CN102289677B - Method for analyzing image based on principal component analysis and method applicable to detection of defects of fabric - Google Patents
Method for analyzing image based on principal component analysis and method applicable to detection of defects of fabric Download PDFInfo
- Publication number
- CN102289677B CN102289677B CN 201110219310 CN201110219310A CN102289677B CN 102289677 B CN102289677 B CN 102289677B CN 201110219310 CN201110219310 CN 201110219310 CN 201110219310 A CN201110219310 A CN 201110219310A CN 102289677 B CN102289677 B CN 102289677B
- Authority
- CN
- China
- Prior art keywords
- vector
- matrix
- subwindow
- groups
- pivot
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Abstract
The invention belongs to the field of image analysis processing, is applicable to the field of automatic detection and control of the surface quality of fabrics, and relates to a method for analyzing an image based on principal component analysis and a method applicable to detection of the defects of fabric. The method for analyzing the image based on the principal component analysis comprises the following steps of: firstly, expanding gray values in an original image sample into two groups of vectors according to rows and lines; secondly, performing template operation on the two groups of vectors, and respectively performing principal component analysis on the two groups of vectors which are subjected to template operation to obtain corresponding principal component matrixes; and finally, performing projection operation on a sample to be detected by using the two principal component matrixes, and calculating the similarity of the sample after projection and the sample before projection to analyze the characteristics of the image. The invention has the advantages that: non-uniform illumination can be eliminated without the conventional pre-processing step; calculation in a detection period is simple; the original fabric sample is respectively expanded according to the rows and the lines and then subjected to template operation, so that the longitude and latitude orientation characteristics of fabric texture can be fully utilized, the defects can be highlighted, and the random interference of the texture is restrained; and detection accuracy rate is improved.
Description
Technical field
The invention belongs to the image analysis processing field, be applied to the automatic Detection ﹠ Controling of textile surface quality field, the present invention relates to a kind of image analysis method based on pivot analysis and be applied to the method that fabric defects detects.
Background technology
Pivot analysis (PCA) or Karhunen-Loeve (KL) conversion are as a kind of important Multivariable Statistical Methods, because its outstanding character, the area of pattern recognition that is widely used is such as recognition of face, data compression.The basic thought of pivot analysis is to obtain one group of identical and mutual incoherent new feature of number with linear transformation from original feature, and front several in these features have comprised former characteristic main information.
In art of image analysis, pivot analysis is mainly used in the analysis of multispectral image and true color image as a kind of multivariate method, and can't directly apply to the analysis of gray level image.The people such as Bharati (2000) spatially carry out gray level image producing multiple image behind the rotary manipulation of the translation of different directions and different angles, then adopt PCA the multiple image that produces to be carried out the analysis of Pixel-level, because method is the analysis of carrying out on the basis of single pixel, and image need to be carried out translation and obtain the multiple image that satisfies pivot analysis with rotation, the operand that relates to is very large.Carry out detection field at fabric defects, the people such as Ozdemir (1996) at first are divided into original image 32 * 32 nonoverlapping subwindows, the gray-scale value of the every row of subwindow is considered as random vector, then the covariance matrix of these stochastic variables is done pivot analysis and obtains eigenwert, and extract first three eigenvalue of maximum and as difference normal and flaw sample index.The proper vector that Kumar (2003) extracts 7 * 7 templates adopts PCA to carry out carrying out Defect Detection behind the dimensionality reduction.The people such as Sezer (2004) utilize PCA that the high dimensional feature vector that extracts is carried out first dimensionality reduction then to carry out Defect Detection with independent component analysis.It should be noted that, the people (2004) such as Kumar (2003) and Sezer just see PCA as a kind of householder method of dimensionality reduction, and the people such as Ozdemir (1996) are though directly adopted pivot analysis to carry out Defect Detection, but the method need to be carried out pivot analysis to each sample, and the operand that relates to is very large; Next is that the method is not considered cloth textured random disturbance, and there is larger error in testing result.
Summary of the invention
Purpose of the present invention is exactly not enough in order to overcome existing detection method, has proposed a kind of image analysis method based on pivot analysis and has been applied to the fabric defects detection method.The present invention directly is divided into a certain size subwindow with image, and take a window as unit rather than on single pixel image is carried out pivot analysis, can greatly reduce calculated amount; Gray-scale value in the subwindow is launched by the mode of row and column, can delineate better the feature of image different directions.For fabric defects, to being arranged, the overlapping resulting sample of division subwindow carries out pivot analysis, can effectively extract the inherent texture and structural characteristic of fabric; After being launched by the row, column direction respectively, the source textile sample carries out template operation, and then carry out respectively pivot analysis, not only take full advantage of cloth textured longitude and latitude orientation characteristic, and be conducive to outstanding flaw and suppress the texture random disturbance, improve Detection accuracy.
A kind of image analysis method based on pivot analysis of the present invention comprises training stage and analysis phase two parts, and concrete steps are:
Training stage
1) flawless image is had the subwindow that is divided into overlappingly the capable n row of m size, it is in order to extract better the structural information of image inherence that overlapping division subwindow is arranged; Be the matrix of the capable n of m row depending on each subwindow, the gray-scale value in the subwindow is launched into the column vector of two groups of m * n dimension by row with by the mode of row, and to look this two groups of m * n dimensional vector be two groups of random vectors, remember that to corresponding two groups of random vectors be x
hAnd x
vTo random vector x
hAnd x
vCarry out template operation with outstanding flaw and suppress random disturbance, with random vector x after the template operation
hAnd x
vCarry out pivot analysis, obtain corresponding pivot matrix and be designated as W
HAnd W
V
2) with the continuous zero lap of flawless image be divided into the subwindow of the capable n of m row size; Gray-scale value in the subwindow is launched into two groups of m * n dimensional vector by row with by the mode that is listed as, remembers to be designated as y to corresponding two groups of m * n dimensional vector
hAnd y
vTo y
hAnd y
vCarry out template operation with outstanding flaw and suppress random disturbance, the result after the corresponding template operation is taken matrix in the right side respectively
And matrix
Be about to f
hAnd f
vProject to matrix
And matrix
On, the column vector that obtains corresponding two groups of new m * n dimension is designated as y '
hAnd y '
vCalculate y
hAnd y '
h, y
vAnd y '
vBetween similarity, obtaining corresponding two similarities is S
hAnd S
v, and note S=S
h+ S
vCalculate the whenever S value of all subwindows, then calculate the cumulative distribution function F (S) of S value, make F (S ')=α, the corresponding S ' of α this moment value is used threshold value T as analyzing
αT is namely arranged
α=S ';
Wherein, α is confidence level, and from probability of making a mistake of probability theory expression, α in the present invention represents a false drop rate, is about to the probability that normal sample is mistaken for the flaw sample.Owing to be based on a certain threshold value T in actual analysis
αCarry out down, can not predict the analytical effect of reality.Therefore, in order to carry out certain prediction to the analytical effect of reality, the present invention selects the corresponding threshold value T under the different α (getting 0～0.15)
αCarry out actual test, set up the relation between α and the actual analysis effect, and then can guarantee actual analytical effect by setting α in the actual analysis afterwards.
Analysis phase
3) with the continuous zero lap of image to be analyzed be divided into the subwindow of the capable n of m row size; Choose a subwindow, and the gray-scale value in the subwindow is launched into the column vector that two groups of m * n ties up by row with by the mode that is listed as, remember to be designated as f to the column vector of corresponding two groups of m * n dimension
hAnd f
vTo being f
hAnd f
vCarry out template operation with outstanding flaw and suppress random disturbance, the result after the corresponding template operation is taken matrix in the right side respectively
And matrix
The column vector that obtains corresponding two groups of new m * n dimension is designated as f '
hAnd f '
vCalculate f
hAnd f '
h, f
vAnd f '
vBetween similarity, obtaining corresponding two similarities is K
hAnd K
v, and note K=K
h+ K
vIf K is less than threshold value T
α, think that then this sample is the flaw sample;
Wherein, the measurement index of described similarity is cosine distance, Euclidean distance or signal to noise ratio (S/N ratio), and very little for the difference between the indefectible sample is that similarity is very high, and larger with the sample difference of flaw, namely similarity is not high; Common m, choosing of n there is no concrete theoretical foundation, mainly depend on the size of flaw own and calculated amount, if m, n are too little, the calculated amount that then relates to can be very large, and if m, n are too large, then can not effectively give prominence to flaw information, cause analysis precision to reduce, the comprehensive actual conditions of the present invention and experimental exploring think that m, n get 32 * 32 the bests.
As preferred technical scheme:
Aforesaid a kind of image analysis method based on pivot analysis, described have overlappingly divide the horizontal level that refers between adjacent two the subwindow reference positions of line direction at a distance of 1～[n/2] individual pixel ([] expression rounds), the continuous non-overlapping dividing mode of adjacent two subwindows of column direction is resulting; Perhaps the upright position between adjacent two the subwindow reference positions of column direction is at a distance of 1～[m/2] individual pixel, and the continuous non-overlapping dividing mode of adjacent two subwindows of line direction is resulting.
Aforesaid a kind of image analysis method based on pivot analysis, described image are that bit depth is 8 gray level images.
It is 3 one dimension template that aforesaid a kind of image analysis method based on pivot analysis, described template operation refer to adopt length:
Do convolution with former vector.
Aforesaid a kind of image analysis method based on pivot analysis, described to random vector x after the template operation
hAnd x
vCarry out pivot analysis, obtain corresponding pivot matrix and be designated as W
HAnd W
V, refer to by finding the solution random vector x
hAnd x
vThe eigenwert of autocorrelation matrix separately and proper vector, and choose respectively front p eigenwert characteristic of correspondence vector as corresponding pivot matrix, exponent number is all the capable p row of m * n; Described pivot analysis algorithm is that the eigenwert that the proper vector of finding the solution the autocorrelation matrix of random vector obtains the pivot matrix or finds the solution the covariance matrix of random vector obtains the pivot matrix.
Aforesaid a kind of image analysis method based on pivot analysis, the measurement index of described similarity is the cosine distance.Cosine is apart from the cosine value of angle between two vectors of expression, and it is defined as:
The present invention also provides a kind of image analysis method based on pivot analysis to be applied to the method that fabric defects detects, and comprises training stage and detection-phase two parts, and concrete steps are:
Training stage
1) flawless textile image is had the subwindow that is divided into overlappingly the capable n row of m size, it is in order to extract better the structural information of image inherence that overlapping division subwindow is arranged; Be the matrix of the capable n of m row depending on each subwindow, the gray-scale value in the subwindow is launched into the column vector of two groups of m * n dimension by row with by the mode of row, and the column vector of looking this two groups of m * n dimension is two groups of random vectors, remember that to corresponding two groups of random vectors be x
hAnd x
vTo random vector x
hAnd x
vCarry out template operation with outstanding flaw and suppress random disturbance, to random vector x after the template operation
hAnd x
vCarry out pivot analysis, obtain corresponding pivot matrix and be designated as W
HAnd W
V
2) with the continuous zero lap of flawless textile image be divided into the subwindow of the capable n of m row size; Gray-scale value in the subwindow is launched into the column vector that two groups of m * n ties up by row with by the mode that is listed as, remembers to be designated as y to the column vector of corresponding two groups of m * n dimension
hAnd y
vTo being y
hAnd y
vCarry out template operation with outstanding flaw and suppress random disturbance, the result after the corresponding template operation is taken matrix in the right side respectively
And matrix
Be about to f
hAnd f
vProject to matrix
And matrix
On, the column vector that obtains corresponding two new m * n dimension is designated as y '
hAnd y '
vCalculate y
hAnd y '
h, y
vAnd y '
vBetween similarity, obtaining corresponding two similarities is S
hAnd S
v, and note S=S
h+ S
vCalculate the whenever S value of all subwindows, then calculate the cumulative distribution function F (S) of S value, make F (S ')=α, α is confidence level and gets 0～0.15, and the corresponding S ' of α this moment value is used threshold value T as analyzing
αT is namely arranged
α=S ';
Detection-phase
3) with the continuous zero lap of textile image to be detected be divided into the subwindow of the capable n of m row size; Choose a subwindow, and the gray-scale value in the subwindow is launched into the column vector that two groups of m * n ties up by row with by the mode that is listed as, remember to be designated as f to the column vector of corresponding two groups of m * n dimension
hAnd f
vTo being f
hAnd f
vCarry out template operation, the result after the corresponding template operation is taken matrix in the right side respectively
And matrix
The column vector that obtains corresponding two new m * n dimension is designated as f '
hAnd f '
vCalculate f
hAnd f '
h, f
vAnd f '
vBetween similarity, obtaining corresponding two similarities is K
hAnd K
v, and note K=K
h+ K
vIf K is less than threshold value T
α, think that then this sample is the flaw sample;
Wherein, the measurement index of described similarity is cosine distance, Euclidean distance or signal to noise ratio (S/N ratio); M * n preferred 32 * 32.
The aforesaid method that is applied to the fabric defects detection, described fabric refers to the plain color woven fabric, it is 8 gray level images that used textile image is bit depth.
Aforesaid a kind of image analysis method based on pivot analysis, described have overlappingly divide the horizontal level that refers between adjacent two the subwindow reference positions of line direction at a distance of 1～[n/2] individual pixel, the continuous non-overlapping dividing mode of adjacent two subwindows of column direction is resulting; Perhaps the upright position between adjacent two the subwindow reference positions of column direction is at a distance of 1～[m/2] individual pixel, and the continuous non-overlapping dividing mode of adjacent two subwindows of line direction is resulting.
Aforesaidly be applied to the method that fabric defects detects, it is characterized in that, it is 3 one dimension template that described template operation refers to adopt length:
Do convolution with former vector.
The aforesaid method that is applied to the fabric defects detection is characterized in that, and is described to random vector x after the template operation
hAnd x
vCarry out pivot analysis, obtain corresponding pivot matrix and be designated as W
HAnd W
V, refer to by finding the solution random vector x
hAnd x
vThe eigenwert of autocorrelation matrix separately and proper vector, and choose respectively front p eigenwert characteristic of correspondence vector as corresponding pivot matrix, exponent number is all the capable p row of m * n; Described pivot analysis algorithm is that the proper vector that the proper vector of finding the solution the autocorrelation matrix of random vector obtains the pivot matrix or finds the solution the covariance matrix of random vector obtains the pivot matrix.
The aforesaid method that is applied to the fabric defects detection is characterized in that the measurement index of described similarity is the cosine distance.
The aforesaid method that is applied to the fabric defects detection, the line direction of described textile image to be detected and column direction are corresponding to weft yarn and the warp thread direction of fabric, or corresponding to warp thread and weft direction, purpose is to give prominence to better warp thread and weft direction orientation characteristic.
Beneficial effect
1, method itself has a negative function to illumination is irregular, does not need traditional image pre-treatment step;
2, algorithm is simple in the calculating of detection-phase;
3, after the source textile sample is launched by the row, column direction respectively, carry out template operation, can not only take full advantage of cloth textured longitude and latitude orientation characteristic, and be conducive to outstanding flaw and suppress the texture random disturbance, improve Detection accuracy.
Description of drawings
Fig. 1 is the division synoptic diagram of subwindow of the present invention on textile image
Fig. 2 is that the present invention has overlapping division subwindow synoptic diagram
Fig. 3 is that the present invention is by going by row expansion mode synoptic diagram
Fig. 4 is that the continuous zero lap of the present invention is divided the subwindow synoptic diagram
Fig. 5 is data set 1 contained flaw image in the embodiment of the invention
Fig. 6 is data set 2 contained flaw images in the embodiment of the invention
Fig. 7 is data set 3 contained flaw images in the embodiment of the invention
Fig. 8 is data set 4 contained flaw images in the embodiment of the invention
Fig. 9 is data set 5 contained flaw images in the embodiment of the invention
Figure 10 is data set 6 contained flaw images in the embodiment of the invention
Figure 11 is the test findings to data set among the embodiment 1
Figure 12 is the test findings to data set among the embodiment 2
Figure 13 is the test findings to data set among the embodiment 3
Figure 14 is the test findings to data set among the embodiment 4
Figure 15 is the test findings to data set among the embodiment 5
Figure 16 is the test findings to data set among the embodiment 6
Embodiment
Below in conjunction with embodiment, further set forth the present invention.Should be understood that these embodiment only to be used for explanation the present invention and be not used in and limit the scope of the invention.Should be understood that in addition those skilled in the art can make various changes or modifications the present invention after the content of having read the present invention's instruction, these equivalent form of values fall within the application's appended claims limited range equally.
The embodiment of the invention adopts six kinds of plain color woven fabric images with different texture background as trial image, these trial image and on fault all from production practices, Fig. 5～image that is respectively six kinds of contained faults of trial image shown in Figure 10.
Concrete performing step is:
Training stage
1) as depicted in figs. 1 and 2, with flawless image with the horizontal level between adjacent two the subwindow reference positions of line direction at a distance of 1 pixel, be divided into the big or small subwindow of 32 row, 32 row adjacent two the continuous zero laps of subwindow of column direction; Be the matrix of 32 row 32 row depending on each subwindow, as shown in Figure 3, gray-scale value in the subwindow is launched into the column vector of two group of 1024 dimension by row with by the mode of row, and to look this two group of 1024 dimensional vector be two groups of random vectors, remembers that to corresponding two groups of random vectors be x
hAnd x
vUsing length is that 3 one dimension template M is to random vector x
hAnd x
vCarry out template operation, find the solution respectively random vector x after the template operation
hAnd x
vAutocorrelation matrix, and get respectively front 10 eigenwert characteristics of correspondence vector as separately pivot matrix, be designated as W
HAnd W
V, exponent number is all 1024 row, 10 row;
2) as shown in Figure 1 and Figure 4, with the continuous zero lap of flawless textile image be divided into the subwindow of 32 row 32 row size; As shown in Figure 3, the gray-scale value in the subwindow is launched into two group of 1024 column vector of tieing up by row with by the mode that is listed as, remembers to be designated as y to the column vector of corresponding two group of 1024 dimension
hAnd y
vTo being y
hAnd y
vCarry out template operation, the result after the corresponding template operation is taken matrix in the right side respectively
And matrix
The column vector that obtains corresponding two 1024 new dimensions is designated as y '
hAnd y '
vCalculate y
hAnd y '
h, y
vAnd y '
vBetween the cosine distance, obtain corresponding two cosine distance and be S
hAnd S
v, and note S=S
h+ S
vCalculate the whenever S value of all subwindows, then calculate the cumulative distribution function F (S) of S value, make F (S ')=α, α is confidence level, and the corresponding S ' of α this moment value is used threshold value T as analyzing
αT is namely arranged
α=S '; The corresponding threshold value T of value (step-length is 0.01) between calculating α gets 0～0.15
0～0.15
Detection-phase
3) as shown in Figure 1 and Figure 4, with the continuous zero lap of textile image to be detected be divided into the subwindow of 32 row 32 row size; Choose a subwindow, as shown in Figure 3, and with the gray-scale value in the subwindow by the row and by row mode be launched into two group 1024 the dimension column vector, remember to corresponding two group 1024 the dimension column vector be designated as f
hAnd f
vTo being f
hAnd f
vCarry out template operation, with the right multiply matrix of result's difference after the corresponding template operation
And matrix
The column vector that obtains corresponding two 1024 new dimensions is designated as f '
hAnd f '
vCalculate f
hAnd f '
h, f
vAnd f '
vBetween the cosine distance, obtain corresponding two cosine distance and be K
hAnd K
v, and note K=K
h+ K
vIf K is less than threshold value T
α, think that then this sample is the flaw sample; With the corresponding threshold value T of all subwindow sample to be measured values between α gets 0～0.15
0～0.15Under test, obtain false drop rate (α %) and loss curve, such as Figure 11～Figure 16.Wherein, false drop rate is that normal sample is declared the number of flaw sample and the ratio of whole samples; Loss is that the flaw sample is judged to the number of normal sample and the ratio of whole samples.
Embodiment
Six used data set samples of embodiment are allocated as follows shown in the table:
Wherein, training set A
HAnd A
VRespectively by step 1) random vector x
hAnd x
vConsist of, sample number unification of the present invention gets 600, is used for calculating corresponding pivot matrix; Training set B
HAnd B
VRespectively by step 2) y
hAnd y
vConsist of, be used for to calculate detect used threshold value T
αTest set D
HAnd D
VRespectively by step 3) in the f that the flaw sample is arranged
hAnd f
vConsist of, be used for to obtain loss.
Figure 11～Figure 16 has provided the actual testing result of implementing used six data sets, wherein the represented false drop rate of horizontal ordinate in each example is by confidence level α direct estimation gained, and the represented loss of ordinate then is at the threshold value T that gets under the different confidence level α
αTest gained (α gets 0～0.15).
Claims (10)
1. the image analysis method based on pivot analysis is characterized in that comprising training stage and analysis phase two parts, and concrete steps are:
Training stage
1) flawless image there is the subwindow that is divided into overlappingly the capable n row of m size; Gray-scale value in the subwindow is launched into the column vector of two groups of m * n dimension by row with by the mode of row, and to look this two groups of m * n dimensional vector be two groups of random vectors, remember that to corresponding two groups of random vectors be x
hAnd x
vTo random vector x
hAnd x
vCarry out template operation, with random vector x after the template operation
hAnd x
vCarry out pivot analysis, obtain corresponding pivot matrix and be designated as W
HAnd W
V
2) with the continuous zero lap of flawless image be divided into the subwindow of the capable n of m row size; Gray-scale value in the subwindow is launched into two groups of m * n dimensional vector by row with by the mode that is listed as, remembers to be designated as y to corresponding two groups of m * n dimensional vector
hAnd y
vTo y
hAnd y
vCarry out template operation, with the right multiply matrix of result's difference after the corresponding template operation
And matrix
The column vector that obtains corresponding two groups of new m * n dimension is designated as y '
hAnd y '
vCalculate y
hAnd y '
h, y
vAnd y '
vBetween similarity, obtaining corresponding two similarities is S
hAnd S
v, and note S=S
h+ S
vCalculate the S value of all subwindows, then calculate the cumulative distribution function F (S) of S value, make F (S ')=α, α is confidence level, and the corresponding S ' of α this moment value is used threshold value T as analyzing
αT is namely arranged
α=S ';
Analysis phase
3) with the continuous zero lap of image to be analyzed be divided into the subwindow of the capable n of m row size; Choose a subwindow, and the gray-scale value in the subwindow is launched into the column vector that two groups of m * n ties up by row with by the mode that is listed as, remember to be designated as f to the column vector of corresponding two groups of m * n dimension
hAnd f
vTo being f
hAnd f
vCarry out template operation, the result after the corresponding template operation is taken matrix in the right side respectively
And matrix
The column vector that obtains corresponding two groups of new m * n dimension is designated as f '
hAnd f '
v
Calculate f
hAnd f '
h, f
vAnd f '
vBetween similarity, obtaining corresponding two similarities is K
hAnd K
v, and note K=K
h+ K
v
If K is less than threshold value T
α, think that then this sample is the flaw sample;
Wherein, the measurement index of described similarity is cosine distance, Euclidean distance or signal to noise ratio (S/N ratio); M * n is 32 * 32.
2. a kind of image analysis method based on pivot analysis according to claim 1, it is characterized in that, described have overlappingly divide the horizontal level that refers between adjacent two the subwindow reference positions of line direction at a distance of 1～[n/2] individual pixel, the continuous non-overlapping dividing mode of adjacent two subwindows of column direction is resulting; Perhaps the upright position between adjacent two the subwindow reference positions of column direction is at a distance of 1～[m/2] individual pixel, and the continuous non-overlapping dividing mode of adjacent two subwindows of line direction is resulting.
3. a kind of image analysis method based on pivot analysis according to claim 1 is characterized in that, described image is that bit depth is 8 gray level images.
4. a kind of image analysis method based on pivot analysis according to claim 1 is characterized in that, it is 3 one dimension template that described template operation refers to adopt length:
Do convolution with former vector.
5. a kind of image analysis method based on pivot analysis according to claim 1 is characterized in that, and is described to random vector x after the template operation
hAnd x
vCarry out pivot analysis, obtain corresponding pivot matrix and be designated as W
HAnd W
V, refer to by finding the solution random vector x
hAnd x
vThe eigenwert of autocorrelation matrix separately and proper vector, and choose respectively front 10 eigenwert characteristics of correspondence vector as corresponding pivot matrix, exponent number is all capable 10 row of m * n; Described pivot matrix is that the proper vector that the proper vector of finding the solution the autocorrelation matrix of random vector obtains the pivot matrix or finds the solution the covariance matrix of random vector obtains the pivot matrix.
6. a kind of image analysis method based on pivot analysis according to claim 1 is characterized in that, the measurement index of described similarity is the cosine distance.
7. a kind of image analysis method based on pivot analysis as claimed in claim 1 is applied to the method that fabric defects detects, and it is characterized in that comprising training stage and detection-phase two parts, and concrete steps are:
Training stage
1) flawless textile image there is the subwindow that is divided into overlappingly the capable n row of m size; Gray-scale value in the subwindow is launched into the column vector of two groups of m * n dimension by row with by the mode of row, and the column vector of looking this two groups of m * n dimension is two groups of random vectors, remembers that to corresponding two groups of random vectors be x
hAnd x
vTo random vector x
hAnd x
vCarry out template operation, to random vector x after the template operation
hAnd x
vCarry out pivot analysis, obtain corresponding pivot matrix and be designated as W
HAnd W
V
2) with the continuous zero lap of flawless textile image be divided into the subwindow of the capable n of m row size; Gray-scale value in the subwindow is launched into the column vector that two groups of m * n ties up by row with by the mode that is listed as, remembers to be designated as y to the column vector of corresponding two groups of m * n dimension
hAnd y
vTo being y
hAnd y
vCarry out template operation, the result after the corresponding template operation is taken matrix in the right side respectively
And matrix
The column vector that obtains corresponding two new m * n dimension is designated as y '
hAnd y '
vCalculate y
hAnd y '
h, y
vAnd y '
vBetween similarity, obtaining corresponding two similarities is S
hAnd S
v, and note S=S
h+ S
vCalculate the S value of all subwindows, then calculate the cumulative distribution function F (S) of S value, make F (S ')=α, α is confidence level, and the corresponding S ' of α this moment value is used threshold value T as analyzing
αT is namely arranged
α=S ';
Detection-phase
3) with the continuous zero lap of textile image to be detected be divided into the subwindow of the capable n of m row size; Choose a subwindow, and the gray-scale value in the subwindow is launched into the column vector that two groups of m * n ties up by row with by the mode that is listed as, remember to be designated as f to the column vector of corresponding two groups of m * n dimension
hAnd f
vTo being f
hAnd f
vCarry out template operation, with the right multiply matrix of result's difference after the corresponding template operation
And matrix
The column vector that obtains corresponding two new m * n dimension is designated as f '
hAnd f '
vCalculate f
hAnd f '
h, f
vAnd f '
vBetween similarity, obtaining corresponding two similarities is K
hAnd K
v, and note K=K
h+ K
vIf K is less than threshold value T
α, think that then this sample is the flaw sample;
Wherein, the measurement index of described similarity is cosine distance, Euclidean distance or signal to noise ratio (S/N ratio); M * n is 32 * 32.
8. according to claim 7ly be applied to the method that fabric defects detects, it is characterized in that, described have overlappingly divide the horizontal level that refers between adjacent two the subwindow reference positions of line direction at a distance of 1～[n/2] individual pixel, the continuous non-overlapping dividing mode of adjacent two subwindows of column direction is resulting; Perhaps the upright position between adjacent two the subwindow reference positions of column direction is at a distance of 1～[m/2] individual pixel, and the continuous non-overlapping dividing mode of adjacent two subwindows of line direction is resulting; It is 3 one dimension template that described template operation refers to adopt length:
Do convolution with former vector; The measurement index of described similarity is the cosine distance.
9. the method that is applied to the fabric defects detection according to claim 7 is characterized in that, and is described to random vector x after the template operation
hAnd x
vCarry out pivot analysis, obtain corresponding pivot matrix and be designated as W
HAnd W
V, refer to by finding the solution random vector x
hAnd x
vThe eigenwert of autocorrelation matrix separately and proper vector, and choose respectively front 10 eigenwert characteristics of correspondence vector as corresponding pivot matrix, exponent number is all capable 10 row of m * n; Described pivot matrix is that the proper vector that the proper vector of finding the solution the autocorrelation matrix of random vector obtains the pivot matrix or finds the solution the covariance matrix of random vector obtains the pivot matrix.
10. according to claim 7ly be applied to the method that fabric defects detects, it is characterized in that, the line direction of described textile image to be detected and column direction are corresponding to weft yarn and the warp thread direction of fabric, or corresponding to warp thread and weft direction, purpose is to give prominence to better warp thread and weft direction orientation characteristic.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN 201110219310 CN102289677B (en) | 2011-08-02 | 2011-08-02 | Method for analyzing image based on principal component analysis and method applicable to detection of defects of fabric |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN 201110219310 CN102289677B (en) | 2011-08-02 | 2011-08-02 | Method for analyzing image based on principal component analysis and method applicable to detection of defects of fabric |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN102289677A CN102289677A (en) | 2011-12-21 |
| CN102289677B true CN102289677B (en) | 2013-05-29 |
Family
ID=45336083
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN 201110219310 Active CN102289677B (en) | 2011-08-02 | 2011-08-02 | Method for analyzing image based on principal component analysis and method applicable to detection of defects of fabric |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102289677B (en) |
Families Citing this family (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103473769B (en) * | 2013-09-05 | 2016-01-06 | 东华大学 | A kind of fabric defects detection method based on svd |
| CN103473770B (en) * | 2013-09-05 | 2016-06-15 | 东华大学 | A kind of linear textiles flaw detection method based on rectangular window projection and singular value decomposition |
| FR3010523B1 (en) * | 2013-09-06 | 2015-09-25 | Safran | PROCESS FOR CHARACTERIZING A PIECE OF COMPOSITE MATERIAL |
| CN103593670B (en) * | 2013-10-14 | 2016-07-06 | 浙江工业大学 | A kind of copper plate/strip detection method of surface flaw based on online limit of sequence learning machine |
| CN107967680B (en) * | 2017-11-24 | 2019-07-09 | 常州大学 | Textile flaw detection method based on peak threshold and composite character |
| CN114841982B (en) * | 2022-05-20 | 2024-01-09 | 浙江纵驰环境科技有限公司 | Air conditioner filter element quality monitoring method based on image processing |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101216435A (en) * | 2008-01-03 | 2008-07-09 | 东华大学 | Fabric flaw automatic detection method based on multi-fractal characteristic parameter |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2006119076A (en) * | 2004-10-25 | 2006-05-11 | Jasco Corp | Apparatus and method for analyzing mapping data |
-
2011
- 2011-08-02 CN CN 201110219310 patent/CN102289677B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101216435A (en) * | 2008-01-03 | 2008-07-09 | 东华大学 | Fabric flaw automatic detection method based on multi-fractal characteristic parameter |
Non-Patent Citations (2)
| Title |
|---|
| 基于时间序列分形特征的织物瑕疵检测;田承泰等;《纺织学报》;20100531;第31卷(第5期);第44~47页 * |
| 田承泰等.基于时间序列分形特征的织物瑕疵检测.《纺织学报》.2010,第31卷(第5期),第44~47页. |
Also Published As
| Publication number | Publication date |
|---|---|
| CN102289677A (en) | 2011-12-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN102289677B (en) | Method for analyzing image based on principal component analysis and method applicable to detection of defects of fabric | |
| Zhu et al. | Yarn-dyed fabric defect detection based on autocorrelation function and GLCM | |
| CN103604809B (en) | A kind of online visible detection method of pattern cloth flaw | |
| CN103886325B (en) | Cyclic matrix video tracking method with partition | |
| CN103234976B (en) | Based on the online visible detection method of tricot machine Fabric Defect of Gabor transformation | |
| CN108090894B (en) | Fabric defect detection method based on distance matching function and perceptual hash algorithm | |
| CN108960255A (en) | Conspicuousness fabric defect detection method based on color similarity and position aggregation | |
| CN102867299B (en) | Image analysis method based on singular value decomposition and method applied to defect detection of fabric | |
| CN112465752A (en) | Improved Faster R-CNN-based small target detection method | |
| CN104318559A (en) | Quick feature point detecting method for video image matching | |
| CN107240086A (en) | A kind of fabric defects detection method based on integration nomography | |
| CN112488211A (en) | Fabric image flaw classification method | |
| CN109949287A (en) | A kind of fabric defects detection method based on adaptivenon-uniform sampling and template correction | |
| CN109002792B (en) | SAR image change detection method based on layered multi-model metric learning | |
| CN113674263A (en) | Small sample defect detection method based on generation type countermeasure network | |
| Li et al. | Small object detection algorithm based on feature pyramid-enhanced fusion SSD | |
| An et al. | Fabric defect detection using deep learning: An Improved Faster R-approach | |
| Liang et al. | Fast hyperspectral image classification combining transformers and SimAM-based CNNs | |
| CN102968786B (en) | A kind of non-supervisory remote sensing images potential target method for detecting area | |
| CN102005049A (en) | Unilateral generalized gaussian model-based threshold method for SAR (Source Address Register) image change detection | |
| CN103455798B (en) | Histogrammic human body detecting method is flowed to based on maximum geometry | |
| CN110516527B (en) | Visual SLAM loop detection improvement method based on instance segmentation | |
| CN103065296B (en) | High-resolution remote sensing image residential area extraction method based on edge feature | |
| Gao et al. | A novel patterned fabric defect detection algorithm based on GHOG and low-rank recovery | |
| CN110503631A (en) | A kind of method for detecting change of remote sensing image |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant |