CN103489182B - A kind of fabric defects detection method based on image projection and svd - Google Patents

Abstract

The invention relates to a fabric defect detection method based on image projection and singular value decomposition. In the training stage, firstly, the flawless fabric image sample is divided into square sub-windows with overlapping; then the obtained sub-windows are projected along the vertical and horizontal directions respectively to obtain the joint projection sequence; finally, the singular value decomposition is performed on the matrix composed of the joint projection sequence to extract Base vector; in the detection stage, the fabric image sample to be detected is divided into square sub-windows without overlap; at the same time, the sub-windows are projected along the vertical and horizontal directions to obtain a joint projection sequence; the basis vector is used to reconstruct the resulting joint projection sequence , and use the reconstruction error to determine whether the sub-window contains defects. The invention makes full use of the warp and weft orientation characteristics of fabric texture and flaws, and analyzes the sequence obtained by projection in the vertical and horizontal directions, which not only greatly reduces the complexity of the method, but also has strong adaptability to different fabric textures and flaw types, especially against linear artifacts.

Description

A Fabric Defect Detection Method Based on Image Projection and Singular Value Decomposition

technical field

The invention belongs to the field of image analysis and processing, and relates to a fabric flaw detection method based on image projection and singular value decomposition, which is applied to the field of automatic detection and control of textile surface quality.

Background technique

Assuming that matrix A is an m×n matrix, singular value decomposition (SVD) can decompose matrix A into three matrices U, Σ and V. Among them, matrix U is an orthogonal matrix of m×n, matrix V is an orthogonal matrix of n×n, matrix Σ is a matrix of m×n, and the elements outside its diagonal are all zero, that is, matrix Σ=diag(σ ₁ , σ ₂ , ..., σ _r ) is a diagonal matrix, r is the rank of matrix A, σ ₁ , σ ₂ , ..., σ _r is the singular value of A matrix. The matrices U and V are called left singular matrix and right singular matrix respectively, and the elements on the diagonal of matrix Σ are in descending order, namely: σ ₁ ≥σ ₂ ≥…≥σ _r ≥0.

Set the diagonal elements of the matrix Σ to σ _k+1 =σ _k+2 =…=σ _r =0 to obtain the reconstruction of matrix A under rank k (k<r) Using the Frobenius norm, its reconstruction error can be expressed as:

${<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; A</span>}<span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; A</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; f</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; r</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

It can be seen from the above formula that the singular value decomposition obtained is the optimal low-rank reconstruction of the A matrix. In the field of image analysis, singular value decomposition is mainly used for image compression, reconstruction and restoration.

In the field of fabric defect detection, Tomczak and Mosorov (2006) used the largest singular value obtained after SVD decomposition to detect fabric defects. Chen and Feng (2010) extracted the singular value mean as a feature for flaw discrimination. Mak and (2010) implemented SVD on the fabric samples, extracted each column of the left singular matrix as a feature sub-image, and projected the image sample to be detected to the feature sub-image, and calculated the sum of the squares of the projected values as a flaw discrimination index.

It is worth noting that the above researchers all directly performed SVD on the original fabric image, and then extracted the corresponding features for flaw detection, did not perform projection operations on the original image and did not make full use of the low-rank reconstruction properties of SVD.

Contents of the invention

The purpose of the present invention is to overcome the shortcomings of existing algorithms, improve the adaptability and real-time performance of the algorithm to different textures and defects, and propose a fabric defect detection method based on image projection and singular value decomposition.

A kind of fabric defect detection method based on image projection and singular value decomposition of the present invention comprises the following steps:

(1) Training stage;

Divide the flawless fabric image sample into square sub-windows with overlapping, and project the sub-windows along the vertical and horizontal directions respectively to obtain a joint projection sequence; then form a matrix with the obtained joint projection sequence, and perform singular value decomposition on the matrix, extract basis vector;

(2) Detection stage;

The fabric image sample to be detected is continuously divided into square sub-windows without overlapping, and the sub-windows are projected along the vertical and horizontal directions to obtain the joint projection sequence y; the base vector pair y obtained in (1) is used to reconstruct y to obtain refactoring of y Calculate reconstruction error And judge whether the sub-window contains defects by reconstructing the error;

The specific implementation is as follows:

In the training phase, the flawless fabric image samples are divided into square sub-windows with a size of w×w in an overlapping manner, wherein 8 pixels≤w≤64 pixels; the overlapping division means that they are in the same row, The latter sub-window is obtained by translating the previous sub-window with a horizontal translation step s. There is an overlap between adjacent sub-windows. The horizontal length of the overlapping portion is w-s. In the adjacent row, the sub-window of the next row is the previous row The sub-windows of the sub-windows are vertically translated with a step length s, and there are overlapping parts between adjacent sub-windows, and the longitudinal length of the overlapping parts is w-s, where 1≤s<w;

Project the sub-windows along the vertical and horizontal directions respectively, that is, calculate the average value of the gray value of all pixels in each column and each row in the sub-window to obtain two projection sequences; then connect the obtained two projection sequences to obtain a joint projection sequence;

Arrange the joint projection sequences corresponding to all sub-windows as the columns of the matrix into a matrix, and perform singular value decomposition on the obtained matrix; extract the first k columns of the left singular matrix as the base vector D, that is, D=[d ₁ , d ₂ ,...,d _k ], where d ₁ , d ₂ ,..., d _k are the column vectors of the first 1 to k columns of the left singular matrix, and 4≤k≤16;

In the detection stage, the fabric image sample to be detected is continuously divided into square sub-windows with a size of w×w, where 8 pixels≤w≤64 pixels; and then the sub-windows are projected along the vertical and horizontal directions respectively according to the method of the training stage , to obtain a joint projection sequence; apply the basis vector to reconstruct the joint projection sequence obtained in the detection stage, and determine whether the sub-window contains defects by calculating the reconstruction error E obtained therefrom.

As a preferred technical solution:

According to the fabric defect detection method based on image projection and singular value decomposition, the fabric is a grayscale image with a bit depth of 8 bits.

As mentioned above, a fabric defect detection method based on image projection and singular value decomposition, the reconstruction refers to the reconstruction obtained under the minimum square error.

As mentioned above, a fabric defect detection method based on image projection and singular value decomposition, the described reconstruction error to determine whether the sub-window contains defects means that when the reconstruction error exceeds a preset threshold, it is considered that the sub-window contains Defects; Determination of the threshold: First, test the non-defective fabric to obtain the cumulative distribution function of its reconstruction error E, and then select the reconstruction error value corresponding to 95% of its cumulative distribution function as the threshold.

As mentioned above, a fabric defect detection method based on image projection and singular value decomposition, the joint projection sequence refers to the two projection sequences obtained by projecting along the vertical and horizontal directions, one in front, the other in the back or a After the other, the sequence obtained in an end-to-end manner.

Beneficial effect:

1. The present invention obtains the preprocessing method of the joint sequence by implementing vertical and horizontal direction projection, which greatly reduces the computational complexity of the method and improves the real-time performance of the method;

2. The method of the present invention is not sensitive to uneven illumination and has better robustness;

3. The method has strong adaptability to different fabric textures and defect types, especially for linear defects.

Description of drawings

The division of the window in Figure 1w×w

Figure 2 Division of horizontal overlapping windows

Figure 3 Division of vertical overlapping windows

Figure 4 is the used defect test image of the present invention

Figure 5 is the final detection result of the test image Figure 4

Fig. 6 image of flaw test used in the present invention

Figure 7 is the final detection result of the test image Figure 6

Fig. 8 image of defect test used in the present invention

Figure 9 is the final detection result of the test image Figure 8

Fig. 10 image of defect test used in the present invention

Figure 11 The final detection results of the test image Figure 10

detailed description

The present invention will be further described below in combination with specific embodiments. It should be understood that these examples are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the teachings of the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by the appended claims of the present application.

A kind of fabric defect detection method based on image projection and singular value decomposition of the present invention comprises the following steps:

(1) Training stage;

Divide the flawless fabric image sample into square sub-windows with overlapping, and project the sub-windows along the vertical and horizontal directions respectively to obtain a joint projection sequence; then form a matrix with the obtained joint projection sequence, and perform singular value decomposition on the matrix, extract basis vector;

(2) Detection stage;

The fabric image sample to be detected is continuously divided into square sub-windows without overlapping, and the sub-windows are projected along the vertical and horizontal directions to obtain the joint projection sequence y; the base vector pair y obtained in (1) is used to reconstruct y to obtain refactoring of y Calculate reconstruction error And judge whether the sub-window contains defects by reconstructing the error;

The specific implementation is as follows:

In the training phase, the flawless fabric image samples are divided into square sub-windows with a size of w×w in an overlapping manner, wherein 8 pixels≤w≤64 pixels; the overlapping division means that they are in the same row, The latter sub-window is obtained by translating the previous sub-window with a horizontal translation step s. There is an overlap between adjacent sub-windows. The horizontal length of the overlapping portion is w-s. In the adjacent row, the sub-window of the next row is the previous row The sub-windows of the sub-windows are vertically translated with a step length s, and there are overlapping parts between adjacent sub-windows, and the longitudinal length of the overlapping parts is w-s, where 1≤s<w;

Project the sub-windows along the vertical and horizontal directions respectively, that is, calculate the average value of the gray value of all pixels in each column and each row in the sub-window to obtain two projection sequences; then connect the obtained two projection sequences to obtain a joint projection sequence;

Arrange the joint projection sequences corresponding to all sub-windows as the columns of the matrix into a matrix, and perform singular value decomposition on the obtained matrix; extract the first k columns of the left singular matrix as the base vector D, that is, D=[d ₁ , d ₂ ,...,d _k ], where d ₁ , d ₂ ,..., d _k are the column vectors of the first 1 to k columns of the left singular matrix, and 4≤k≤16;

In the detection stage, the fabric image sample to be detected is continuously divided into square sub-windows with a size of w×w, where 8 pixels≤w≤64 pixels; and then the sub-windows are projected along the vertical and horizontal directions respectively according to the method of the training stage , to obtain a joint projection sequence; apply the basis vector to reconstruct the joint projection sequence obtained in the detection stage, and determine whether the sub-window contains defects by calculating the reconstruction error E obtained therefrom.

As a preferred technical solution:

According to the fabric defect detection method based on image projection and singular value decomposition, the fabric is a grayscale image with a bit depth of 8 bits.

As mentioned above, a fabric defect detection method based on image projection and singular value decomposition, the reconstruction refers to the reconstruction obtained under the minimum square error.

As mentioned above, a fabric defect detection method based on image projection and singular value decomposition, the described reconstruction error to determine whether the sub-window contains defects means that when the reconstruction error exceeds a preset threshold, it is considered that the sub-window contains Defects; Determination of the threshold: First, test the non-defective fabric to obtain the cumulative distribution function of its reconstruction error E, and then select the reconstruction error value corresponding to 95% of its cumulative distribution function as the threshold.

As mentioned above, a fabric defect detection method based on image projection and singular value decomposition, the joint projection sequence refers to the two projection sequences obtained by projecting along the vertical and horizontal directions, one in front, the other in the back or a After the other, the sequence obtained in an end-to-end manner.

Example 1

Training phase:

(1) Divide the flawless fabric image into sub-windows with a size of 32×32 (pixels) in an overlapping manner, and the horizontal and vertical translation steps are both 1 pixel. As shown in Figure 3;

(2) Project the sub-windows along the vertical and horizontal directions respectively, that is, calculate the average value of the gray value of all pixels in each column and each row in the sub-window, and obtain two projection sequences; then, connect the obtained two projection sequences After that, the joint projection sequence is obtained;

(3) Arrange the joint projection sequences corresponding to all sub-windows as the columns of the matrix into a matrix, and perform singular value decomposition on the obtained matrix; extract the first 4 columns of the left singular matrix as the base vector D.

Determination of the threshold: Firstly, the flawless fabric is tested to obtain the cumulative distribution function of its reconstruction error E, and then the reconstruction error value corresponding to the cumulative distribution function of 95% is selected as the threshold, and the threshold used is 17.

detection stage

(1) The image to be detected, as shown in Figure 4, is continuously divided into sub-windows with a size of 32×32 (pixels) square;

(2) According to the method in the training phase, the sub-windows are projected vertically and horizontally respectively to obtain a joint projection sequence;

(3) Apply the base vector to reconstruct the joint projection sequence obtained in the detection stage, and determine whether the sub-window contains defects by calculating the reconstruction error E obtained. The actual detection results are shown in Figure 5, where the black square Boxes represent subwindows judged to contain blemishes.

Example 2

Training phase:

(1) Divide the flawless fabric image into sub-windows with a size of 32×32 (pixels) in an overlapping manner, and the horizontal and vertical translation steps are both 1 pixel. As shown in Figure 3;

(2) Project the sub-windows along the vertical and horizontal directions respectively, that is, calculate the average value of the gray value of all pixels in each column and each row in the sub-window, and obtain two projection sequences; then, connect the obtained two projection sequences After that, the joint projection sequence is obtained;

(3) Arrange the joint projection sequences corresponding to all sub-windows as the columns of the matrix into a matrix, and perform singular value decomposition on the obtained matrix; extract the first 16 columns of the left singular matrix as the base vector D.

Determination of the threshold: Firstly, the flawless fabric is tested to obtain the cumulative distribution function of its reconstruction error E, and then the reconstruction error value corresponding to the cumulative distribution function of 95% is selected as the threshold, and the threshold used is 27.

detection stage

(1) The image to be detected, as shown in Figure 6, is continuously divided into sub-windows with a size of 32×32 (pixels) square;

(2) According to the method in the training phase, the sub-windows are projected vertically and horizontally respectively to obtain a joint projection sequence;

(3) Apply the base vector to reconstruct the joint projection sequence obtained in the detection stage, and determine whether the sub-window contains defects by calculating the reconstruction error E obtained. The actual detection results are shown in Figure 7, where the black square Boxes represent subwindows judged to contain blemishes.

Example 3

Training phase:

(1) Divide the flawless fabric image into sub-windows with a size of 32×32 (pixels) in an overlapping manner, and the horizontal and vertical translation steps are both 31 pixels. As shown in Figure 3;

(2) Project the sub-windows along the vertical and horizontal directions respectively, that is, calculate the average value of the gray value of all pixels in each column and each row in the sub-window, and obtain two projection sequences; then, connect the obtained two projection sequences After that, the joint projection sequence is obtained;

(3) Arrange the joint projection sequences corresponding to all sub-windows as the columns of the matrix into a matrix, and perform singular value decomposition on the obtained matrix; extract the first 4 columns of the left singular matrix as the base vector D.

Determination of the threshold: Firstly, the flawless fabric is tested to obtain the cumulative distribution function of its reconstruction error E, and then the reconstruction error value corresponding to the cumulative distribution function of 95% is selected as the threshold, and the threshold used is 28.

detection stage

(1) Divide the image to be detected, as shown in Figure 8, into sub-windows with a size of 32×32 (pixels) in a continuous and non-overlapping manner;

(2) According to the method in the training phase, the sub-windows are projected vertically and horizontally respectively to obtain a joint projection sequence;

(3) Apply the base vector to reconstruct the joint projection sequence obtained in the detection stage, and determine whether the sub-window contains defects by calculating the reconstruction error E obtained. The actual detection results are shown in Figure 9, where the black square Boxes represent subwindows judged to contain blemishes.

Example 4

Training phase:

(1) Divide the flawless fabric image into sub-windows with a size of 32×32 (pixels) in an overlapping manner, and the horizontal and vertical translation steps are both 31 pixels. As shown in Figure 3;

(2) Project the sub-windows along the vertical and horizontal directions respectively, that is, calculate the average value of the gray value of all pixels in each column and each row in the sub-window, and obtain two projection sequences; then, connect the obtained two projection sequences After that, the joint projection sequence is obtained;

(3) Arrange the joint projection sequences corresponding to all sub-windows as the columns of the matrix into a matrix, and perform singular value decomposition on the obtained matrix; extract the first 16 columns of the left singular matrix as the base vector D.

Determination of the threshold: Firstly, the flawless fabric is tested to obtain the cumulative distribution function of its reconstruction error E, and then the reconstruction error value corresponding to the cumulative distribution function of 95% is selected as the threshold, and the threshold used is 17.

detection stage

(1) The image to be detected, as shown in Figure 10, is continuously divided into sub-windows with a size of 32×32 (pixels) square;

(2) According to the method in the training phase, the sub-windows are projected vertically and horizontally respectively to obtain a joint projection sequence;

(3) Apply the base vector to reconstruct the joint projection sequence obtained in the detection stage, and determine whether the sub-window contains defects by calculating the reconstruction error E obtained. The actual detection results are shown in Figure 11, where the black square Boxes represent subwindows judged to contain blemishes.

Claims (3)

1., based on a fabric defects detection method for image projection and svd, it is characterized in that comprising the following steps:

(1) training stage;

Indefectible textile image sample is had and is divided into square subwindow overlappingly,

By subwindow edge direction projection in length and breadth respectively, namely calculate in subwindow the mean value of all pixel gray-scale values often arranging and often go, obtain two projection sequence; Then joint projection sequence is obtained after two of gained projection sequence being connected; Then the joint projection sequence of gained is formed a matrix, and svd is implemented to this matrix, extract base vector;

(2) detection-phase;

Textile image sample to be detected continuous zero lap Ground Split is divided into square subwindow, and by the projection of subwindow along direction in length and breadth, obtains joint projection sequences y; Apply the base vector of gained in (1) y is reconstructed, obtain the reconstruct of y calculate reconstructed error and judge whether subwindow comprises flaw by reconstructed error;

Be implemented as follows:

In the training stage, described indefectible textile image sample is had and is divided into the square subwindow that size is w × w overlappingly, wherein 8 pixels≤w≤64 pixel; Described have division overlappingly to refer in the same row, a rear subwindow is that previous subwindow transverse translation step-length s obtains, lap is had between adjacent subwindow, the lateral length of described lap is w-s, in an adjacent row, the subwindow of next line is that the subwindow longitudinal translation step-length s of lastrow obtains, and has lap between adjacent subwindow, the longitudinal length of described lap is w-s, wherein 1≤s<w;

Joint projection sequence corresponding for all subwindows is arranged as a matrix as matrix column, and svd is implemented to the matrix of gained; The front k extracting left singular matrix arranges as base vector D, i.e. D=[d ₁, d ₂..., d _k], wherein d ₁, d ₂..., d _kfor the column vector of 1 to k row before left singular matrix, and there are 4≤k≤16;

At detection-phase, the continuous zero lap of textile image sample to be detected is divided into the square subwindow that size is w × w, wherein 8 pixels≤w≤64 pixel; Then according to the method for training stage, by subwindow edge direction projection in length and breadth respectively, namely calculate in subwindow the mean value of all pixel gray-scale values often arranging and often go, obtain two projection sequence; Then joint projection sequence is obtained after two of gained projection sequence being connected; Apply the joint projection sequence of described base vector to detection-phase gained to be reconstructed, and judge whether subwindow comprises flaw by the reconstructed error E calculating its gained;

Described reconstruct refers to the reconstruct of gained under least squares error;

Described judge whether subwindow comprises flaw and refer to and then think that subwindow comprises flaw when reconstructed error exceedes the threshold value preset by reconstructed error; The determination of threshold value: first test indefectible fabric, obtains the Cumulative Distribution Function of its reconstructed error E, then chooses reconstructed error value corresponding to its Cumulative Distribution Function 95% as threshold value.

2. a kind of fabric defects detection method based on image projection and svd according to claim 1, is characterized in that, described fabric to be bit depth the be gray level image of 8.

3. a kind of fabric defects detection method based on image projection and svd according to claim 1, it is characterized in that, described joint projection sequence, refer to two projection sequence along direction projection gained in length and breadth, one front, another rear or one rear, another front with the sequence of end to end mode gained.