Image Defect Recognition Based on âSuper Fuzzy ... - CiteSeerX
JOURNAL OF MULTIMEDIA, VOL. 5, NO. 2, APRIL 2010
181
Image Defect Recognition Based on “Super Fuzzy” Characteristic Zhe LIU and Xiuchen WANG Zhongyuan University of Technology, Zhengzhou, China Email: [email protected], [email protected]
Abstract—In this paper, we propose a new defects recognition algorithm for dynamic image based on “super fuzzy” feature. With this algorithm, the image is divided into some variable windows, and the eigenvector of each window is constructed. We introduce “super fuzzy” vector to make window vectors “super fuzzy” processing, thus the window feature has “super fuzzy” characteristic with the difference of the primary and secondary. Also we present window coefficient to adjust recognition speed and accuracy according to different images. Furthermore, objective function, membership function and clustering center calculation function of fuzzy clustering algorithm with window coefficient and “super fuzzy” vector are proposed in this paper. At last, we take example for fabric defects detection with this algorithm, list recognition results, discuss recognition result influence by “super fuzzy” feature and size change of window, and make some comparison with other algorithms. The conclusion shows that this algorithm can recognize more categories of image abnormal regions with high-accuracy, high-speed, no-training and extensive application. Index Terms—super fuzzy feature, variable window, fuzzy, clustering, recognition, defect region
I.
INTRODUCTION
Image defect region recognition is an important research issue of computer pattern recognition with wide application prospect. The aim is to recognize automatically abnormal region with rule mutation according to whole characteristic rule of image. This research result can apply to computer-aided product detection, quality evaluation and other fields in every industry, and has key meaning to improve quality detection automation in enterprises. Many scholars have done some research about it. Now, there are two main methods. The first approach is space domain algorithms. In these algorithms, the image’s gray value is extracted for analysis and recognition. Main algorithms are geometric method [1-4], co-occurrence matrix [5-6], neural network method [7-9], probability statistic [10-11], and so on. In geometric method, the basis is unstable because of variable defects shape, the defect recognition accuracy is difficult to reach, and involves in massive calculation. In co-occurrence matrix,
© 2010 ACADEMY PUBLISHER doi:10.4304/jmm.5.2.181-188
image’s quality requirement is very high. In order to construct multi-dimensional feature matrix and analyze its characteristic, its calculation is very large too, so the recognition speed and accuracy are not idea. The neural network approach once was research hotspot, but its calculation is large and recognition needs pre-training. When to recognize static fabric sometimes need more time, and can not adapt to dynamic fabric detection. In probability statistic, because the probability of all gray value must be calculated and compared, and the probability statistic is more depend on the image quality, the problem of recognition speed and accuracy is existed too. The second approach is frequency domain algorithms. In these algorithms, image is transformed to certain spectrum for research and analysis. At first, Fourier transform is main algorithm [12-14], But Fourier transform algorithm is so difficult to recognize dynamic window’s characteristic that it couldn’t apply to practice application of dynamic detection. Later improved Fourier transform is appear, the more representative approach is Gabor algorithm [15-17]. Though Gabor algorithm has the computing capacity for dynamic window, the problem of large computing is also existed. Wavelet algorithm [18-24] was research hotspot for a time, and obtained some achievements on individual defects detection. But some defects are distinct appear after image transforming to spectrum in frequency domain algorithms, this transform is quite insensitive to many defects. That is, the result after transform could not further recognize defects. Moreover, spectrum transforms involves complex transform function. It increases calculation load and influences the recognition speed. More importantly, even though the image is made spectrum transform, the secondary recognition for spectrum characteristic has to face, which will further influence the recognition speed and accuracy. Aiming at these instances, this paper proposes a novel recognition algorithm with fuzzy clustering based on “super fuzzy” characteristic. This recognition algorithm can recognize defect regions of dynamic image. The image is divided into some windows and the window coefficient is introduced for describing image rule. The characteristic vector of each window is constructed and the “super fuzzy” vector is introduced so that each window characteristic has primary and secondary
182
JOURNAL OF MULTIMEDIA, VOL. 5, NO. 2, APRIL 2010
difference. On this basis, some matching objective functions and membership functions of fuzzy clustering are proposed in this paper. This algorithm is tested by recognizing fabric defects, result shows that this algorithm can recognize many kinds of image’s defect regions with high-accuracy, high-speed, extensive application and no re-training.
rule. If pixel points are carried through clustering one by one, it will arouse clustering confusion. Recognition accuracy influence by change of region window size is also great, so window factor must be considered during recognition process. To sum up, we propose a pattern recognition algorithm with fuzzy clustering based on “super fuzzy” characteristic and variable window. The thought is shown in Figure1.
II. ALGORITHM THOUGHT Image’s defect region is produced owing to rule characteristic change of this region. In daily life, one person with no professional knowledge even a sensible child can distinguish defect region of random images quickly and easily by naked eye. After studying we find that recognition process by human (is very simple). This process contains three periods. Firstly, people looks through an image wholly and search the main rule of image quickly. Secondly, the image is segmented and its rule is found one by one. Finally, abnormal region is found abiding by main image’s rule. This process is a fuzzy judgment all the way, but it always has a very accurate result. This paper tries to study a new pattern recognition algorithm, so that computer can recognize defect region quickly and accurately based on fuzzy judgment abiding by above human recognition process. We think that any images have many rule characteristics, and these characteristics have difference with primary and secondary. For example, when we recognize fabric defects, the characteristic of silk is obvious luster, so abnormal region of defects represents abnormal luster firstly. Luster of cotton fabric changes slow, but its structure texture rule change can make obvious change of fabric appearance, so abnormal region of defects represents difference of fabric structure texture firstly. Abnormal region of defects in wool fabric represents the difference of fabric partial roughness. Therefore, we introduce a concept of “super fuzzy” characteristic, window can describe with main features instead of whole feature. the calculation load and time is greatly keep down. The theory of “super fuzzy” characteristic can explain vividly with an example. If we want to distinguish five people from different provinces, and every people has filled in a table with province, city, section, street and unit. Do we recognize them one by one according to their detail information or according to their fuzzy province? This problem is very simple, we only make “super fuzzy”, and not consider latter detail. Only the “super fuzzy” feature—province, we can distinguish the source region of each person without any error. Furthermore, if three persons come from the same province, we can use the second “super fuzzy” feature—city to compare these three persons. The result is also clear. This process, which is a simplify process of complex recognition problem and greatly shorten recognition time, is happen in many cases in daily life. In addition, image recognition is recognized abiding by regions one by one. Only certain region can present certain rule, single pixel point can not show the image’s
© 2010 ACADEMY PUBLISHER
Fig.1 Algorithm thought
III. RECOGNITION ALGORITHM BASED ON “SUPER FUZZY” CHARACTERISTIC
A. Multi-characteristic Model of Window If the image f ( x, y ) is divided into N × M number variable grids according to lateral interval ∆x and longitudinal interval ∆y , Τij denotes eigenvector in the i row and the j array region, which has n factors, respectively denotes n characteristic parameters of local region with λ1 , λ2 ,…, λn . If the degree of membership of each region is uij ，the image fuzzy set is expressed as: N M
Μ=
uij
UU T i =1 j =1
（1）
ij
Considering multi-characteristic of region，the formula (1) is decomposed to fuzzy set with multi-characteristic. The fuzzy set is: N M
n
Μ=
∑ UU T .λ ) (
k =1 i =1 j =1
Where
uijk
uijk
（2）
ij k
denotes the degree of membership of
the k characteristic in the i row and the Tij .λk denotes the k characteristic the
j array i
region， row and the
j array region. Thus the calculating formula of the degree of membership is given as:
uijk = 1 +
Tij .λk u
（3）
Where u denotes the maximum value of the characteristic. Figure 2 shows a multi-characteristic model diagram of an image. Here only six characteristics are selected for analysis, they are denoted by λ1 , λ2 , λ3 , λ4 , λ5 , λ6 . The k
JOURNAL OF MULTIMEDIA, VOL. 5, NO. 2, APRIL 2010
color area of each characteristic represents the degree of membership of this characteristic.
183
This paper proposes a novel fuzzy clustering algorithm. The “super fuzzy” vector is introduced before constructing objective function. The characteristic influence on clustering is changed according to this vector. Also the window coefficient is introduced, and calculating speed is determined according the change of image actual need and recognition accuracy. In the universes X , the fuzzy set Tij is Tij
={ λ1, λ2 ,..., λn } ,
vector E is E = ( m1 ， m2 ，…， n
Fig2. Schematic diagram of multi-characteristic
mn
)
⊂ [0,1]
prescriptive B. “Super Fuzzy” Processing of Multi-characteristic Fuzzy clustering algorithm belongs to unsupervised learning in pattern recognition. It need not pre-training; can directly achieve the goal of automatic classification by machine learning. The most successful and extensive application is the Fuzzy C-Mean algorithm(FCM), which was proposed by Dunn in 1974 and extended by Bezdek. But this algorithm has many obvious disadvantages, such as it only fits to find clusters with globosity shape, this algorithm is more sensitive to voice data, algorithm couldn’t sure to converge the minimum point of objective function, and so on. Later many researchers improved the FCM and presented some new fuzzy clustering algorithms. Ozdemir[25], Menard[26], and Kaymak et al.[27] increased the euclidean metric stability under noise environment by improving the metric form of fuzzy clustering algorithm, and reduced the algorithm’s initial value and shape size’s sensitivity of cluster. Schneider [28], Liu [29], Zhang et al. [30]changed the membership function to fit actual different influence of each data point. Ichihashi [31], Yasuda [32], and Wei et al.[33] presented a fuzzy clustering algorithm under maximum entropy theory by introducing the entropy. Ma [34] and Wang et al. [35] did some research of optimizing the fuzzy clustering algorithm and making it more effective. However, in existing fuzzy clustering algorithms, general and regional feature rule of image is omitted. The cluster analysis is only by local pixel point’s features not by rule region’s features with the distinction between the primary and the secondary, so some points in defect regions and some points in normal regions are both clustered into the same cluster. At the same time, these algorithms ignore the clustering influence by the rule region’s size, thus the clustering analysis is done in a very thick data set. These two disadvantages make existing algorithms not only waste the calculating time but also influence the recognition accuracy and application scope. In those traditional fuzzy clustering algorithms, an image is considered as n number characteristic vectors as Tij (i = 1,2..., N , j = 1,2,..., M ) , and divides into c number clusters. Clustering center of each cluster is obtained， and objective function achieves minimum. This algorithm is carried though only under the condition of characteristic averaging ， and the image is not make windows segmentation.
© 2010 ACADEMY PUBLISHER
∑m
, and
s =1
s =1
vector
Tij
. The scalar product of and
E
Tij • E
is
{ m1λ1, m2λ2 ,..., mnλn } , the whole characteristic λ
ij =
=
n
∑ (m λ ) s s
2
.
s =1
These factors m1 , m2 ,…, mn are called “super fuzzy” factors corresponding each fuzzy sets, and E is called “super fuzzy” vector. The aim to introduce “super fuzzy” vector is to adjust the proportion of each characteristic in every window according to different image characteristic, and these characteristics have the difference of the primary and secondary. The most important characteristic influencing image quality is regarded as main characteristic according to different situation, so that continued clustering analysis will more accurate and easy. As shown in Figure3, a, b and c respectively shows new characteristic composition of six characteristics shown in Figure2 after fuzzy processing with different “super fuzzy” vector. The color area still denotes its degree of membership. Figure 3-a shows the result of “super fuzzy” feature, in which λ1 is the main characteristic, λ3 、 λ5 are the secondary features. Figure 3-b shows “super fuzzy” feature, in which λ1 is the main characteristic, and other characteristics are not considered. Figure 3-c shows “super fuzzy” feature, in which λ1 is the main characteristic, λ5 is the secondary feature.
Fig3. Result of “super fuzzy” characteristic
C.
Variable Window Coefficient Introducing
The image f ( x, y) is divided into N × M number variable grids according to lateral interval ∆x and longitudinal interval
∆y
. If ε
=
M +N M ×N
, here ε
is the window
coefficient. The more particular the image is divided, the little the ε value is, the higher the precision of clustering is. In general reason, the change of the number N and M is not acute for the same image, the value ε will not arouse an acute fluctuant result. Figure 4 shows the clustering result after different window segmentation. Obviously, the more particular the
184
JOURNAL OF MULTIMEDIA, VOL. 5, NO. 2, APRIL 2010
window is segmented, the higher the precision of defect region is found. The more rough the window is segmented, such as window segmentation with “5×7”, the more imprecise clustering result is. It must be point out that some different pixel points, which belong to image’s defect region and normal region, also appear some local feature consistent or approximate. If the window is divided too particular, some pixel points in defect region and normal region will be consider as in a same class. Therefore, the window segmentation must be subject to image’s local rule, not too particular.
In order that clustering center calculation is more near main characteristic, the “super fuzzy” vector is introduced as: M ×N
∑
vi = k =1
1
uikε ( E • xk ) (i = 1,2,..., c)
1
N
（6）
∑u ε
ik
k =1
uik =
1 c
⎛d ⎜ ik ⎜ d jk i =1 ⎝
⎞ ⎟ ⎟ ⎠
∑
(i = 1,2,..., c; k = 1,2,..., n)
ε
（7）
IV. EXAMPLE DEMONSTRATION
Fig4. Clustering influence by window segmentation
D.
Clustering Algorithm
The clustering process we have used is shown in Figure5.
Computer recognition for fabric defect is representation of image’s defect recognition. Up to now, there is no mature method to solve this problem. We develop a minitype computer automatic recognition system for fabric defect using above fuzzy clustering algorithm based on “super fuzzy” characteristic. The transmitting speed of this system is 60 meter/minute and the width is 1.1 meter. Dynamic fabric image is transmitted to the computer using a high definition camera， the recognition software is programmed with matlab7.0. In order to validate this fuzzy clustering algorithm, a program is programmed especially to output the middle result of image recognition, and defect regions are marked by red dot. A. Fabric Image’s Multi- Characteristic Construction
Fig5. Clustering process
Considering the multi-characteristic and window coefficient, objective function is constructed as: M ×N c
Om (u , v) =
∑∑
1
uik ε d ( xk , vt )
（4）
Six characteristics showing multi-characteristic of fabric image’s rule are constructed. λ1 denotes gray mean value of pixel points in each window, λ2 denotes fluctuate degree of each window’s inner gray, λ3 and λ4 denote texture roughness degree of each region, λ5 and λ6 denotes difference between the maximum local gray value and minimum local gray value in the region, they show the uniformity of texture. If window’s pixel size of the i row and the j array are N ij × M ij , Gkl denotes gray value of each pixel point, G denotes gray mean value, they can be expressed as following:
k =1 i =1 c
where，
∑u
ik = 1 ， uik ∈ (0,1)
i =1
Here “super fuzzy” vector is introduced and variation in cluster is obtained: 2 d ( xk , vi ) = E • xk − E • vi （5） When variation in cluster is calculated, formula（5） shows that the window characteristic has the difference of the primary and secondary considering “super fuzzy” vector. Fuzzy clustering can be made according to different image characteristic, and calculating time can be saved and recognition accuracy can be improved.
N ij M ij
∑∑ G
（8）
M ij × N ij N ij M ij
∑∑ (G
kl − G )
k =1 l =1
λ2 =
M ij N ij
2
（9）
M ij
λ3 = −
∑v
p ln v p
（10）
p =1 N ij
λ4 = −
∑w p =1
© 2010 ACADEMY PUBLISHER
kl
λ1 = k =1 l =1
p ln w p
（11）
JOURNAL OF MULTIMEDIA, VOL. 5, NO. 2, APRIL 2010
185
characteristics. The “super fuzzy” vector of middle two
where： N ij
∑
vp =
H kl
k =1 N ij M ij
∑∑ H
， wp = kl
k =1 l =1
λ5 =
l =1
k (l −
∆x ) 2
, Gkl , G
k (l +
∆x ) 2
λ6 =
k =1
(k −
∆y )l 2
∑∑ H
. Here only has 4 factors, other
features λ5 、 λ6 are deleted. The “super fuzzy” vector of the forth graph is similar to Graph 6.
l =1 N ij M ij
) − min(G
k (l −
M ij − 1
N ij
∑ (max(G
∑
H kl
1 1 1 1 , , , ) 4 4 4 4
kl
k =1 l =1
M ij
∑ (max(G
graphs is: E = (
M ij
, Gkl , G
(k +
∆y )l 2
) − min(G
N ij − 1
(k −
∆x ) 2
∆y )l 2
, Gkl , G
, Gkl , G
k (l +
(k +
∆x ) 2
∆y )l 2
))
（12）
))
（13）
B. General Result
Fig.7 Recognition Result of changing feature number in “super fuzzy”
Here only image’s part with defect is intercepted and result comparison is displayed due to the limited space. As shown in Figure 6, Each intercepted image’s size is 8cm×8cm, and 20×15 number windows are segmented. The graphs which name with “-1” are the former graphs, the graphs which name with “-2” are recognition graph of middle result. the graphs which name with “-3” are the position region graph by the system orientation after clustering calculation. If we want to continue to extract defect shape, only make analysis in defect regions and extract the defect silhouette. In these results, the “super fuzzy” vector is: E = ( 1 , 1 , 1 , 1 , 1 , 1 ) . 6 6 6 6 6 6
vector
When the number of characteristic is unchanged and “super fuzzy” vector is changed, fuzzy clustering recognition results are shown in Figure 8. The “super 13 13 1 ,0, , ,0,0) ，the 15 30 30 1 1 is: E = (0,0,0,0, , ) , the 2 2
fuzzy” vector in Graph 8-a is: E = ( “super fuzzy” vector in Graph 8-b
“super fuzzy” vector in Graph 8-c is same to Figure 6. From these results, we find that recognition result shown in graph 8-b can obtain the same recognition result as shown in graph 8-c only considering features λ5 , λ6 . In this way, we can simple recognition calculating work greatly. But recognition result shown in Graph8-a is different to Graph 8-c due to the difference of “super fuzzy” vector. All these show that the main characteristics of Graph 8-sample are λ5 and λ6 .
Fig.6. Fabric Defect Recognition Result Using Clustering Algorithm on “super fuzzy” feature
C.
Recognition Result Influence by “Super Fuzzy” Fig.8 Recognition Result of Different “super fuzzy” features
Feature Change rule of fabric image with different structure is different, so that characteristic describing each fabric’s region is different. Commonly, the number of characteristic is many for thicker fabric, and be contrary for thinner fabric. The more characteristic number is selected, the higher the fabric recognition accuracy is, and the slower the recognition speed is. The difference of selected “super fuzzy” vector will reduce the number of image’s characteristic. That is the secondary feature of image is not considered, and only the main feature is considered. Therefore, fabric recognition result will be influenced. As is shown in Figure 7, graphs are the fuzzy clustering recognition result by selecting 4 different
© 2010 ACADEMY PUBLISHER
D.
Influence of Variable Window
The size of window also influences the recognition result. Commonly, the bigger the size of window is, the bigger the recognition result error is, whereas the smaller the size of window is, the higher the accuracy of recognition result is. But the size of window is too small to show the region rule and recognize defects. Figure 9 lists the influence of window’s change on recognition result. Therefore, the minimum value of window is determined by actual image, and the principle is that reflecting fully local rule of image is better.
186
JOURNAL OF MULTIMEDIA, VOL. 5, NO. 2, APRIL 2010
experiment was carried out 20 times. In reference [16], a new defect detection scheme was proposed, which consists of an odd symmetric real-valued Gabor filter, an even symmetric real-valued Gabor filter and one smoothing filter. They selected 39 defects samples in their experiments, including warp-lacking, stain, Oil stain, holes, and refuse defects, the overall recognition accuracy was 96.2%. In reference[25], they used wavelet transform to recognize fabric defects, including hole, oil stain, stripe, folding mark, and so on, they tested 45 samples, and all made pre-training, the accuracy was 100%. The accuracy comparison between these algorithms is shown in Table 3.
Fig.9 Recognition Result of Different Window Size
V.
ALGORITHM COMPARISON
TABLE3. RECOGNITION ACCURACY COMPARISON BETWEEN PROPOSED
A. Recognition Accuracy
ALGORITHM AND FREQUENCY DOMAIN ALGORITHMS
Using fabric defection system designed by the algorithm in this paper, selecting 1000 meters defective sample fabrics, we detect them and make the comparison with manual detection. The result is shown in Table 1. TABLE1. MANUAL AND COMPUTER DETECTION COMPARISON Image defect class warp-lacking Stain Oil stain Holes Refuse
Manual detection 67 128 87 25 44
Computer detection 68 123 88 25 43
Accuracy (%) 98.5 96.1 98.9 100 97.7
As shown in Table 2, we select some representative literatures in space domain algorithms for fabric defect recognition and make the comparisons with the algorithm in this paper. In reference [4], they used local binary patterns (LBPS) to recognize the fabric defects, their algorithm is simple, but the influence of recognition result by window’s size is big, and other kinds of defects was not discussed too. In reference [5], they defined label co-occurrence matrix, used detection method of abnormal point, which could distinguish normal texture and defect better. They mainly recognized defects of stain, 20 TABLE 2 RECOGNITION ACCURACY COMPARISON BETWEEN PROPOSED ALGORITHM AND MAIN SPACE DOMAIN ALGORITHMS
Algorithm Reference[4] Reference[5] Reference[11] proposed
Warp lacking 98 --98.5
Defect Type Oil holes stain -96 97 93 --92 ----96.1 98.9 100 stain
refuse 95 ---97.7
samples were selected and the recognition accuracy was 93%. In reference [11], they proposed multiscal localization method with statistics principle and recognized defects of blur. They concluded two results, the accuracy was 71% when rough localization, the accuracy was 92% when exact localization. In frequency domain methods, Reference [9] used a feasible approach for the recognition of fabric defects based on discrete wavelet transform and back-propagation neural network, they have recognized 6 kinds of defects of fabrics, the recognition accuracy was 99.2%, when
© 2010 ACADEMY PUBLISHER
Algorithm Reference[9] Reference[16] Reference[25] proposed
warp-lacking 100 96.2 100 98.5
Defect Type Oil stain 100 95 96.2 96.2 100 100 96.1 98.9 stain
holes
refuse
100 96.2 100 100
100 96.2 100 97.7
From Table1 to Table 3, we find that the recognition accuracy of algorithm proposed in this paper is more excellent, and the recognition accuracy is better than that of space domain algorithms. From above data, recognition result is worse than that of frequency domain algorithms, but the data from frequency domain algorithms is the recognition result of static image or the recognition result with very slow speed, so their result is not apply to dynamic detection and their accuracy has no reference meaning. B. Recognition Speed From Table3, the recognition effect of frequency domain algorithms is better, but most algorithms need large calculation and pre-training. Their recognition speed is slow and efficiency is low. More importantly, from frequency domain algorithms in reference [12-24], we don’t find any evidence about dynamic image detection application, and the data are obtained from static image. These results are not applied to fabric detection with quick moving. The algorithm proposed in this paper recognizes defects only by “super fuzzy” feature of image region, which can save more workload. Table 4 lists the recognition speed data from experiments. TABLE4. RECOGNITION ACCURACY INFLUENCE BY FABRIC SPEED Fabric speed 30m/min 60m/min 120/min
warp-lacking 98.5 98.5 93.5
Defect Type Oil stain holes stain 98.4 100 100 96.1 98.9 100 93.0 95.6 92.0
refuse 97.7 97.7 93.2
From Table 4, when the fabric speed is 60m/min, the recognition accuracy using algorithm proposed in this paper is high. Especially when the speed is 30m/min, the accuracy is reach two 100%. When the speed is 120m/min, the recognition accuracy began to reduce. But in reference [1-24], only reference [4] mentioned that
JOURNAL OF MULTIMEDIA, VOL. 5, NO. 2, APRIL 2010
their fabric speed was 40m/min，other literatures did not mention the fabric speed. All these are proved that these algorithms for dynamic fabric recognition must be demonstrated. C.
Category of Recognition
The image’s defects for analysis in reference [1] to reference [11] are mostly according to 1 to 2 kinds of defects, and only 6 defects are analyzed and recognized in reference [4] and reference [2]. Reference [12] to reference [24] display that there are 1 to 2 defects for analysis and recognition. Only reference [13], reference [18] and reference [21] declared that their algorithms could recognize more defects categories, but they did not give convincing data explanation. In the algorithm proposed in this paper, defects category is not especial demand, and the key of the algorithm is the comparison and clustering between “super fuzzy” features. As long as different region from normal texture rule is found, we can consider that this region is defect region, no considering the defect category in this region. Therefore, this algorithm has generality, the category of recognition defects is extensive. We list the number of defects categories with main algorithms as shown in Table 5. TABLE5. THE NUMBER OF RECOGNITION DEFECTS CATEGORIES Algorithm
Reference
Geometric method Co-occurrence matrix Neural network Probability statistic Fourier Gabor Wavelet Proposed
[1]-[4] [5][6] [7][8][9] [10][11] [12][14] [15]-[17] [18-24]
The number of defects categories 1—6 1-2 Pre-training situation 1 1-6 1-8 1-8 23
VI. CONCLUSION In this paper, we segment out the image into some windows and construct its multi-characteristic vector, which can describe the local rule of image. The “super fuzzy” vector is introduced, so that each characteristic of window has the difference of the primary and secondary, and recognition efficiency and accuracy of image’s defect region is improved. The window coefficient is introduced, so that recognition speed and accuracy can be adjusted according to different image’s characteristics. A fuzzy clustering algorithm is presented considering “super fuzzy” vector and window coefficient, some functions and formulas from this algorithm can recognize many kinds of image’s defects regions with high accuracy, high speed, broad application, and no pre-training. REFERENCES [1] D.r. Rohrmus, “Invariant and adaptive geometrical texture features for defect detection and classification”, Pattern Recognition, vol.38, pp.1546-1599, Feb 2003. [2] Lu Yun, Zhang Jingmiao and Jiang Jianwei, “Fabric Defect
© 2010 ACADEMY PUBLISHER
187
Detection Method Based on Image Distance Difference”, Microcomputer Information, vol.23, pp.303-306, Apri 2007. [3] Wang Feng, Jiao Guotai and Du Ye, “Method of Fabric Defects Detection Based on Mathematical Morphology”, Journal of Test and Measurement Technology, vol.21,pp.515-518, Jun 2007. [4] F. Tajeripour, E. Kabir, and A. Sheikhi, “Fabric Defect Detection Using Modified Local Binary Patterns”, EURASIP Journal on Advances in Signal Processing, vol.1,pp.1-12,Jan 2008 [5] Zou Chao, Zhu desen and Xiao Li, “Textural defect detection based on label co-occurrence matrixm”, Journal of Huazhong University of Science and Technology(Nature Science), vol.34,pp.25-28,Jun 2006. [6] Zhang Y F and Bresee R R., “Fabric defect detection and classification using image analysis”, Textile Research Journal,vol.65,pp.1-9,Jan 1995. [7] Chen Pei-wen, “Classifying textile faults with a backpropagation neural network using power spectral”, Textile Research Journal, vol.68, pp.121-126, Feb 1998. [8] Choonjong Kwak, Jose A.Ventura and Karim Tofang-sazi, “A neural network approach for defect identification and classification on leather fabric”, Journal of Intelligent Manufacturing, vol.11,pp.485-499,May 2000. [9] Liu Jianli and Zuo Baoqi, “Identification of fabric defects based on discrete wavelet transform and back-propagation neural network”, Journal of the Textile Institute, vol.98,pp.355-362,Apri 2007. [10] Shuang-Wu Zhu, Hong-Yang Hao, Peng-Yang Li et al, “Fabric Defects Segmentation Approach Based on Texture Primitive”, 2007 International Conference on Machine Learning and Cybernetics, vol.3, pp.1596-1600, Aug 2007,. [11] Arunkumar Gururajan, Hamed Sari-sarraf and Eric F. Hequet, “Statistical approach to unsupervised defect detection and multiscale localization in two-texture images”, Optical Engineering, vol.47,pp.027202(1)-027202(10),Feb 2008. [12] Chan CH and Grantham KHP, “Fabric defect detection by fourier analysis”, IEEE Transaction on Industry Application, vol.36, pp.1267-1276, May 2000. [13] Wood E J., “Applying fouier and associated transforms to pattern characterization in textiles”, Textile Res. J., vol.60, pp.212-222, Apri 1990. [14] Ravandi S A H,Toriumi K., “Fourier transform analysis of plain woven fabric appearance”, Textile Res. J., vol.65,pp.678-683, Jul 1995. [15] Bodnarova, Bennamoun and S.Latham, “Optimal Gabor filters for textile flaw detection”, Pattern Recognition,vol.35,pp.2973-2991,Nov 2002. [16] K.L. Mak and P. Peng, “An automated inspection system for textile fabrics based on Gabor filters”, Robotics and Computer-Integrated Manufacturing, vol.24,pp.359-369,Jan 2008. [17] Bodnarova A,Bennamoun M and Latham S., “Optimal gabor filters for textile flaw detection”, Pattern Recognition, vol.35, pp.2973-2991, Jan 2002. [18] Aruvazhagan S and Ganesan L., “Texture classification using wavelet transform”, Pattern Recognition letters,
188
vol.24,pp.1513-1521,Dec 2003. [19] Jasper W J,Garnier S J and Potlapalli H, “Texture characterization and defect detection using adaptive wavelets”, Optical Engineering, vol.35,pp.3140-3149,Nov 1996. [20] Portilla J and Simoncelli E.P.A, “Parametric TextureModel based on Joint Statistics of Complex Wavelet Coefficients”, International Journal of Computer Vision, vol.40, pp.49-61, Jan 2002. [21] Soo Chang Kim and Tae Jin Kang, “Texture classification and segmentation using incomplete tree structured wavelet packet frame and Gaussian mixture model”, Imaging Systems and Techniques, vol.13, pp.46-51, May 2005. [22] Shen Jianqiang,Geng Zhaofeng and Zou Xuan, “Detection Method for Fabric Texture Direction Based on Wavelet Transform”, Computer Engineering, vol.33, pp.182-184, Jan 2007. [23] W. Jasper, J.Joines, and J. brenzovich, “Fabric defect detection using a genetic algorithm tuned wavelet filter”, Journal of the Textile Institute, vol.96, pp.43-54, Jan 2005. [24]Chang-Chiun Huang and Tong-Fu Lin. “Image Inspection of Nonwoven Defects Using Wavelet Transforms and Neural Networks”. Fibers and Polymers, vol.9, pp.633-638, May 2008 [25] Ozdemir D and Akarun L., “A fuzzy algorithm for color quantization of images”, Pattern Recognition, vol.35, pp.1785-1791, Oct 2002. [26] Menard M, CourboulaY V and DardignaC P., “Possibilistic and probabilistic fuzzy clustering: Unification within the framework of the non-extensive thermostatistics”, Pattern Recognition, vol.36,pp.1325-1342,Jun 2003. [27] Kaymak U and Setne M., “Fuzzy clustering with volume prototypes and adaptive cluster merging”, IEEE Trans. on Fuzzy Systems, vol.10,pp.706-712, Jun 2002. [28] Schneider A., “Weighted possibilistic clustering algorithms”, Proc. of the 9th IEEE Int’l Conf. on Fuzzy Systems, vol.1,pp.176-180, 2000. [29] Liu SH and Lin JS, “Vector quantization in DCT domain using fuzzy possibilistic c-means based on penalized and compensated constraints”, Pattern Recognition, vol.35,pp.2201-2211, Oct 2002. [30] Zhang JS and Leung YW, “Improved possibilistic c-means clustering algorithms”, IEEE Trans. on Fuzzy Systems, vol.12, pp.209-217, Feb 2004. [31] Ichihashi H, Miyagishi K and Honda K, “Fuzzy C-means clustering with regularization by K-L information”, Proc. of the 10th IEEE Int’l Conf. on Fuzzy Systems, vol.2,pp.924-927, Aug 2001. [32] Yasuda M, Furuhashi T, Matsuzaki m and Okuma S, “A study on statistical mechanical characeteristics of fuzzy clustering”, Proc. Of IEEE Int’l Conf. on Systems, Man and Cybemetics,pp.2415-2420, Apri 2001. [33] Wei C and Fahn C, “The multisynapse neural network and its application to fuzzy clustering”, IEEE Trans. on Neural Networks, vol.13, pp.600-618, Mar 2002. [34] Ma Jun and Hao Lu, “An Optimal Algorithm for Fuzzy Classification Problem”, Journal of Software, vol.12,pp.578-581,Apri 2001.
© 2010 ACADEMY PUBLISHER
JOURNAL OF MULTIMEDIA, VOL. 5, NO. 2, APRIL 2010
[35] Wang JH, Shen Z and Hu YF, “An applicable and efficient clustering algorithm”, Journal of Software, vol.15,pp.697-705, May 2004. Zhe LIU was born in China, March 1972. He is associate professor in Zhongyuan University of Technology and is pursuing Doctor Degree in Tanjin Polytechnic University now. He is working on the theoretic and application research of computer vision all along. During these years, more than 10 high quality papers he wrote had been published, and five scientific research projects of the Province he masterminded have been finished. He also has obtained two national patented inventions. Some papers and research projects has won the government prizes. Xiuchen WANG was born in China, October 1972. She has obtained the master degree from Xian Technology University. She is an associate professor of Zhongyuan University of technology now. Her research interest is computer application and pattern recognition. During these years, she has delivered more than ten important papers, and has finished three scientific research projects from the Province. Some papers and research projects have won the government prizes.
181
Image Defect Recognition Based on “Super Fuzzy” Characteristic Zhe LIU and Xiuchen WANG Zhongyuan University of Technology, Zhengzhou, China Email: [email protected], [email protected]
Abstract—In this paper, we propose a new defects recognition algorithm for dynamic image based on “super fuzzy” feature. With this algorithm, the image is divided into some variable windows, and the eigenvector of each window is constructed. We introduce “super fuzzy” vector to make window vectors “super fuzzy” processing, thus the window feature has “super fuzzy” characteristic with the difference of the primary and secondary. Also we present window coefficient to adjust recognition speed and accuracy according to different images. Furthermore, objective function, membership function and clustering center calculation function of fuzzy clustering algorithm with window coefficient and “super fuzzy” vector are proposed in this paper. At last, we take example for fabric defects detection with this algorithm, list recognition results, discuss recognition result influence by “super fuzzy” feature and size change of window, and make some comparison with other algorithms. The conclusion shows that this algorithm can recognize more categories of image abnormal regions with high-accuracy, high-speed, no-training and extensive application. Index Terms—super fuzzy feature, variable window, fuzzy, clustering, recognition, defect region
I.
INTRODUCTION
Image defect region recognition is an important research issue of computer pattern recognition with wide application prospect. The aim is to recognize automatically abnormal region with rule mutation according to whole characteristic rule of image. This research result can apply to computer-aided product detection, quality evaluation and other fields in every industry, and has key meaning to improve quality detection automation in enterprises. Many scholars have done some research about it. Now, there are two main methods. The first approach is space domain algorithms. In these algorithms, the image’s gray value is extracted for analysis and recognition. Main algorithms are geometric method [1-4], co-occurrence matrix [5-6], neural network method [7-9], probability statistic [10-11], and so on. In geometric method, the basis is unstable because of variable defects shape, the defect recognition accuracy is difficult to reach, and involves in massive calculation. In co-occurrence matrix,
© 2010 ACADEMY PUBLISHER doi:10.4304/jmm.5.2.181-188
image’s quality requirement is very high. In order to construct multi-dimensional feature matrix and analyze its characteristic, its calculation is very large too, so the recognition speed and accuracy are not idea. The neural network approach once was research hotspot, but its calculation is large and recognition needs pre-training. When to recognize static fabric sometimes need more time, and can not adapt to dynamic fabric detection. In probability statistic, because the probability of all gray value must be calculated and compared, and the probability statistic is more depend on the image quality, the problem of recognition speed and accuracy is existed too. The second approach is frequency domain algorithms. In these algorithms, image is transformed to certain spectrum for research and analysis. At first, Fourier transform is main algorithm [12-14], But Fourier transform algorithm is so difficult to recognize dynamic window’s characteristic that it couldn’t apply to practice application of dynamic detection. Later improved Fourier transform is appear, the more representative approach is Gabor algorithm [15-17]. Though Gabor algorithm has the computing capacity for dynamic window, the problem of large computing is also existed. Wavelet algorithm [18-24] was research hotspot for a time, and obtained some achievements on individual defects detection. But some defects are distinct appear after image transforming to spectrum in frequency domain algorithms, this transform is quite insensitive to many defects. That is, the result after transform could not further recognize defects. Moreover, spectrum transforms involves complex transform function. It increases calculation load and influences the recognition speed. More importantly, even though the image is made spectrum transform, the secondary recognition for spectrum characteristic has to face, which will further influence the recognition speed and accuracy. Aiming at these instances, this paper proposes a novel recognition algorithm with fuzzy clustering based on “super fuzzy” characteristic. This recognition algorithm can recognize defect regions of dynamic image. The image is divided into some windows and the window coefficient is introduced for describing image rule. The characteristic vector of each window is constructed and the “super fuzzy” vector is introduced so that each window characteristic has primary and secondary
182
JOURNAL OF MULTIMEDIA, VOL. 5, NO. 2, APRIL 2010
difference. On this basis, some matching objective functions and membership functions of fuzzy clustering are proposed in this paper. This algorithm is tested by recognizing fabric defects, result shows that this algorithm can recognize many kinds of image’s defect regions with high-accuracy, high-speed, extensive application and no re-training.
rule. If pixel points are carried through clustering one by one, it will arouse clustering confusion. Recognition accuracy influence by change of region window size is also great, so window factor must be considered during recognition process. To sum up, we propose a pattern recognition algorithm with fuzzy clustering based on “super fuzzy” characteristic and variable window. The thought is shown in Figure1.
II. ALGORITHM THOUGHT Image’s defect region is produced owing to rule characteristic change of this region. In daily life, one person with no professional knowledge even a sensible child can distinguish defect region of random images quickly and easily by naked eye. After studying we find that recognition process by human (is very simple). This process contains three periods. Firstly, people looks through an image wholly and search the main rule of image quickly. Secondly, the image is segmented and its rule is found one by one. Finally, abnormal region is found abiding by main image’s rule. This process is a fuzzy judgment all the way, but it always has a very accurate result. This paper tries to study a new pattern recognition algorithm, so that computer can recognize defect region quickly and accurately based on fuzzy judgment abiding by above human recognition process. We think that any images have many rule characteristics, and these characteristics have difference with primary and secondary. For example, when we recognize fabric defects, the characteristic of silk is obvious luster, so abnormal region of defects represents abnormal luster firstly. Luster of cotton fabric changes slow, but its structure texture rule change can make obvious change of fabric appearance, so abnormal region of defects represents difference of fabric structure texture firstly. Abnormal region of defects in wool fabric represents the difference of fabric partial roughness. Therefore, we introduce a concept of “super fuzzy” characteristic, window can describe with main features instead of whole feature. the calculation load and time is greatly keep down. The theory of “super fuzzy” characteristic can explain vividly with an example. If we want to distinguish five people from different provinces, and every people has filled in a table with province, city, section, street and unit. Do we recognize them one by one according to their detail information or according to their fuzzy province? This problem is very simple, we only make “super fuzzy”, and not consider latter detail. Only the “super fuzzy” feature—province, we can distinguish the source region of each person without any error. Furthermore, if three persons come from the same province, we can use the second “super fuzzy” feature—city to compare these three persons. The result is also clear. This process, which is a simplify process of complex recognition problem and greatly shorten recognition time, is happen in many cases in daily life. In addition, image recognition is recognized abiding by regions one by one. Only certain region can present certain rule, single pixel point can not show the image’s
© 2010 ACADEMY PUBLISHER
Fig.1 Algorithm thought
III. RECOGNITION ALGORITHM BASED ON “SUPER FUZZY” CHARACTERISTIC
A. Multi-characteristic Model of Window If the image f ( x, y ) is divided into N × M number variable grids according to lateral interval ∆x and longitudinal interval ∆y , Τij denotes eigenvector in the i row and the j array region, which has n factors, respectively denotes n characteristic parameters of local region with λ1 , λ2 ,…, λn . If the degree of membership of each region is uij ，the image fuzzy set is expressed as: N M
Μ=
uij
UU T i =1 j =1
（1）
ij
Considering multi-characteristic of region，the formula (1) is decomposed to fuzzy set with multi-characteristic. The fuzzy set is: N M
n
Μ=
∑ UU T .λ ) (
k =1 i =1 j =1
Where
uijk
uijk
（2）
ij k
denotes the degree of membership of
the k characteristic in the i row and the Tij .λk denotes the k characteristic the
j array i
region， row and the
j array region. Thus the calculating formula of the degree of membership is given as:
uijk = 1 +
Tij .λk u
（3）
Where u denotes the maximum value of the characteristic. Figure 2 shows a multi-characteristic model diagram of an image. Here only six characteristics are selected for analysis, they are denoted by λ1 , λ2 , λ3 , λ4 , λ5 , λ6 . The k
JOURNAL OF MULTIMEDIA, VOL. 5, NO. 2, APRIL 2010
color area of each characteristic represents the degree of membership of this characteristic.
183
This paper proposes a novel fuzzy clustering algorithm. The “super fuzzy” vector is introduced before constructing objective function. The characteristic influence on clustering is changed according to this vector. Also the window coefficient is introduced, and calculating speed is determined according the change of image actual need and recognition accuracy. In the universes X , the fuzzy set Tij is Tij
={ λ1, λ2 ,..., λn } ,
vector E is E = ( m1 ， m2 ，…， n
Fig2. Schematic diagram of multi-characteristic
mn
)
⊂ [0,1]
prescriptive B. “Super Fuzzy” Processing of Multi-characteristic Fuzzy clustering algorithm belongs to unsupervised learning in pattern recognition. It need not pre-training; can directly achieve the goal of automatic classification by machine learning. The most successful and extensive application is the Fuzzy C-Mean algorithm(FCM), which was proposed by Dunn in 1974 and extended by Bezdek. But this algorithm has many obvious disadvantages, such as it only fits to find clusters with globosity shape, this algorithm is more sensitive to voice data, algorithm couldn’t sure to converge the minimum point of objective function, and so on. Later many researchers improved the FCM and presented some new fuzzy clustering algorithms. Ozdemir[25], Menard[26], and Kaymak et al.[27] increased the euclidean metric stability under noise environment by improving the metric form of fuzzy clustering algorithm, and reduced the algorithm’s initial value and shape size’s sensitivity of cluster. Schneider [28], Liu [29], Zhang et al. [30]changed the membership function to fit actual different influence of each data point. Ichihashi [31], Yasuda [32], and Wei et al.[33] presented a fuzzy clustering algorithm under maximum entropy theory by introducing the entropy. Ma [34] and Wang et al. [35] did some research of optimizing the fuzzy clustering algorithm and making it more effective. However, in existing fuzzy clustering algorithms, general and regional feature rule of image is omitted. The cluster analysis is only by local pixel point’s features not by rule region’s features with the distinction between the primary and the secondary, so some points in defect regions and some points in normal regions are both clustered into the same cluster. At the same time, these algorithms ignore the clustering influence by the rule region’s size, thus the clustering analysis is done in a very thick data set. These two disadvantages make existing algorithms not only waste the calculating time but also influence the recognition accuracy and application scope. In those traditional fuzzy clustering algorithms, an image is considered as n number characteristic vectors as Tij (i = 1,2..., N , j = 1,2,..., M ) , and divides into c number clusters. Clustering center of each cluster is obtained， and objective function achieves minimum. This algorithm is carried though only under the condition of characteristic averaging ， and the image is not make windows segmentation.
© 2010 ACADEMY PUBLISHER
∑m
, and
s =1
s =1
vector
Tij
. The scalar product of and
E
Tij • E
is
{ m1λ1, m2λ2 ,..., mnλn } , the whole characteristic λ
ij =
=
n
∑ (m λ ) s s
2
.
s =1
These factors m1 , m2 ,…, mn are called “super fuzzy” factors corresponding each fuzzy sets, and E is called “super fuzzy” vector. The aim to introduce “super fuzzy” vector is to adjust the proportion of each characteristic in every window according to different image characteristic, and these characteristics have the difference of the primary and secondary. The most important characteristic influencing image quality is regarded as main characteristic according to different situation, so that continued clustering analysis will more accurate and easy. As shown in Figure3, a, b and c respectively shows new characteristic composition of six characteristics shown in Figure2 after fuzzy processing with different “super fuzzy” vector. The color area still denotes its degree of membership. Figure 3-a shows the result of “super fuzzy” feature, in which λ1 is the main characteristic, λ3 、 λ5 are the secondary features. Figure 3-b shows “super fuzzy” feature, in which λ1 is the main characteristic, and other characteristics are not considered. Figure 3-c shows “super fuzzy” feature, in which λ1 is the main characteristic, λ5 is the secondary feature.
Fig3. Result of “super fuzzy” characteristic
C.
Variable Window Coefficient Introducing
The image f ( x, y) is divided into N × M number variable grids according to lateral interval ∆x and longitudinal interval
∆y
. If ε
=
M +N M ×N
, here ε
is the window
coefficient. The more particular the image is divided, the little the ε value is, the higher the precision of clustering is. In general reason, the change of the number N and M is not acute for the same image, the value ε will not arouse an acute fluctuant result. Figure 4 shows the clustering result after different window segmentation. Obviously, the more particular the
184
JOURNAL OF MULTIMEDIA, VOL. 5, NO. 2, APRIL 2010
window is segmented, the higher the precision of defect region is found. The more rough the window is segmented, such as window segmentation with “5×7”, the more imprecise clustering result is. It must be point out that some different pixel points, which belong to image’s defect region and normal region, also appear some local feature consistent or approximate. If the window is divided too particular, some pixel points in defect region and normal region will be consider as in a same class. Therefore, the window segmentation must be subject to image’s local rule, not too particular.
In order that clustering center calculation is more near main characteristic, the “super fuzzy” vector is introduced as: M ×N
∑
vi = k =1
1
uikε ( E • xk ) (i = 1,2,..., c)
1
N
（6）
∑u ε
ik
k =1
uik =
1 c
⎛d ⎜ ik ⎜ d jk i =1 ⎝
⎞ ⎟ ⎟ ⎠
∑
(i = 1,2,..., c; k = 1,2,..., n)
ε
（7）
IV. EXAMPLE DEMONSTRATION
Fig4. Clustering influence by window segmentation
D.
Clustering Algorithm
The clustering process we have used is shown in Figure5.
Computer recognition for fabric defect is representation of image’s defect recognition. Up to now, there is no mature method to solve this problem. We develop a minitype computer automatic recognition system for fabric defect using above fuzzy clustering algorithm based on “super fuzzy” characteristic. The transmitting speed of this system is 60 meter/minute and the width is 1.1 meter. Dynamic fabric image is transmitted to the computer using a high definition camera， the recognition software is programmed with matlab7.0. In order to validate this fuzzy clustering algorithm, a program is programmed especially to output the middle result of image recognition, and defect regions are marked by red dot. A. Fabric Image’s Multi- Characteristic Construction
Fig5. Clustering process
Considering the multi-characteristic and window coefficient, objective function is constructed as: M ×N c
Om (u , v) =
∑∑
1
uik ε d ( xk , vt )
（4）
Six characteristics showing multi-characteristic of fabric image’s rule are constructed. λ1 denotes gray mean value of pixel points in each window, λ2 denotes fluctuate degree of each window’s inner gray, λ3 and λ4 denote texture roughness degree of each region, λ5 and λ6 denotes difference between the maximum local gray value and minimum local gray value in the region, they show the uniformity of texture. If window’s pixel size of the i row and the j array are N ij × M ij , Gkl denotes gray value of each pixel point, G denotes gray mean value, they can be expressed as following:
k =1 i =1 c
where，
∑u
ik = 1 ， uik ∈ (0,1)
i =1
Here “super fuzzy” vector is introduced and variation in cluster is obtained: 2 d ( xk , vi ) = E • xk − E • vi （5） When variation in cluster is calculated, formula（5） shows that the window characteristic has the difference of the primary and secondary considering “super fuzzy” vector. Fuzzy clustering can be made according to different image characteristic, and calculating time can be saved and recognition accuracy can be improved.
N ij M ij
∑∑ G
（8）
M ij × N ij N ij M ij
∑∑ (G
kl − G )
k =1 l =1
λ2 =
M ij N ij
2
（9）
M ij
λ3 = −
∑v
p ln v p
（10）
p =1 N ij
λ4 = −
∑w p =1
© 2010 ACADEMY PUBLISHER
kl
λ1 = k =1 l =1
p ln w p
（11）
JOURNAL OF MULTIMEDIA, VOL. 5, NO. 2, APRIL 2010
185
characteristics. The “super fuzzy” vector of middle two
where： N ij
∑
vp =
H kl
k =1 N ij M ij
∑∑ H
， wp = kl
k =1 l =1
λ5 =
l =1
k (l −
∆x ) 2
, Gkl , G
k (l +
∆x ) 2
λ6 =
k =1
(k −
∆y )l 2
∑∑ H
. Here only has 4 factors, other
features λ5 、 λ6 are deleted. The “super fuzzy” vector of the forth graph is similar to Graph 6.
l =1 N ij M ij
) − min(G
k (l −
M ij − 1
N ij
∑ (max(G
∑
H kl
1 1 1 1 , , , ) 4 4 4 4
kl
k =1 l =1
M ij
∑ (max(G
graphs is: E = (
M ij
, Gkl , G
(k +
∆y )l 2
) − min(G
N ij − 1
(k −
∆x ) 2
∆y )l 2
, Gkl , G
, Gkl , G
k (l +
(k +
∆x ) 2
∆y )l 2
))
（12）
))
（13）
B. General Result
Fig.7 Recognition Result of changing feature number in “super fuzzy”
Here only image’s part with defect is intercepted and result comparison is displayed due to the limited space. As shown in Figure 6, Each intercepted image’s size is 8cm×8cm, and 20×15 number windows are segmented. The graphs which name with “-1” are the former graphs, the graphs which name with “-2” are recognition graph of middle result. the graphs which name with “-3” are the position region graph by the system orientation after clustering calculation. If we want to continue to extract defect shape, only make analysis in defect regions and extract the defect silhouette. In these results, the “super fuzzy” vector is: E = ( 1 , 1 , 1 , 1 , 1 , 1 ) . 6 6 6 6 6 6
vector
When the number of characteristic is unchanged and “super fuzzy” vector is changed, fuzzy clustering recognition results are shown in Figure 8. The “super 13 13 1 ,0, , ,0,0) ，the 15 30 30 1 1 is: E = (0,0,0,0, , ) , the 2 2
fuzzy” vector in Graph 8-a is: E = ( “super fuzzy” vector in Graph 8-b
“super fuzzy” vector in Graph 8-c is same to Figure 6. From these results, we find that recognition result shown in graph 8-b can obtain the same recognition result as shown in graph 8-c only considering features λ5 , λ6 . In this way, we can simple recognition calculating work greatly. But recognition result shown in Graph8-a is different to Graph 8-c due to the difference of “super fuzzy” vector. All these show that the main characteristics of Graph 8-sample are λ5 and λ6 .
Fig.6. Fabric Defect Recognition Result Using Clustering Algorithm on “super fuzzy” feature
C.
Recognition Result Influence by “Super Fuzzy” Fig.8 Recognition Result of Different “super fuzzy” features
Feature Change rule of fabric image with different structure is different, so that characteristic describing each fabric’s region is different. Commonly, the number of characteristic is many for thicker fabric, and be contrary for thinner fabric. The more characteristic number is selected, the higher the fabric recognition accuracy is, and the slower the recognition speed is. The difference of selected “super fuzzy” vector will reduce the number of image’s characteristic. That is the secondary feature of image is not considered, and only the main feature is considered. Therefore, fabric recognition result will be influenced. As is shown in Figure 7, graphs are the fuzzy clustering recognition result by selecting 4 different
© 2010 ACADEMY PUBLISHER
D.
Influence of Variable Window
The size of window also influences the recognition result. Commonly, the bigger the size of window is, the bigger the recognition result error is, whereas the smaller the size of window is, the higher the accuracy of recognition result is. But the size of window is too small to show the region rule and recognize defects. Figure 9 lists the influence of window’s change on recognition result. Therefore, the minimum value of window is determined by actual image, and the principle is that reflecting fully local rule of image is better.
186
JOURNAL OF MULTIMEDIA, VOL. 5, NO. 2, APRIL 2010
experiment was carried out 20 times. In reference [16], a new defect detection scheme was proposed, which consists of an odd symmetric real-valued Gabor filter, an even symmetric real-valued Gabor filter and one smoothing filter. They selected 39 defects samples in their experiments, including warp-lacking, stain, Oil stain, holes, and refuse defects, the overall recognition accuracy was 96.2%. In reference[25], they used wavelet transform to recognize fabric defects, including hole, oil stain, stripe, folding mark, and so on, they tested 45 samples, and all made pre-training, the accuracy was 100%. The accuracy comparison between these algorithms is shown in Table 3.
Fig.9 Recognition Result of Different Window Size
V.
ALGORITHM COMPARISON
TABLE3. RECOGNITION ACCURACY COMPARISON BETWEEN PROPOSED
A. Recognition Accuracy
ALGORITHM AND FREQUENCY DOMAIN ALGORITHMS
Using fabric defection system designed by the algorithm in this paper, selecting 1000 meters defective sample fabrics, we detect them and make the comparison with manual detection. The result is shown in Table 1. TABLE1. MANUAL AND COMPUTER DETECTION COMPARISON Image defect class warp-lacking Stain Oil stain Holes Refuse
Manual detection 67 128 87 25 44
Computer detection 68 123 88 25 43
Accuracy (%) 98.5 96.1 98.9 100 97.7
As shown in Table 2, we select some representative literatures in space domain algorithms for fabric defect recognition and make the comparisons with the algorithm in this paper. In reference [4], they used local binary patterns (LBPS) to recognize the fabric defects, their algorithm is simple, but the influence of recognition result by window’s size is big, and other kinds of defects was not discussed too. In reference [5], they defined label co-occurrence matrix, used detection method of abnormal point, which could distinguish normal texture and defect better. They mainly recognized defects of stain, 20 TABLE 2 RECOGNITION ACCURACY COMPARISON BETWEEN PROPOSED ALGORITHM AND MAIN SPACE DOMAIN ALGORITHMS
Algorithm Reference[4] Reference[5] Reference[11] proposed
Warp lacking 98 --98.5
Defect Type Oil holes stain -96 97 93 --92 ----96.1 98.9 100 stain
refuse 95 ---97.7
samples were selected and the recognition accuracy was 93%. In reference [11], they proposed multiscal localization method with statistics principle and recognized defects of blur. They concluded two results, the accuracy was 71% when rough localization, the accuracy was 92% when exact localization. In frequency domain methods, Reference [9] used a feasible approach for the recognition of fabric defects based on discrete wavelet transform and back-propagation neural network, they have recognized 6 kinds of defects of fabrics, the recognition accuracy was 99.2%, when
© 2010 ACADEMY PUBLISHER
Algorithm Reference[9] Reference[16] Reference[25] proposed
warp-lacking 100 96.2 100 98.5
Defect Type Oil stain 100 95 96.2 96.2 100 100 96.1 98.9 stain
holes
refuse
100 96.2 100 100
100 96.2 100 97.7
From Table1 to Table 3, we find that the recognition accuracy of algorithm proposed in this paper is more excellent, and the recognition accuracy is better than that of space domain algorithms. From above data, recognition result is worse than that of frequency domain algorithms, but the data from frequency domain algorithms is the recognition result of static image or the recognition result with very slow speed, so their result is not apply to dynamic detection and their accuracy has no reference meaning. B. Recognition Speed From Table3, the recognition effect of frequency domain algorithms is better, but most algorithms need large calculation and pre-training. Their recognition speed is slow and efficiency is low. More importantly, from frequency domain algorithms in reference [12-24], we don’t find any evidence about dynamic image detection application, and the data are obtained from static image. These results are not applied to fabric detection with quick moving. The algorithm proposed in this paper recognizes defects only by “super fuzzy” feature of image region, which can save more workload. Table 4 lists the recognition speed data from experiments. TABLE4. RECOGNITION ACCURACY INFLUENCE BY FABRIC SPEED Fabric speed 30m/min 60m/min 120/min
warp-lacking 98.5 98.5 93.5
Defect Type Oil stain holes stain 98.4 100 100 96.1 98.9 100 93.0 95.6 92.0
refuse 97.7 97.7 93.2
From Table 4, when the fabric speed is 60m/min, the recognition accuracy using algorithm proposed in this paper is high. Especially when the speed is 30m/min, the accuracy is reach two 100%. When the speed is 120m/min, the recognition accuracy began to reduce. But in reference [1-24], only reference [4] mentioned that
JOURNAL OF MULTIMEDIA, VOL. 5, NO. 2, APRIL 2010
their fabric speed was 40m/min，other literatures did not mention the fabric speed. All these are proved that these algorithms for dynamic fabric recognition must be demonstrated. C.
Category of Recognition
The image’s defects for analysis in reference [1] to reference [11] are mostly according to 1 to 2 kinds of defects, and only 6 defects are analyzed and recognized in reference [4] and reference [2]. Reference [12] to reference [24] display that there are 1 to 2 defects for analysis and recognition. Only reference [13], reference [18] and reference [21] declared that their algorithms could recognize more defects categories, but they did not give convincing data explanation. In the algorithm proposed in this paper, defects category is not especial demand, and the key of the algorithm is the comparison and clustering between “super fuzzy” features. As long as different region from normal texture rule is found, we can consider that this region is defect region, no considering the defect category in this region. Therefore, this algorithm has generality, the category of recognition defects is extensive. We list the number of defects categories with main algorithms as shown in Table 5. TABLE5. THE NUMBER OF RECOGNITION DEFECTS CATEGORIES Algorithm
Reference
Geometric method Co-occurrence matrix Neural network Probability statistic Fourier Gabor Wavelet Proposed
[1]-[4] [5][6] [7][8][9] [10][11] [12][14] [15]-[17] [18-24]
The number of defects categories 1—6 1-2 Pre-training situation 1 1-6 1-8 1-8 23
VI. CONCLUSION In this paper, we segment out the image into some windows and construct its multi-characteristic vector, which can describe the local rule of image. The “super fuzzy” vector is introduced, so that each characteristic of window has the difference of the primary and secondary, and recognition efficiency and accuracy of image’s defect region is improved. The window coefficient is introduced, so that recognition speed and accuracy can be adjusted according to different image’s characteristics. A fuzzy clustering algorithm is presented considering “super fuzzy” vector and window coefficient, some functions and formulas from this algorithm can recognize many kinds of image’s defects regions with high accuracy, high speed, broad application, and no pre-training. REFERENCES [1] D.r. Rohrmus, “Invariant and adaptive geometrical texture features for defect detection and classification”, Pattern Recognition, vol.38, pp.1546-1599, Feb 2003. [2] Lu Yun, Zhang Jingmiao and Jiang Jianwei, “Fabric Defect
© 2010 ACADEMY PUBLISHER
187
Detection Method Based on Image Distance Difference”, Microcomputer Information, vol.23, pp.303-306, Apri 2007. [3] Wang Feng, Jiao Guotai and Du Ye, “Method of Fabric Defects Detection Based on Mathematical Morphology”, Journal of Test and Measurement Technology, vol.21,pp.515-518, Jun 2007. [4] F. Tajeripour, E. Kabir, and A. Sheikhi, “Fabric Defect Detection Using Modified Local Binary Patterns”, EURASIP Journal on Advances in Signal Processing, vol.1,pp.1-12,Jan 2008 [5] Zou Chao, Zhu desen and Xiao Li, “Textural defect detection based on label co-occurrence matrixm”, Journal of Huazhong University of Science and Technology(Nature Science), vol.34,pp.25-28,Jun 2006. [6] Zhang Y F and Bresee R R., “Fabric defect detection and classification using image analysis”, Textile Research Journal,vol.65,pp.1-9,Jan 1995. [7] Chen Pei-wen, “Classifying textile faults with a backpropagation neural network using power spectral”, Textile Research Journal, vol.68, pp.121-126, Feb 1998. [8] Choonjong Kwak, Jose A.Ventura and Karim Tofang-sazi, “A neural network approach for defect identification and classification on leather fabric”, Journal of Intelligent Manufacturing, vol.11,pp.485-499,May 2000. [9] Liu Jianli and Zuo Baoqi, “Identification of fabric defects based on discrete wavelet transform and back-propagation neural network”, Journal of the Textile Institute, vol.98,pp.355-362,Apri 2007. [10] Shuang-Wu Zhu, Hong-Yang Hao, Peng-Yang Li et al, “Fabric Defects Segmentation Approach Based on Texture Primitive”, 2007 International Conference on Machine Learning and Cybernetics, vol.3, pp.1596-1600, Aug 2007,. [11] Arunkumar Gururajan, Hamed Sari-sarraf and Eric F. Hequet, “Statistical approach to unsupervised defect detection and multiscale localization in two-texture images”, Optical Engineering, vol.47,pp.027202(1)-027202(10),Feb 2008. [12] Chan CH and Grantham KHP, “Fabric defect detection by fourier analysis”, IEEE Transaction on Industry Application, vol.36, pp.1267-1276, May 2000. [13] Wood E J., “Applying fouier and associated transforms to pattern characterization in textiles”, Textile Res. J., vol.60, pp.212-222, Apri 1990. [14] Ravandi S A H,Toriumi K., “Fourier transform analysis of plain woven fabric appearance”, Textile Res. J., vol.65,pp.678-683, Jul 1995. [15] Bodnarova, Bennamoun and S.Latham, “Optimal Gabor filters for textile flaw detection”, Pattern Recognition,vol.35,pp.2973-2991,Nov 2002. [16] K.L. Mak and P. Peng, “An automated inspection system for textile fabrics based on Gabor filters”, Robotics and Computer-Integrated Manufacturing, vol.24,pp.359-369,Jan 2008. [17] Bodnarova A,Bennamoun M and Latham S., “Optimal gabor filters for textile flaw detection”, Pattern Recognition, vol.35, pp.2973-2991, Jan 2002. [18] Aruvazhagan S and Ganesan L., “Texture classification using wavelet transform”, Pattern Recognition letters,
188
vol.24,pp.1513-1521,Dec 2003. [19] Jasper W J,Garnier S J and Potlapalli H, “Texture characterization and defect detection using adaptive wavelets”, Optical Engineering, vol.35,pp.3140-3149,Nov 1996. [20] Portilla J and Simoncelli E.P.A, “Parametric TextureModel based on Joint Statistics of Complex Wavelet Coefficients”, International Journal of Computer Vision, vol.40, pp.49-61, Jan 2002. [21] Soo Chang Kim and Tae Jin Kang, “Texture classification and segmentation using incomplete tree structured wavelet packet frame and Gaussian mixture model”, Imaging Systems and Techniques, vol.13, pp.46-51, May 2005. [22] Shen Jianqiang,Geng Zhaofeng and Zou Xuan, “Detection Method for Fabric Texture Direction Based on Wavelet Transform”, Computer Engineering, vol.33, pp.182-184, Jan 2007. [23] W. Jasper, J.Joines, and J. brenzovich, “Fabric defect detection using a genetic algorithm tuned wavelet filter”, Journal of the Textile Institute, vol.96, pp.43-54, Jan 2005. [24]Chang-Chiun Huang and Tong-Fu Lin. “Image Inspection of Nonwoven Defects Using Wavelet Transforms and Neural Networks”. Fibers and Polymers, vol.9, pp.633-638, May 2008 [25] Ozdemir D and Akarun L., “A fuzzy algorithm for color quantization of images”, Pattern Recognition, vol.35, pp.1785-1791, Oct 2002. [26] Menard M, CourboulaY V and DardignaC P., “Possibilistic and probabilistic fuzzy clustering: Unification within the framework of the non-extensive thermostatistics”, Pattern Recognition, vol.36,pp.1325-1342,Jun 2003. [27] Kaymak U and Setne M., “Fuzzy clustering with volume prototypes and adaptive cluster merging”, IEEE Trans. on Fuzzy Systems, vol.10,pp.706-712, Jun 2002. [28] Schneider A., “Weighted possibilistic clustering algorithms”, Proc. of the 9th IEEE Int’l Conf. on Fuzzy Systems, vol.1,pp.176-180, 2000. [29] Liu SH and Lin JS, “Vector quantization in DCT domain using fuzzy possibilistic c-means based on penalized and compensated constraints”, Pattern Recognition, vol.35,pp.2201-2211, Oct 2002. [30] Zhang JS and Leung YW, “Improved possibilistic c-means clustering algorithms”, IEEE Trans. on Fuzzy Systems, vol.12, pp.209-217, Feb 2004. [31] Ichihashi H, Miyagishi K and Honda K, “Fuzzy C-means clustering with regularization by K-L information”, Proc. of the 10th IEEE Int’l Conf. on Fuzzy Systems, vol.2,pp.924-927, Aug 2001. [32] Yasuda M, Furuhashi T, Matsuzaki m and Okuma S, “A study on statistical mechanical characeteristics of fuzzy clustering”, Proc. Of IEEE Int’l Conf. on Systems, Man and Cybemetics,pp.2415-2420, Apri 2001. [33] Wei C and Fahn C, “The multisynapse neural network and its application to fuzzy clustering”, IEEE Trans. on Neural Networks, vol.13, pp.600-618, Mar 2002. [34] Ma Jun and Hao Lu, “An Optimal Algorithm for Fuzzy Classification Problem”, Journal of Software, vol.12,pp.578-581,Apri 2001.
© 2010 ACADEMY PUBLISHER
JOURNAL OF MULTIMEDIA, VOL. 5, NO. 2, APRIL 2010
[35] Wang JH, Shen Z and Hu YF, “An applicable and efficient clustering algorithm”, Journal of Software, vol.15,pp.697-705, May 2004. Zhe LIU was born in China, March 1972. He is associate professor in Zhongyuan University of Technology and is pursuing Doctor Degree in Tanjin Polytechnic University now. He is working on the theoretic and application research of computer vision all along. During these years, more than 10 high quality papers he wrote had been published, and five scientific research projects of the Province he masterminded have been finished. He also has obtained two national patented inventions. Some papers and research projects has won the government prizes. Xiuchen WANG was born in China, October 1972. She has obtained the master degree from Xian Technology University. She is an associate professor of Zhongyuan University of technology now. Her research interest is computer application and pattern recognition. During these years, she has delivered more than ten important papers, and has finished three scientific research projects from the Province. Some papers and research projects have won the government prizes.
