gray-level co-occurrence matrix of the target image is obtained, denoted as the correlation difference; the normalized value of the absolute difference between the entropy value of the gray-level co-occurrence matrix of the pixel within the first window and the entropy value of the gray-level co-occurrence matrix of the target image is obtained, denoted as the entropy difference; the gray-level change feature is multiplied by the sum of the correlation difference and the entropy difference to obtain the suspected defect degree of the pixel.
[0024] The specific calculation model for the degree of suspected defects is as follows: , Where Ai represents the suspected defect level of the i-th pixel; dimax represents the maximum value of the grayscale gradient of each pixel within the first window; This represents the standard deviation of the grayscale gradient of each pixel within the first window. The grayscale change feature indicates that the larger the maximum value of the grayscale gradient or the more drastic the change, the more it matches the characteristics of the webbing defect area; Cr and En are the correlation value and entropy value of the grayscale co-occurrence matrix of the global image (target image), respectively; Cri and Eni are the correlation value and entropy value of the grayscale co-occurrence matrix of the first window, respectively. For the difference in correlation, Due to the difference in entropy values, The greater the difference in texture regularity or texture complexity between the first window and the global texture, the greater the likelihood of a local defect at that pixel. `norm` represents the normalization function. This allows us to determine the suspected defect level of each pixel in the target image.
[0025] Step S2: Obtain the detail texture anomaly factor of a pixel based on the LBP feature difference, frequency characteristic difference, and similarity between the first window of a pixel and the target image, as well as the similarity between the first window and surrounding first windows; enlarge the first window of a pixel to obtain the second window of the pixel; obtain the periodic structure disruption factor of the pixel based on the spectral difference, gradient direction difference, and difference in projection curves in the horizontal and vertical directions between the second window of a pixel and the target image; multiply the detail texture anomaly factor and the periodic structure disruption factor of a pixel and take the square root to obtain the adjustment factor of the pixel.
[0026] In ribbon image defect detection, normal and defective regions show significant differences in gradient and texture. Based on the degree of local gradient anomaly and texture consistency, the suspected defect level of each pixel can be initially obtained, laying the foundation for subsequent analysis. While the above method obtains the suspected defect level of each pixel, this preliminary judgment may be inaccurate or prone to misjudgment. Therefore, it is necessary to further combine microscopic and macroscopic features, calculating the detail texture anomaly factor and the periodic structure disruption factor respectively, to comprehensively evaluate pixel features from different scales. Integrating these two factors with the suspected defect level allows for the comprehensive assessment of multiple information aspects, more accurately determining the defect level of the target pixel, and improving detection accuracy and robustness.
[0027] After initially obtaining the suspected defect level of each pixel through the above steps, a more in-depth analysis of its local features is needed. At the microscopic level, i.e., from the perspective of local detail texture, by calculating multi-scale LBP feature differences and Gabor filter response differences, subtle changes in texture within the local window can be deeply explored. These changes can keenly reflect anomalies in detail texture, thus yielding detail texture anomaly factors. At the macroscopic level, i.e., from the perspective of global periodic structure, using Fourier spectrum analysis and principal direction consistency calculations, the degree of structural consistency between local regions and the global structure is considered to determine whether the periodic structure has been disrupted, thereby deriving the periodic structure disruption factor. These two approaches complement each other, comprehensively evaluating pixel features at different scales and providing a more accurate basis for subsequent adaptive defect parameter calculations.
[0028] The detail texture anomaly factor focuses on a detailed analysis of various texture and structural features within a local window. First, by calculating the differences between local and global LBP features at multiple scales, it captures the degree of deviation in texture patterns at different scales, reflecting the consistency between local and global textures. LBP at different scales can capture texture information within different ranges; defects generally exhibit anomalies at multiple scales. Second, by utilizing the differences in Gabor filter bank responses, it analyzes the differences in frequency characteristics between local and global textures from different directions and scales, revealing the periodic changes in local and global image textures. Third, through local self-similarity measurement, it assesses the similarity of internal structures within local regions; defective regions typically have low self-similarity. Finally, these features are fused to comprehensively reflect the consistency between local and global textures, thus obtaining the detail texture anomaly factor.
[0029] The detail texture anomaly factor of a pixel is obtained based on the difference in LBP features and frequency characteristics between the first window of a pixel and the target image, as well as the similarity between the first window and surrounding first windows.
[0030] Specifically, the LBP algorithm is used to obtain the LBP histogram of the first window of a pixel and the LBP histogram of the target image at different scales. The chi-square distance at each scale is obtained based on the LBP histogram of the first window of the pixel and the LBP histogram of the target image at each scale. Gabor filters of different directions and scales are used to process the pixels within the first window of the pixel and the pixels in the target image, respectively, to obtain the mean window response amplitude and the mean global response amplitude of each filter. Pixels within the first window of the pixel are grouped into a central image block. Pixels within the eight neighborhoods of the central image block are grouped into separate image blocks, denoted as the neighborhood image blocks of the central image block. The mean normalized cross-correlation of the central image block and its neighborhood image blocks is obtained, denoted as the average cross-correlation of the pixel. The detail texture anomaly factor of the pixel is obtained using the chi-square distance at each scale, the mean window response amplitude and the mean global response amplitude of each filter, and the average cross-correlation of the pixel.
[0031] The detail texture anomaly factor of a pixel is obtained by using the chi-square distance at each scale, the mean window response amplitude of each filter, the mean global response amplitude, and the average cross-correlation of the pixel. Specifically, the normalized value of the mean of the absolute values of the differences between the mean window response amplitude of each filter and the mean global response amplitude is obtained and multiplied by the normalized value of the mean chi-square distance at each scale to obtain the local texture anomaly feature of the pixel; the normalized value of the difference between the first preset value and the absolute value of the average cross-correlation of the pixel is obtained and added to the local texture anomaly feature of the pixel to obtain the detail texture anomaly factor of the pixel.
[0032] The specific calculation model for the detail texture anomaly factor is as follows: , Where Bi is the detail texture anomaly factor corresponding to the i-th pixel; s represents the s-th scale. The scales of the LBP algorithm are R=1, 2, 3, and the number of sampling points are 8, 16, 24 (unified mode), respectively. sum(s)=3, that is, three scales. For each scale, the LBP histogram of the first window and the LBP histogram of the target image are calculated, and then the chi-square distance of the s-th scale is obtained. ; This represents the mean difference between the local and global multi-scale LBP features of the pixel. The larger the value, the more the local features of the pixel match the multi-scale abnormal features of the webbing defect area. J = 4, representing the number of Gabor filter orientations, and K = 3, representing the number of Gabor filter scales. There are four orientations (0°, 45°, 90°, 135°) and three scales (wavelengths of 4, 8, and 16 respectively), resulting in a total of 12 filters. Gabor filters of different orientations and scales are used to process pixels within the first window of the given pixel and pixels in the target image, respectively. The mean window response amplitude and the mean global response amplitude of each filter are obtained. sum(J*K) represents the number of filters, i.e., 3*4=12. The greater the difference between the first window and the global mean, the greater the difference between the local and global periodic texture of that pixel, and the more likely defects such as pattern and print errors will appear locally, which will destroy the overall periodic texture of the image. The average cross-correlation of the pixel is calculated by taking the first window of the pixel as a central image block, and then taking the other first windows within the eight neighborhoods of the central image block as the neighboring image blocks of the central image block using the method of taking the eight neighboring pixels of the pixel. There are a total of 8 neighboring image blocks. The calculation method of the normalized cross-correlation between the central image block and the neighboring image blocks is existing technology and will not be described in detail here. The first preset value is 1. The smaller the average cross-correlation, the weaker the local self-similarity of the pixel, and the more likely it is to be a webbing defect area. Therefore, the absolute value of the average cross-correlation of the pixel is subtracted from the first preset value here. The local texture anomaly features of this pixel represent local texture anomalies, so multiplication is used. This allows us to obtain the detail texture anomaly factor for each pixel.
[0033] The periodic structure disruption factor focuses on analyzing the degree of disruption of the global structure by local windows from the perspective of the overall structure. First, Fourier spectrum analysis compares the differences in frequency components between the local and global structures, revealing deviations in the local structure in the frequency domain. Ribbon textures are typically periodic, and the spectrum reflects this periodicity; defects disrupt this periodicity. Second, a principal direction consistency metric is used to assess the differences in gradient direction distribution between the local and global structures, reflecting the uniformity of structural direction. Third, projection contour analysis compares the similarity of the projection curves of the local and global structures in the horizontal and vertical directions, revealing the consistency of the structural contours. Ribbon textures usually have a clear direction (such as warp direction), and defects generally cause changes in direction. Finally, these features are normalized and weighted to comprehensively reflect the consistency between the local and global structures. The weighting is similar to that of the local texture consistency factor, thus obtaining the global structure consistency factor. These two factors complement each other, jointly improving the accuracy of defect detection.
[0034] When obtaining the periodic structural disruption factor of the target pixel, more emphasis is placed on structural analysis. Therefore, the first window corresponding to the pixel needs to be further expanded to 3n*3n, which is called the second window. A larger window can better capture its local structural features.
[0035] Therefore, the periodic structure disruption factor of a pixel is obtained based on the spectral difference, gradient direction difference, and difference in projection curves in the horizontal and vertical directions between the second window of a pixel and the target image.
[0036] Specifically, a two-dimensional Fourier transform is performed on the grayscale values of pixels within the second window of a pixel, and the zero frequency is shifted to the center of the spectrum. Other frequency components are symmetrically distributed according to high and low frequencies to obtain the spectrum. A predetermined number of frequency points with the largest amplitude in the spectrum are arranged in ascending order of amplitude to form the frequency point sequence of the second window. Similarly, the same analysis is performed on the target image to obtain the frequency point sequence of the target image. Two frequency points with the same position in their respective sequences are combined into a frequency point pair. The Euclidean distance between the positions of the two frequency points in a frequency point pair in the spectrum is obtained, and multiplied by the absolute value of the difference in amplitude between the two frequency points in the frequency point pair to obtain the distance and amplitude feature difference of the frequency point pair. A grayscale gradient direction histogram of the second window is constructed based on the grayscale gradient direction angle of each pixel in the second window, and the grayscale gradient direction angle histogram is obtained. The gray-level gradient direction angle with the highest frequency is used as the main gray-level gradient direction angle of the second window. Similarly, the main gray-level gradient direction angle of the target image is obtained. The absolute value of the difference between the main gray-level gradient direction angles of the second window and the target image is obtained and normalized to obtain the gradient direction angle difference. The similarity between the horizontal projection curve of the pixel in the second window and the horizontal projection curve of the pixel in the target image is obtained and recorded as the horizontal projection similarity. The similarity between the vertical projection curve of the pixel in the second window and the vertical projection curve of the pixel in the target image is obtained and recorded as the vertical projection similarity. The mean values of the horizontal projection similarity and the vertical projection similarity are normalized to obtain the comprehensive similarity. The normalized value of the mean value of the distance and amplitude feature differences of all frequency point pairs, the gradient direction angle difference, and the comprehensive similarity are multiplied to obtain the periodic structure destruction factor of the pixel.
[0037] The specific calculation model for the periodic structure failure factor is as follows: , in, represents the periodic structure disruption factor of the i-th pixel, and norm represents the normalization operation; This represents the Euclidean distance between the positions of the two frequency points in the frequency spectrum of the a-th frequency pair. This represents the amplitude of the frequency point corresponding to the second window of the i-th pixel relative to the a-th frequency point. This represents the amplitude of the frequency point corresponding to the target image in this frequency point pair, and sum(A) represents the number of frequency point pairs. In this application, the preset number of frequency points is 10. Here, the frequency point sequence in the second window and the frequency point sequence in the target image are both taken from the main frequency components of the spectrum, which are representative. The normalized value of the mean of the distance and amplitude feature differences of all frequency point pairs indicates that the greater the spectral difference between the second window and the global image, the higher the degree of local periodic destruction of the pixel, and the more likely it is to be a webbing defect area. Xi and X represent the main gray-level gradient direction angles of the second window and the target image, respectively, in radians. norm(|Xi-X|) represents the difference in gradient direction angles, indicating that the greater the difference in the main direction between the local area of the second window and the overall image, the greater the possibility of anomalies in the local structure. DTWci represents the similarity between the horizontal projection curve of the pixels in the second window and the horizontal projection curve of the pixels in the target image, which is also known as horizontal projection similarity. DTWvi represents the similarity between the vertical projection curve of the pixels in the second window and the vertical projection curve of the pixels in the target image, which is also known as vertical projection similarity. To assess the overall similarity, the lower the horizontal and vertical projection similarity between the second window and the overall image, the greater the likelihood that the local structure of that pixel has an directional anomaly, and the more likely it is to be a webbing anomaly region. The horizontal and vertical projection similarities are obtained using the DTW algorithm.
[0038] After obtaining the detail texture anomaly factor and periodic structure disruption factor for each pixel, combining these two factors yields the adjustment factor for the current pixel. A larger adjustment factor indicates a higher probability that the pixel belongs to a webbing defect region, and a more reliable assessment of the suspected defect. Specifically, the adjustment factor for a pixel is obtained by multiplying its detail texture anomaly factor and periodic structure disruption factor and taking the square root. Using the square root balances the contributions of macroscopic and microscopic factors, avoiding excessive amplification from multiplication or linear superposition from addition. This allows us to obtain the adjustment factor for each pixel.
[0039] Step S3: Multiply the suspected defect level of a pixel by the adjustment factor of the pixel and normalize to obtain the defect level of the pixel; use the coordinates and defect levels of each pixel in the target image to cluster all pixels to obtain different clusters; obtain the defect region in the target image according to the defect level of the pixels in each cluster.
[0040] In step S1, the suspected defect level of each pixel is obtained. The suspected defect level of a pixel is then multiplied by its adjustment factor and normalized to obtain the pixel's actual defect level. The higher the suspected defect level of a pixel and the larger its corresponding adjustment factor, the greater the likelihood that it is a genuine webbing defect, and the greater its defect level.
[0041] This allows us to construct a feature vector for each pixel by combining its coordinates and defect severity. Then, based on the Euclidean distance between the feature vectors of any two pixels, mean-shift clustering is used to divide the pixels of the target image into multiple clusters, with each cluster representing a set of pixels with similar defect features.
[0042] Defect regions in the target image are obtained based on the defect severity of pixels in each cluster. Specifically, the average defect severity of all pixels in the cluster with the largest number of pixels is obtained and denoted as the background defect severity; the average defect severity of all pixels in a cluster is obtained and denoted as the average defect severity of that cluster; clusters with an average defect severity greater than the background defect severity are denoted as candidate clusters; candidate clusters with an average defect severity greater than a preset defect severity threshold are selected as defect clusters; and a region is formed in the target image using pixels from a defect cluster, denoted as the defect region.
[0043] The defect severity threshold is set according to the level of inspection precision of the current batch of webbing. The higher the level of inspection precision, the lower the threshold. Finally, by mapping each defect cluster onto the original image, the defect area can be accurately located and segmented.
[0044] Finally, after acquiring the defective areas of the webbing, multi-dimensional features, such as defect shape, texture, and color distribution, are extracted and combined with basic information such as defect location and area to construct feature vectors. Subsequently, a machine learning model, such as a support vector machine (SVM) or random forest, is trained using labeled defect sample data to classify different types of defects (such as broken warp, missing weft, and stains). The model parameters are optimized through cross-validation to ensure classification accuracy. Finally, the model is deployed to achieve automatic identification and category determination of new defective areas, thus assisting in the intelligent management and control of defects in webbing production.
[0045] Example 2: This embodiment provides an intelligent defect identification system for ribbon manufacturing, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When executed by the processor, the computer program implements the steps of an intelligent defect identification method for ribbon manufacturing. Since Embodiment 1 has already described a detailed method for intelligent defect identification in ribbon manufacturing, it will not be elaborated upon further here.
[0046] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.
[0047] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.
[0048] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for intelligent identification of defects in webbing manufacturing, characterized in that, The method includes: Acquire a surface image of the webbing and convert it to grayscale to obtain the target image; establish a window centered on a pixel in the target image as the first window for that pixel; obtain the suspected defect level of the pixel based on the grayscale changes and texture features of the pixels within the first window; The detail texture anomaly factor of a pixel is obtained based on the difference in LBP features, frequency characteristics, and similarity between the first window of a pixel and the target image, as well as the similarity between the first window and surrounding first windows. The first window of a pixel is enlarged to obtain the second window of the pixel. The periodic structure disruption factor of a pixel is obtained based on the difference in spectrum, gradient direction, and projection curves in the horizontal and vertical directions between the second window of a pixel and the target image. The adjustment factor of a pixel is obtained by multiplying the detail texture anomaly factor and the periodic structure disruption factor and taking the square root. The defect level of a pixel is obtained by multiplying the suspected defect level of the pixel by the adjustment factor of the pixel and normalizing it; all pixels in the target image are clustered using the coordinates and defect levels of each pixel to obtain different clusters; and the defect regions in the target image are obtained according to the defect levels of the pixels in each cluster.
2. The intelligent defect identification method for webbing manufacturing according to claim 1, characterized in that, The step of obtaining the suspected defect level of a pixel based on the grayscale changes and texture features of the pixel within the first window includes: The gray-level change feature is obtained by summing the normalized value of the maximum gray-level gradient of each pixel within the first window of a pixel with the normalized value of the standard deviation of the gray-level gradient of each pixel. The normalized value of the absolute difference between the correlation value of the gray-level co-occurrence matrix of the pixel within the first window and the correlation value of the gray-level co-occurrence matrix of the target image is obtained and denoted as the correlation difference. The normalized value of the absolute difference between the entropy value of the gray-level co-occurrence matrix of the pixel within the first window and the entropy value of the gray-level co-occurrence matrix of the target image is obtained and denoted as the entropy difference. The gray-level change feature is multiplied by the sum of the correlation difference and the entropy difference to obtain the suspected defect degree of the pixel.
3. The intelligent defect identification method for webbing manufacturing according to claim 1, characterized in that, The step of obtaining the detail texture anomaly factor of a pixel based on the LBP feature difference, frequency characteristic difference, and similarity between the first window of a pixel and the target image, as well as the similarity between the first window and surrounding first windows, includes: The LBP algorithm is used to obtain the LBP histogram of the first window of a pixel and the LBP histogram of the target image at different scales. The chi-square distance at each scale is obtained based on the LBP histogram of the first window of the pixel and the LBP histogram of the target image at each scale. Gabor filters of different directions and scales are used to process the pixels within the first window of the pixel and the pixels in the target image, respectively, to obtain the mean window response amplitude and the mean global response amplitude of each filter. Pixels within the first window of the pixel are grouped into a central image block. Pixels within the eight neighborhoods of the central image block are grouped into separate image blocks, denoted as the neighborhood image blocks of the central image block. The mean normalized cross-correlation of the central image block and its neighborhood image blocks is obtained, denoted as the average cross-correlation of the pixel. The detail texture anomaly factor of the pixel is obtained using the chi-square distance at each scale, the mean window response amplitude and the mean global response amplitude of each filter, and the average cross-correlation of the pixel.
4. The intelligent defect identification method for webbing manufacturing according to claim 3, characterized in that, The method of obtaining the detail texture anomaly factor of a pixel by utilizing the chi-square distance at each scale, the mean window response amplitude of each filter, the mean global response amplitude, and the average cross-correlation of the pixel includes: The normalized value of the mean absolute value of the difference between the mean window response amplitude of each filter and the mean global response amplitude is obtained, and multiplied by the normalized value of the mean chi-square distance of each scale to obtain the local texture anomaly feature of the pixel; the normalized value of the difference between the first preset value and the absolute value of the average cross-correlation of the pixel is obtained, and added to the local texture anomaly feature of the pixel to obtain the detail texture anomaly factor of the pixel.
5. The intelligent defect identification method for webbing manufacturing according to claim 1, characterized in that, The step of obtaining the periodic structure disruption factor of a pixel based on the spectral difference, gradient direction difference, and differences in the projection curves in the horizontal and vertical directions between the second window of a pixel and the target image includes: A two-dimensional Fourier transform is performed on the grayscale values of pixels within a second window for each pixel. The zero frequency is shifted to the center of the spectrum, and other frequency components are symmetrically distributed according to high and low frequencies to obtain the spectrum. A predetermined number of frequency points with the largest amplitude in the spectrum are arranged in ascending order of amplitude to form the frequency point sequence of the second window. Similarly, the same analysis is performed on the target image to obtain its frequency point sequence. Two frequency points with the same position in their respective sequences are grouped into a frequency point pair. The Euclidean distance between the positions of the two frequency points in a frequency point pair in the spectrum is obtained and multiplied by the absolute value of the difference in amplitude between the two frequency points in the pair to obtain the distance and amplitude feature difference of the frequency point pair. A grayscale gradient direction histogram of the second window is constructed based on the grayscale gradient direction angle of each pixel, and the frequency in the grayscale gradient direction angle histogram is obtained. The largest gray-level gradient direction angle is used as the main gray-level gradient direction angle of the second window. Similarly, the main gray-level gradient direction angle of the target image is obtained. The absolute value of the difference between the main gray-level gradient direction angles of the second window and the target image is obtained and normalized to obtain the gradient direction angle difference. The similarity between the horizontal projection curve of the pixel in the second window and the horizontal projection curve of the pixel in the target image is obtained and recorded as the horizontal projection similarity. The similarity between the vertical projection curve of the pixel in the second window and the vertical projection curve of the pixel in the target image is obtained and recorded as the vertical projection similarity. The mean values of the horizontal projection similarity and the vertical projection similarity are normalized to obtain the comprehensive similarity. The normalized value of the mean value of the distance and amplitude feature differences of all frequency point pairs, the gradient direction angle difference, and the comprehensive similarity are multiplied to obtain the periodic structure destruction factor of the pixel.
6. The intelligent defect identification method for webbing manufacturing according to claim 1, characterized in that, The step of obtaining the defect region in the target image based on the defect degree of pixels in various clusters includes: The average defect level of all pixels in the cluster with the largest number of pixels is obtained and denoted as the background defect level. The average defect level of all pixels in a cluster is obtained and denoted as the average defect level of that cluster. Clusters with an average defect level greater than the background defect level are denoted as candidate clusters. Candidate clusters with an average defect level greater than a preset defect level threshold are selected as defect clusters. A region is formed in the target image using the pixels in a defect cluster and denoted as the defect region.
7. An intelligent defect identification system for ribbon manufacturing, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent defect identification method for webbing manufacturing as described in any one of claims 1-6.