An image recognition method for mesh cloth

By analyzing the same-row mesh area and multi-dimensional feature vectors in the fiber mesh image, combined with Otsu threshold segmentation and clustering algorithm, the problem of defect recognition in glass fiber mesh image recognition is solved, the recognition accuracy and efficiency are improved, and the Hardware requirements.

CN118762211BActive Publication Date: 2024-11-22NANTONG HAOKE NEW MATERIAL TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411244222.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-06
Publication Date
2024-11-22
Estimated Expiration
2044-09-06

AI Technical Summary

Technical Problem

The prior art is difficult to effectively identify small defects such as wrinkles and deformations in glass fiber mesh image recognition, and has high requirements for system hardware, and lacks recognition accuracy and speed.

Method used

By acquiring the fiber mesh cloth image, the horizontal overlap and warp yarn row offset of adjacent target recognition meshes in the same column mesh area set are analyzed, and the overall warp yarn deviation of the fiber mesh cloth is determined by combining the multi-dimensional feature vector. and analyze the uniformity of local distance distribution from a horizontal angle, and use technologies such as Otsu threshold segmentation and clustering algorithm for identification.

Benefits of technology

It improves the accuracy and efficiency of glass fiber mesh image recognition, can better identify defects such as lint and cracking, reduces hardware requirements, and enhances the generalization ability of the identification model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118762211B_ABST
    Figure CN118762211B_ABST
Patent Text Reader

Abstract

The present application relates to the field of image recognition technology, and proposes an image recognition method for mesh cloth, including: obtaining a fiber mesh cloth image; obtaining a set of mesh areas in the same column based on the centroid pixel point of the target recognition mesh in the fiber mesh cloth image and the horizontal distance between the mesh areas; determining the warp yarn row deviation based on the horizontal overlap between two adjacent mesh areas in the set of mesh areas in the same column; determining the overall warp yarn skewness based on the warp yarn row deviation and the horizontal center deviation of the mesh area in the set of mesh areas in the same column; determining the local distance distribution uniformity based on the difference in the horizontal distance between the pixels in two adjacent mesh areas in all the mesh area sets in the same column; obtaining the recognition result of the fiber mesh cloth image based on the overall warp yarn skewness and the local distance distribution uniformity. The present application determines the recognition label by analyzing the distribution characteristics of the warp and weft yarns on the fiber mesh cloth image, thereby improving the accuracy of image recognition.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of image recognition, and in particular to an image recognition method for mesh cloth. Background Art

[0002] Glass fiber mesh is a kind of mesh woven from glass fiber. It usually uses alkali-free glass fiber as raw material and is made through weaving, coating and other processes. Glass fiber mesh has the characteristics of high temperature resistance, corrosion resistance, wear resistance, high strength and fire resistance, and is widely used in various fields. Due to the difficulty of production process and quality control of glass fiber mesh, the product quality varies.

[0003] By performing image recognition on glass fiber mesh cloth, glass fiber mesh cloth with quality defects can be identified to ensure that the quality of the produced glass fiber mesh cloth meets the standards and requirements and improve market competitiveness. Image recognition models based on convolutional neural networks, such as FasterRCNN, ResNet, etc., can automatically capture local features of the image for image recognition tasks and have high image recognition accuracy. However, this method requires a large number of data sets and training sets, has high requirements on the system hardware, and has a slow image recognition speed. The image recognition method based on support vector machines has a high model generalization ability under the training of small sample data sets, and has low requirements on the system hardware, but the recognition accuracy of this method depends on the selection and design of image features. And because glass fiber mesh cloth has a certain weaving method, when small defects such as wrinkles and deformations appear on the glass fiber mesh cloth, traditional image features are difficult to obtain good recognition results. Summary of the invention

[0004] The present application provides an image recognition method for mesh cloth to solve the problem of difficulty in selecting image features when recognizing the image of glass fiber mesh cloth. The technical solution adopted is as follows:

[0005] An embodiment of the present application provides an image recognition method for a mesh cloth, the method comprising the following steps:

[0006] Acquire a fiber mesh cloth image;

[0007] Acquire the target recognition mesh in the fiber mesh cloth image according to the preset range interval, and obtain the mesh area set in the same column based on the centroid pixel point of the target recognition mesh and the horizontal distance between the mesh areas in each mesh area set in the same row;

[0008] Determine the warp row offset between two adjacent target identification meshes in the same column mesh area set based on the horizontal overlap between the two adjacent target identification meshes;

[0009] Determine the overall warp yarn deflection of the fiber mesh image based on the warp yarn row deviation between two adjacent target identification meshes in each mesh area set in the same column and the mesh area horizontal center deviation of the mesh area set in the same column;

[0010] Determine the local distance distribution uniformity of the fiber mesh image based on the difference in horizontal distances between pixels in two adjacent target recognition meshes in all mesh area sets in the same row;

[0011] The recognition result of the fiber mesh cloth image is obtained based on the overall deflection of the warp yarns and the uniformity of the local distance distribution of the fiber mesh cloth image.

[0012] Preferably, the method for obtaining the target recognition mesh in the fiber mesh cloth image according to the preset range interval is:

[0013] The grayscale values ​​of all pixels in the fiber mesh cloth image are used as input, the segmentation threshold is obtained using the Otsu threshold segmentation algorithm, each pixel with a grayscale value greater than the segmentation threshold is regarded as a foreground pixel, and each connected domain obtained based on all foreground pixels using the connected domain analysis method is regarded as a mesh area;

[0014] Each mesh area that is not within the preset range is regarded as a target recognition mesh.

[0015] Preferably, the method for obtaining the mesh area set in the same column based on the centroid pixel point of the target recognition mesh and the horizontal distance between the mesh areas in each mesh area set in the same row is:

[0016] The pixel point at the vertex of the lower left corner of the fiber mesh image is taken as the origin, the coordinates of the centroid points of all target recognition meshes are taken as input, and the clustering results of all centroid points are obtained using the clustering algorithm;

[0017] The set of target recognition meshes corresponding to all centroid points in each cluster in the clustering result is regarded as a set of mesh areas in the same row;

[0018] The absolute value of the difference between the horizontal coordinates of the centroids of the two target recognition meshes is used as the horizontal distance between the two target recognition meshes, and the average value of the horizontal distances between all target recognition meshes in each mesh area set in the same row is used as the horizontal threshold of each mesh area set in the same row;

[0019] The set consisting of all target recognition meshes whose horizontal distances are less than the horizontal threshold in each mesh area set in the same row is regarded as a mesh area set in the same column.

[0020] Preferably, the method for determining the warp yarn row deviation between two adjacent target identification meshes in the same column mesh area set based on the horizontal overlap between the two adjacent target identification meshes is:

[0021] The number of pixel pairs whose horizontal coordinate difference is 0 in two adjacent target recognition meshes in each mesh area set in the same column is taken as the horizontal overlap distance between the two adjacent target recognition meshes;

[0022] The maximum absolute value of the difference between the horizontal coordinates of all pixels in each target recognition mesh in each mesh area set in the same column is calculated as the width of each target recognition mesh, the sum of the widths of two adjacent target recognition meshes in each mesh area set in the same column is used as the denominator, and the ratio of the horizontal overlap distance of the two adjacent target recognition meshes to the denominator is used as the horizontal overlap degree of the two adjacent target recognition meshes;

[0023] The difference between the maximum value of all horizontal overlaps in each mesh area set in the same column and the horizontal overlaps of two adjacent target recognition meshes in each mesh area set in the same column is taken as the warp row offset between the two adjacent target recognition meshes.

[0024] Preferably, the method for determining the overall warp yarn deflection of the fiber mesh image based on the warp yarn row deviation between two adjacent target identification meshes in each mesh area set in the same column and the mesh area horizontal center offset of the mesh area set in the same column is:

[0025] Determine the horizontal center offset of each mesh area set in the same column based on the multi-dimensional recognition vector of all mesh areas in the fiber mesh cloth image;

[0026] Calculate the variance of the warp yarn row deviation between all two adjacent target identification meshes in each mesh area set in the same column, take the sum of the mean of the warp yarn row deviation between all two adjacent target identification meshes in each mesh area set in the same column and 1 as the denominator, and take the ratio of the variance to the denominator as the first multiplication factor;

[0027] The product of the first multiplication factor and the horizontal center offset of each mesh area set in the same column is taken as the warp yarn skewness on both sides of each mesh area set in the same column, and the average of the warp yarn skewness on both sides of all mesh area sets in the same column is taken as the overall warp yarn skewness of the fiber mesh cloth image.

[0028] Preferably, the method for determining the horizontal center offset of each mesh area set in the same column based on the multi-dimensional recognition vector of all mesh areas in the fiber mesh cloth image is:

[0029] A multi-dimensional recognition vector of each target recognition mesh is constructed based on the boundary information of each target recognition mesh in the RGB space and the color vector angle of the pixel points in each target recognition mesh;

[0030] The normalized result of the Euclidean distance between the centroids of any two target recognition meshes in each mesh area set in the same column is used as the first weight, and the product of the first weight and the DTW distance between the color difference concentration vectors of any two target recognition meshes in each mesh area set in the same column is used as the first difference value;

[0031] The difference between 1 and the first weight is used as the second weight, and the product of the second weight and the DTW distance between the local side information vectors of any two target recognition meshes in each mesh area set in the same column is used as the second difference value;

[0032] The sum of the first difference value and the second difference value is used as the regional feature recognition difference between any two target recognition meshes in each mesh area set in the same column;

[0033] The accumulation result of the coefficient of variation of the set of cosine similarities between the multi-dimensional recognition vectors of each target recognition mesh and the other target recognition meshes in each mesh area set in the same column on each mesh area set in the same column is taken as the multi-dimensional recognition instability coefficient of each mesh area set in the same column;

[0034] The absolute value of the difference between the mean of the horizontal coordinates of the centroid points of all target recognition meshes in each mesh area set in the same column and the horizontal coordinates of the centroid point of each target recognition mesh is taken as the same-column distance, and the product of the same-column distance and the maximum value of all regional feature recognition differences corresponding to each target recognition mesh is accumulated on each mesh area set in the same column as the horizontal center offset of the mesh area of ​​each mesh area set in the same column.

[0035] Preferably, the method for acquiring the multi-dimensional recognition feature vector of the target recognition mesh is:

[0036] Perform DCT transformation on each target recognition mesh, and use the vector composed of all high-frequency coefficients obtained by DCT transformation of each target recognition mesh in descending order as the local side information vector of each target recognition mesh;

[0037] Determine the corresponding position of each target recognition mesh in the RGB image of the glass fiber mesh, take the horizontal direction as the reference direction, calculate the color vector angle between any two pixel points in the corresponding position, take the angle value of each color vector angle and the frequency of occurrence of each color vector angle as a two-tuple, and use the sequence composed of all the two-tuples in ascending order of the color vector angle as the color difference concentration vector of each target recognition mesh;

[0038] The vector composed of the color difference concentration vector and the local edge information vector of each target recognition mesh is used as the multi-dimensional recognition vector of each target recognition mesh.

[0039] Preferably, the method for determining the uniformity of the local distance distribution of the fiber mesh image based on the difference in the horizontal distance between the pixels in two adjacent target recognition meshes in all mesh area sets in the same row is:

[0040] The minimum value of the ratio of the maximum horizontal distances between the pixels in two adjacent target recognition meshes in each mesh area set in the same row is taken as the local distance uniformity between the two adjacent target recognition meshes;

[0041] The sum of the local distance uniformity between all target recognition meshes in each mesh area set in the same line and the average of the accumulated results over all mesh area sets in the same line are taken as the local distance distribution uniformity of the fiber mesh cloth image.

[0042] Preferably, the method for obtaining the maximum horizontal distance between pixels in the target recognition mesh is:

[0043] The maximum value of the absolute value of the difference between the ordinates of all pixels in each target recognition mesh is taken as the maximum horizontal distance between the pixels in each target recognition mesh.

[0044] Preferably, the method for obtaining the recognition result of the fiber mesh cloth image based on the overall deflection of the warp yarns and the uniformity of the local distance distribution of the fiber mesh cloth image is:

[0045] The sum of the overall skewness of the warp yarns of the fiber mesh cloth image and 1 is used as the denominator, and the ratio of the local distance distribution uniformity of the fiber mesh cloth image to the denominator is used as the quality recognition coefficient of the fiber mesh cloth image;

[0046] The quality recognition coefficient and recognition label of the fiber mesh cloth image are used as training to complete the training of the recognition model, and the recognition result of the fiber mesh cloth image is obtained by using the recognition model.

[0047] The beneficial effects of the present application are as follows: the present application determines the warp yarn row deviation between two adjacent target recognition meshes in the same column mesh area set in the fiber mesh cloth image, and determines the overall warp yarn skewness of the fiber mesh cloth image in combination with the multi-dimensional feature vector of the target recognition mesh. The overall warp yarn skewness takes into account the influence of defects such as fuzzing and cracking on the position offset and image information between the target recognition meshes in the adjacent warp yarn determination area, and can adaptively adjust the weights of different feature vectors according to the size of the Euclidean distance; secondly, the uniformity of the distance between the target recognition meshes corresponding to the adjacent multiple warp yarns in the same column mesh area set is analyzed from a horizontal perspective, so as to determine the uniformity of the local distance distribution of the fiber mesh cloth image, which can avoid the influence of the horizontal and vertical staggered structure in the fiber mesh cloth image on the recognition of subtle defects, so that the recognition model trained with the subsequent quality recognition coefficient has better recognition effect when facing the fiber mesh cloth image. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0049] Figure 1 A schematic flow chart of an image recognition method for a mesh cloth provided in one embodiment of the present application;

[0050] Figure 2 A flowchart of an implementation method for an image recognition method of a mesh cloth provided in one embodiment of the present application;

[0051] Figure 3 A schematic diagram of the horizontal overlapping distance between two adjacent target recognition meshes provided in one embodiment of the present application. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0053] See also Figure 1 , which shows a flow chart of an image recognition method for a mesh cloth provided by an embodiment of the present application, the method comprising the following steps:

[0054] Step S001, obtaining a fiber mesh cloth image.

[0055] In this application, a glass fiber mesh cloth with a plain weave structure is used as an identification object. Specifically, the glass fiber mesh cloth is composed of untwisted roving and twisted spun yarn arranged horizontally and vertically in an alternating manner, wherein the fibers of the untwisted roving are arranged in parallel and are relatively loose, and have good wettability to resin, but are prone to cracking, pilling, and are difficult to weave, while the twisted spun yarn is just the opposite. Therefore, two strands of twisted spun yarn are usually used to bind the untwisted roving to fix the shape and size of the glass fiber mesh cloth, thereby ensuring the strength and stability of the glass fiber mesh cloth.

[0056] Furthermore, before acquiring the image of the glass fiber mesh cloth, a black color with a large color difference from the glass fiber mesh cloth is selected as the background, the glass fiber mesh cloth is fixed so that it is naturally spread out, and the meshes in the glass fiber mesh cloth are horizontally distributed. The fixed glass fiber mesh cloth is photographed using a CCD camera to acquire an RGB image of the glass fiber mesh cloth, and the obtained RGB image is converted into a grayscale image.

[0057] During the image acquisition process, noise caused by factors such as the camera and the environment will appear. In order to avoid the influence of noise on subsequent image processing, the grayscale image is used as input, and the grayscale image is denoised using median filtering. The denoised image is recorded as a fiber mesh cloth image. Median filtering is a well-known technology and the specific process will not be repeated here.

[0058] At this point, the fiber mesh cloth image is obtained for subsequent acquisition of image features.

[0059] Step S002, obtaining a mesh area set in the same column based on the centroid pixel point of the target recognition mesh and the horizontal distance between mesh areas in the same row mesh area set; determining the warp row offset between two adjacent target recognition meshes in the mesh area set in the same column based on the horizontal overlap between the two adjacent target recognition meshes.

[0060] The uneven distribution of yarns in the glass fiber mesh will cause the areas where the yarns are too dense to be too thick, while the rest of the areas will be too thin, thus affecting the balance and stability of the entire glass fiber mesh. The deviation of the yarns will cause the shape and size of the glass fiber mesh to be irregular, which will lead to problems such as wrinkling and deformation of the glass fiber mesh. Therefore, this application realizes the recognition of the fiber mesh image based on the distribution characteristics of the warp yarns and weft yarns in the fiber mesh image. The implementation process of the whole solution is as follows: Figure 2 shown.

[0061] Furthermore, due to the large color difference between the glass fiber mesh cloth and the background, the grayscale values ​​of all pixels in the fiber mesh cloth image are used as input, and the segmentation threshold is obtained using the Otsu threshold segmentation algorithm. Each pixel with a grayscale value greater than the segmentation threshold is used as a foreground pixel, and each connected domain obtained based on all foreground pixels using the connected domain analysis method is used as a mesh area. The Otsu threshold segmentation algorithm and connected domain analysis are both well-known technologies, and the specific process will not be repeated. Secondly, when the untwisted coarse yarn in the glass fiber mesh cloth cracks and fuzzes, it will cause the nearby meshes to be too small, and when the yarn breaks, it will cause multiple mesh areas to be connected, so the mesh area is preliminarily screened by setting the area interval.

[0062] Specifically, professionals screen out glass fiber mesh cloth that meets the standards, obtain an image of the glass fiber mesh cloth, obtain all mesh areas in the image, calculate the area of ​​each mesh area, and calculate the average of all areas. and standard deviation ,use The lower limit of the mesh area is calculated by the criterion and upper limit ,in , , will be lower bounded by , Upper limit Determined interval As the range of mesh area, The criterion is a well-known technology, and the specific process is not repeated. Secondly, each mesh area in the fiber mesh cloth image is compared with the range interval, and each mesh area that is not within the range interval is regarded as a target identification mesh.

[0063] Furthermore, the pixel point at the vertex of the lower left corner of the fiber mesh cloth image is taken as the origin, the coordinates of the centroid points of all target recognition meshes are taken as input, the difference between the vertical coordinates of the two centroid points is taken as the metric distance, and the K-means clustering algorithm is used to divide all the centroid points into K clusters. The K-means clustering algorithm is a well-known technology, and the specific process is not repeated. For any cluster cluster, the set of target recognition meshes corresponding to all centroid points in each cluster cluster is taken as a mesh area set in the same row. For any mesh area set in the same row, the absolute value of the difference between the horizontal coordinates of the centroid points of two target recognition meshes in each mesh area set in the same row is taken as the horizontal distance between the two target recognition meshes, the mean of the horizontal distances between all target recognition meshes in each mesh area set in the same row is taken as the horizontal threshold of each mesh area set in the same row, and the set of all target recognition meshes in each mesh area set in the same row whose horizontal distance is less than the horizontal threshold is taken as a mesh area set in the same column. The purpose of such clustering is to identify the degree of skewness and uniformity of the warp and weft yarns by analyzing the edge features of adjacent warp and weft yarns and according to the column and row distribution of the target recognition mesh.

[0064] For any mesh area set in the same column, taking the first mesh area set in the same column as an example, the difference between the horizontal coordinates of the pixels in any two adjacent target recognition meshes in the first mesh area set in the same column is counted, and the number of pixel pairs with a horizontal coordinate difference of 0 in the two adjacent target recognition meshes is taken as the horizontal overlap distance of the two adjacent target recognition meshes, such as Figure 3 shown.

[0065] Based on the above analysis, the warp line deviation is constructed here to characterize the degree of deviation of the warp lines corresponding to two adjacent target recognition meshes in each mesh area set in the same column. The warp line deviation between the kth and k+1th target recognition meshes in the first mesh area set in the same column is calculated as follows:

[0066]

[0067]

[0068] In the formula, is the horizontal overlap between the kth and k+1th target recognition meshes, is the horizontal overlap distance between the kth and k+1th target recognition meshes, , are the widths of the kth and k+1th target recognition meshes in the first mesh area set in the same column, respectively. The width is equal to the maximum absolute value of the difference between the horizontal coordinates of all pixels in the target recognition mesh;

[0069] is the warp line offset between the kth and k+1th target recognition meshes in the first mesh area set in the same column, It is the maximum value of all horizontal overlaps in the first set of mesh areas in the same column.

[0070] The larger the horizontal overlapping area of ​​the kth and k+1th target recognition meshes in the first mesh area set in the same column, the larger the horizontal overlapping distance. The larger the value of The larger the value of , the smaller the deflection of the warp yarns corresponding to the kth and k+1th target recognition meshes; that is, The smaller the value of , the more likely it is that the coordinates of the centroid points of the kth and k+1th target recognition meshes are located on the same vertical line.

[0071] At this point, the warp yarn row deviation between two adjacent target recognition meshes is obtained, which is used for the subsequent calculation of the overall warp yarn deflection of the fiber mesh cloth image.

[0072] Step S003, determining the overall warp yarn skewness of the fiber mesh image based on the warp yarn row deviation between two adjacent target identification meshes in each mesh area set in the same column and the mesh area horizontal center offset of the mesh area set in the same column.

[0073] Furthermore, describing the image space features only by the centroid is too rough, and clustering only by the centroid will result in recognition errors, because the untwisted coarse yarns at different target recognition meshes have different degrees of cracking and fuzzing. Therefore, it is necessary to combine the multi-dimensional edge information features of the target recognition mesh for subsequent recognition processes.

[0074] Specifically, for any target mesh area, take the bth target mesh area in the first set of mesh areas in the same column as an example, perform DCT (Discrete Cosine Transform) transformation on the bth target mesh area, discard the low-frequency coefficients, and retain the high-frequency coefficients. The purpose of doing so is that the high-frequency coefficients generally store the boundary and texture information of the image, which can be used to identify the cracking, fluffing and breakage of the mesh area edge. DCT transformation is a well-known technology, and the specific process is not repeated. The vector composed of all high-frequency coefficients of the bth target mesh area in descending order is used as the local edge information vector of the bth target mesh area.

[0075] Furthermore, if edge cracking, fluffing or breaking occurs, then the pixels that differ in color from the pixels at the flat part inside the mesh area are often concentrated at the cracking, fluffing and breaking parts, that is, the color vector angles of the pixels with color differences will also be concentrated in a range. Therefore, the corresponding position of the b-th target mesh area is determined in the collected RGB image, the horizontal direction is used as the reference direction, the color vector angle between any two pixels in the corresponding position is calculated, and the number of occurrences of each color vector angle is counted, and the array consisting of the angle value of each color vector angle and the frequency of each color vector angle is taken as a binary group, and the sequence composed of all binary groups in ascending order of the color vector angle is taken as the color difference concentration vector of the b-th target mesh area. The color vector angle is a well-known technology, and the specific process is not repeated. The vector composed of the color difference concentration vector and the local edge information vector of the b-th target mesh area is taken as the multi-dimensional recognition vector of the b-th target mesh area.

[0076] Secondly, when the target recognition meshes in the first set of mesh regions in the same column have fuzzing, cracking, etc., the warp yarns corresponding to different target recognition meshes in the first set of mesh regions in the same column will be offset to varying degrees, and the centroid of the target recognition mesh will also be offset to varying degrees under the influence of the change of edge pixels. The Euclidean distance between any two centroids of the target recognition meshes in the first set of mesh regions in the same column is calculated respectively, and all the Euclidean distances are Z-score normalized. Z-score normalization is a well-known technology, and the specific process is not repeated here.

[0077] Based on the above analysis, the horizontal center offset of the mesh area is constructed here to characterize the degree to which the center of mass of different target recognition meshes in each mesh area set in the same column is affected by defects. Calculate the horizontal center offset of the mesh area of ​​the first mesh area set in the same column:

[0078]

[0079]

[0080]

[0081] In the formula, is the difference in regional feature recognition between the i-th and b-th target recognition meshes in the first mesh region set in the same column, is the normalized result of the Euclidean distance between the centroids of the i-th and b-th target recognition meshes, , are the local side information vectors of the ith and bth target recognition meshes, respectively, , are the color difference concentration vectors of the ith and bth target recognition meshes, is the DTW distance function, , They are and , and The DTW distance between them, DTW distance is a well-known technology, and the specific process will not be repeated;

[0082] is the multidimensional identification instability coefficient of the first mesh area set in the same column, is a set of cosine similarities between the multidimensional recognition vectors of the i-th target recognition mesh and all other target recognition meshes. Is a collection The coefficient of variation of , N is the number of target recognition meshes in the first mesh area set in the same column;

[0083] is the horizontal center offset of the mesh area of ​​the first mesh area set in the same column, is the horizontal coordinate of the centroid of the i-th target recognition mesh, is the mean of the horizontal coordinates of the centroids of all target recognition meshes in the first mesh area set in the same column, It is the maximum value of the regional feature recognition difference between the i-th and all other target recognition meshes in the first mesh area set in the same column.

[0084] Among them, the closer the spatial positions of the i-th and b-th target recognition meshes in the first mesh area set in the same column are, the smaller the normalized result of the Euclidean distance between the centroids is, and the first weight The smaller the value, the greater the second weight The larger the value of is, if there is a defective area in one of the target recognition meshes between the i-th and b-th target recognition meshes, the greater the impact on the boundary and texture information of the other target recognition mesh, and the metric result between the local edge information vectors is given. Greater weight, second difference value The larger the value of ; similarly, if the spatial positions of the i-th and b-th target recognition meshes are closer, the greater the impact of the defect area on the color vector angle of the other target recognition mesh when one of the i-th and b-th target recognition meshes has a defect area, the more the measurement result between the color difference concentration vectors should be given Larger weight, first difference value The larger the value of is, the more locations where defects occur in the first mesh area set in the same column, and the more discrete the distribution of the cosine similarity between the multidimensional recognition vectors of the i-th target recognition mesh and all other target recognition meshes. The larger the value of The larger the value, the smaller the probability that the centroid of the target recognition mesh in the first mesh area in the same column is located on a vertical line, and the smaller the difference in the horizontal coordinates of the centroids of all target recognition meshes. The larger the value of The larger the value is, the more serious the impact of fluffing, cracking and other phenomena on the target recognition mesh in the first mesh area set in the same column will be.

[0085] Furthermore, the overall warp yarn deflection of the fiber mesh image is determined based on the warp yarn row deviation between all two adjacent target identification meshes in each mesh area set in the same column and the mesh area horizontal center offset of each mesh area set in the same column:

[0086]

[0087]

[0088] In the formula, is the warp yarn deflection on both sides of the a-th mesh area set in the same column, is the mean value of the warp line deviation between all two adjacent target recognition meshes in the mesh area set of the same column a, The variance of the warp line offset between all two adjacent target recognition meshes in the a-th mesh area set in the same column, is the horizontal center offset of the mesh area of ​​the ath mesh area set in the same column;

[0089] A is the overall skewness of the warp yarns of the fiber mesh image, and M is the number of mesh area sets in the same column.

[0090] Among them, the less the image features in the a-th mesh area set are affected by the defect phenomenon, the smaller the skewness between the two warps corresponding to the a-th mesh area set, and the more concentrated the distribution of the warp line deviation between all two adjacent target recognition meshes in the a-th mesh area set. The smaller the value of The smaller the value of The smaller the value of is, the less the target recognition mesh in the mesh area set in the same column a is affected by the phenomena such as fuzzing and cracking. The smaller the value of is, that is, the smaller the value of A is, the lower the possibility of defects in the fiber mesh cloth image is, and the image features of the fiber mesh cloth image will be less affected.

[0091] At this point, the overall skewness of the warp yarns of the fiber mesh cloth image is obtained, which is used to subsequently obtain the quality recognition coefficient of the fiber mesh cloth image.

[0092] Step S004, determining the uniformity of the local distance distribution of the fiber mesh cloth image based on the difference in horizontal distances between pixel points in two adjacent target recognition meshes in the same mesh area set.

[0093] According to the above steps, the deflection of the warp yarn in the fiber mesh cloth image is analyzed. Secondly, in order to avoid the influence of the horizontal and vertical staggered arrangement of untwisted coarse yarn and twisted fine yarn on the analysis results of the warp yarn distribution characteristics and obtain more accurate image recognition results, this application continues to analyze the uniformity of the warp yarn in the fiber mesh cloth image.

[0094] Specifically, for any set of mesh areas in the same line, taking the jth mesh area set in the same line as an example, the absolute value of the difference between the vertical coordinates of all pixels in each target recognition mesh in the jth mesh area set in the same line is calculated, and the maximum value of the absolute values ​​of all the differences in each target recognition mesh is taken as the maximum horizontal distance between the pixels in each target recognition mesh.

[0095] The local distance distribution uniformity of the fiber mesh image is determined based on the maximum horizontal distance between the pixels in each target recognition mesh in the set of all mesh areas in the same row:

[0096]

[0097]

[0098] In the formula, is the local distance uniformity between the hth and h+1th target recognition meshes in the jth peer mesh region set, is the minimum function, , are the maximum horizontal distances between pixels in the hth and h+1th target recognition meshes, respectively;

[0099] v is the uniformity of the local distance distribution of the fiber mesh cloth image, m is the number of mesh area sets in the same line, and n is the number of target recognition meshes in the jth mesh area set in the same line.

[0100] Among them, when the distance between the three warp yarns corresponding to the hth and h+1th target recognition meshes in the jth same mesh area set is closer, the pixel points in the hth and h+1th target recognition meshes are more concentrated, and the difference between the vertical coordinates of the pixel points is smaller. , The closer the value of is, The closer the value of is to 1; that is, the larger the value of v is, the more consistent the distances between adjacent warp yarns corresponding to the same mesh area set are, and the more uniform the overall distribution of the warp yarns is.

[0101] At this point, the local distance distribution uniformity of the fiber mesh cloth image is obtained, which is used to calculate the quality recognition coefficient of the fiber mesh cloth image in combination with the overall skewness of the warp yarns of the fiber mesh cloth image.

[0102] Step S005 , obtaining a recognition result of the fiber mesh cloth image based on the overall deflection of the warp yarns and the uniformity of the local distance distribution of the fiber mesh cloth image.

[0103] According to the above steps, the degree to which the warp yarns in the fiber mesh cloth image are affected by the defects is evaluated from the vertical and horizontal dimensions. Secondly, the quality recognition coefficient of the fiber mesh cloth image is calculated based on the overall skewness and local distance distribution uniformity of the warp yarns in the fiber mesh cloth image:

[0104]

[0105] Where S is the quality recognition coefficient of the fiber mesh cloth image, and A and v are the overall skewness and local distance distribution uniformity of the warp yarn of the fiber mesh cloth image, respectively.

[0106] Furthermore, the above steps are repeated to obtain the quality recognition coefficients of B flawless and defective fiber mesh cloth images, respectively, and the identification labels of the flawless and defective fiber mesh cloth images are marked as 1 and 0, respectively, and the size of B takes an empirical value of 100. Secondly, the vector composed of the quality recognition coefficient and the identification label of each fiber mesh cloth image is used as a recognition vector, and the SVM (Support Vector Machine) model trained with the recognition vector is used as the recognition model of the fiber mesh cloth image. The model training is a well-known technology, and the specific process is not repeated here. For each glass fiber mesh cloth produced subsequently, the quality recognition coefficient of each glass fiber mesh cloth is obtained based on the above process and input into the recognition model to obtain the recognition result of each glass fiber mesh cloth. The recognition result is divided into two types: qualified quality and unqualified quality.

[0107] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the principles of the present application shall be included in the protection scope of the present application.

Claims

1. An image recognition method for a mesh cloth, characterized in that: The method comprises the following steps: Acquire a fiber mesh cloth image; Acquire the target recognition mesh in the fiber mesh cloth image according to the preset range interval, and obtain the mesh area set in the same column based on the centroid pixel point of the target recognition mesh and the horizontal distance between the mesh areas in each mesh area set in the same row; Determine the warp row offset between two adjacent target identification meshes in the same column mesh area set based on the horizontal overlap between the two adjacent target identification meshes; Determine the overall warp yarn deflection of the fiber mesh image based on the warp yarn row deviation between two adjacent target identification meshes in each mesh area set in the same column and the mesh area horizontal center deviation of the mesh area set in the same column; Determine the local distance distribution uniformity of the fiber mesh image based on the difference in horizontal distances between pixels in two adjacent target recognition meshes in all mesh area sets in the same row; The recognition result of the fiber mesh cloth image is obtained based on the overall deflection of the warp yarns and the uniformity of the local distance distribution of the fiber mesh cloth image; The method for determining the warp yarn row deviation between two adjacent target identification meshes based on the horizontal overlap between the two adjacent target identification meshes in the same column mesh area set is: The number of pixel pairs whose horizontal coordinate difference is 0 in two adjacent target recognition meshes in each mesh area set in the same column is taken as the horizontal overlap distance between the two adjacent target recognition meshes; The maximum absolute value of the difference between the horizontal coordinates of all pixels in each target recognition mesh in each mesh area set in the same column is calculated as the width of each target recognition mesh, the sum of the widths of two adjacent target recognition meshes in each mesh area set in the same column is used as the denominator, and the ratio of the horizontal overlap distance of the two adjacent target recognition meshes to the denominator is used as the horizontal overlap degree of the two adjacent target recognition meshes; The maximum value of all horizontal overlaps in each mesh area set in the same column and the difference between the horizontal overlaps of two adjacent target recognition meshes in each mesh area set in the same column are taken as the warp row offset between the two adjacent target recognition meshes; The method for determining the overall warp yarn deflection of the fiber mesh image based on the warp yarn row deviation between two adjacent target identification meshes in each mesh area set in the same column and the mesh area horizontal center offset of the mesh area set in the same column is: Determine the horizontal center offset of each mesh area set in the same column based on the multi-dimensional recognition vector of all mesh areas in the fiber mesh cloth image; Calculate the variance of the warp yarn row deviation between all two adjacent target identification meshes in each mesh area set in the same column, take the sum of the mean of the warp yarn row deviation between all two adjacent target identification meshes in each mesh area set in the same column and 1 as the denominator, and take the ratio of the variance to the denominator as the first multiplication factor; The product of the first multiplication factor and the horizontal center offset of each mesh area set in the same column is used as the warp yarn deflection on both sides of each mesh area set in the same column, and the average of the warp yarn deflections on both sides of all mesh area sets in the same column is used as the overall warp yarn deflection of the fiber mesh cloth image; The method for determining the uniformity of the local distance distribution of the fiber mesh image based on the difference in the horizontal distance between the pixels in two adjacent target recognition meshes in all mesh area sets in the same row is: The minimum value of the ratio of the maximum horizontal distances between the pixels in two adjacent target recognition meshes in each mesh area set in the same row is taken as the local distance uniformity between the two adjacent target recognition meshes; The local distance uniformity between all target recognition meshes in each mesh area set in the same line is taken as the average value of the cumulative result of all mesh area sets in the same line as the local distance distribution uniformity of the fiber mesh cloth image; The method for obtaining the recognition result of the fiber mesh cloth image based on the overall deflection of the warp yarns and the uniformity of the local distance distribution of the fiber mesh cloth image is: The sum of the overall skewness of the warp yarns of the fiber mesh cloth image and 1 is used as the denominator, and the ratio of the local distance distribution uniformity of the fiber mesh cloth image to the denominator is used as the quality recognition coefficient of the fiber mesh cloth image; The quality recognition coefficient and recognition label of the fiber mesh cloth image are used as training to complete the training of the recognition model, and the recognition result of the fiber mesh cloth image is obtained by using the recognition model.

2. The image recognition method for mesh cloth according to claim 1, characterized in that: The method for obtaining the target recognition mesh in the fiber mesh cloth image according to the preset range interval is: The grayscale values ​​of all pixels in the fiber mesh cloth image are used as input, the segmentation threshold is obtained using the Otsu threshold segmentation algorithm, each pixel with a grayscale value greater than the segmentation threshold is regarded as a foreground pixel, and each connected domain obtained based on all foreground pixels using the connected domain analysis method is regarded as a mesh area; Each mesh area that is not within the preset range is regarded as a target recognition mesh.

3. The image recognition method for mesh cloth according to claim 1, characterized in that: The method for obtaining the mesh area set in the same column based on the centroid pixel point of the target recognition mesh and the horizontal distance between the mesh areas in each mesh area set in the same row is: The pixel point at the vertex of the lower left corner of the fiber mesh image is taken as the origin, the coordinates of the centroid points of all target recognition meshes are taken as input, and the clustering results of all centroid points are obtained using the clustering algorithm; The set of target recognition meshes corresponding to all centroid points in each cluster in the clustering result is regarded as a set of mesh areas in the same row; The absolute value of the difference between the horizontal coordinates of the centroids of the two target recognition meshes is used as the horizontal distance between the two target recognition meshes, and the average value of the horizontal distances between all target recognition meshes in each mesh area set in the same row is used as the horizontal threshold of each mesh area set in the same row; The set consisting of all target recognition meshes whose horizontal distances are less than the horizontal threshold in each mesh area set in the same row is regarded as a mesh area set in the same column.

4. The image recognition method for mesh cloth according to claim 1, characterized in that: The method for determining the horizontal center offset of each mesh area set in the same column based on the multi-dimensional recognition vector of all mesh areas in the fiber mesh cloth image is: A multi-dimensional recognition vector of each target recognition mesh is constructed based on the boundary information of each target recognition mesh in the RGB space and the color vector angle of the pixel points in each target recognition mesh; The normalized result of the Euclidean distance between the centroids of any two target recognition meshes in each mesh area set in the same column is used as the first weight, and the product of the first weight and the DTW distance between the color difference concentration vectors of any two target recognition meshes in each mesh area set in the same column is used as the first difference value; The difference between 1 and the first weight is used as the second weight, and the product of the second weight and the DTW distance between the local side information vectors of any two target recognition meshes in each mesh area set in the same column is used as the second difference value; The sum of the first difference value and the second difference value is used as the regional feature recognition difference between any two target recognition meshes in each mesh area set in the same column; The accumulation result of the coefficient of variation of the set of cosine similarities between the multi-dimensional recognition vectors of each target recognition mesh and the other target recognition meshes in each mesh area set in the same column on each mesh area set in the same column is taken as the multi-dimensional recognition instability coefficient of each mesh area set in the same column; The absolute value of the difference between the mean of the horizontal coordinates of the centroid points of all target recognition meshes in each mesh area set in the same column and the horizontal coordinates of the centroid point of each target recognition mesh is taken as the same-column distance, and the product of the same-column distance and the maximum value of all regional feature recognition differences corresponding to each target recognition mesh is accumulated on each mesh area set in the same column as the horizontal center offset of the mesh area of ​​each mesh area set in the same column.

5. The image recognition method for mesh cloth according to claim 4, characterized in that: The method for obtaining the multi-dimensional recognition feature vector of the target recognition mesh is: Perform DCT transformation on each target recognition mesh, and use the vector composed of all high-frequency coefficients obtained by DCT transformation of each target recognition mesh in descending order as the local side information vector of each target recognition mesh; Determine the corresponding position of each target recognition mesh in the RGB image of the glass fiber mesh, take the horizontal direction as the reference direction, calculate the color vector angle between any two pixel points in the corresponding position, take the angle value of each color vector angle and the frequency of occurrence of each color vector angle as a two-tuple, and use the sequence composed of all the two-tuples in ascending order of the color vector angle as the color difference concentration vector of each target recognition mesh; The vector composed of the color difference concentration vector and the local edge information vector of each target recognition mesh is used as the multi-dimensional recognition vector of each target recognition mesh.

6. The image recognition method for mesh cloth according to claim 1, characterized in that: The method for obtaining the maximum horizontal distance between pixels in the target recognition mesh is: The maximum value of the absolute value of the difference between the ordinates of all pixels in each target recognition mesh is taken as the maximum horizontal distance between the pixels in each target recognition mesh.