You are on page 1of 11

Pattern Recognition Letters 32 (2011) 392402

Contents lists available at ScienceDirect

Pattern Recognition Letters


journal homepage: www.elsevier.com/locate/patrec

Unsupervised range-constrained thresholding


Zuoyong Li a,, Jian Yang b, Guanghai Liu c, Yong Cheng d, Chuancai Liu b
a

Department of Computer Science, Minjiang University, Fuzhou 350108, China School of Computer Science and Technology, Nanjing University of Science and Technology, Nanjing 210094, China c School of Computer Science and Information Technology, Guangxi Normal University, Guilin 541004, China d School of Communication Engineering, Nanjing Institute of Technology, Nanjing 211167, China
b

a r t i c l e

i n f o

a b s t r a c t
Three range-constrained thresholding methods are proposed in the light of human visual perception. The new methods rst implement gray level range-estimation, using image statistical characteristics in the light of human visual perception. An image transformation is followed by virtue of estimated ranges. Criteria of conventional thresholding approaches are then applied to the transformed image for threshold selection. The key issue in the process lies in image transformation which is based on unsupervised estimation for gray level ranges of object and background. The transformation process takes advantage of properties of human visual perception and simplies an original image, which is helpful for image thresholding. Three new methods were compared with their counterparts on a variety of images including nondestructive testing ones, and the experimental results show its effectiveness. 2010 Elsevier B.V. All rights reserved.

Article history: Received 5 November 2009 Available online 26 September 2010 Communicated by Y.J. Zhang Keywords: Thresholding Image segmentation Human visual perception Standard deviation Unsupervised estimation

1. Introduction Image segmentation intends to extract an object from a background based on some pertinent characteristics such as gray level, color, texture and location (Tao et al., 2008). It is a critical preprocessing step in image analysis and computer vision (Huang and Wang, 2009; Sen and Pal, 2010). Among the existing segmentation techniques, thresholding is one of the most popular approaches in terms of simplicity, robustness and accuracy. Implicit assumption in image thresholding is that object (foreground) and background have distinctive gray levels. Thresholding serves a variety of applications, such as biomedical image analysis (Hu et al., 2006; Min and Park, 2009), character identication (Huang et al., 2008; Nomura et al., 2009; Pai et al., 2010) and industrial inspection (Ng, 2006). Thresholding techniques fall into bilevel and multilevel category (Coudray et al., 2010; Horng, 2010; Malyszko and Stepaniuk, 2010; Wang et al., 2010) according to the number of segments. The former assumes an image to be composed of two components (i.e., object and background), with an aim of nding an appropriate threshold for distinguishing both parts. Thresholding result is a binary image where all pixels with gray levels higher than determined threshold are classied into foreground and the rest of pixels assigned to background, or vice versa. The latter category supposes that an image consists of multiple parts, each having homogeneous gray level. Obviously, multiple thresholds should
Corresponding author. Tel.: +86 13906926400; fax: +86 059183761607.
E-mail addresses: fzulzytdq@126.com, fzulzytdq@yahoo.com.cn (Z. Li). 0167-8655/$ - see front matter 2010 Elsevier B.V. All rights reserved. doi:10.1016/j.patrec.2010.09.020

be chosen to group pixels with gray level within a specied range into one class. It can be regarded as an extension of the bilevel one. Thresholding can also be classied into parametric and nonparametric approaches from another perspective (Bazi et al., 2007; Sahoo and Arora, 2004; Tizhoosh, 2005). In the parametric approach, gray level distribution of an image is assumed to obey a given statistical model, and optimal parameter estimation for the model is sought by using image histogram. Fitted model is used to approximate practical distribution. Bottom of valley in the model is regarded as the appropriate location of the optimal threshold. This usually involves a nonlinear estimation of intensive computation. The nonparametric method determines the optimal threshold by optimizing certain criterion, such as between-class variance (Otsu, 1979), variance (Hou et al., 2006) and entropy (Kapur et al., 1985; Pun, 1980). The nonparametric approach is proved to be more robust and accurate. Many thresholding approaches have been developed over the last few years (Albuquerque et al., 2004; Kwon, 2004; Ramesh et al., 1995; Sahoo et al., 1988; Wang et al., 2008). For example, Bazi et al. (2007) proposed a parametric method, which nds the optimal threshold through parameter estimation based on the assumption that object and background follow a generalized Gaussian distribution. Otsus method (1979) chooses the threshold by maximizing the between-class variance of both object and background. Sahoo et al. (1988) revealed that Otsus method is one of the better threshold selection approaches for general real world images with regard to uniformity and shape measures. However, Otsus method exhibits a weakness of tending to classify

Z. Li et al. / Pattern Recognition Letters 32 (2011) 392402

393

an image into two parts of similar size regardless of the practical situation. After exploring the underlying reason for Otsus weakness, Hou et al. (2006) presented an approach based on minimum class variance, which can be regarded as a generalization of Otsus. In image thresholding, one of the most efcient techniques is entropy-based approach, which regards a gray level image histogram as a probability distribution. In Puns method (1980), the threshold is determined by maximizing the posteriori entropy of object and background. Kapur et al. (1985) found some aws in Puns derivations and proposed a corrected version. In 2004, Albuquerque et al. presented an approach based on Tsallis entropy. Tsallis entropy is applied as a general entropy form for information theory. In addition, Rameshs method (1995) nds the threshold by minimizing function approximation error for image histogram with a bilevel function. Wang et al. (2008) determined the threshold by optimizing a criterion function deduced by image histogram and Parzen window technique. A thorough survey over thresholding is provided in the literature (Sezgin and Sankur, 2004). Conventional nonparametric approaches utilize some criteria to search the optimal threshold from all the gray levels, and neglect properties of human visual perception. This makes them suffer a limitation of being unable to obtain satisfactory results when segmenting some images. To eliminate the above limitation, three unsupervised range-constrained thresholding methods are presented in this paper. Based on properties of human visual perception, they rst nd gray level ranges of object and background in an unsupervised way by utilizing statistical characteristics of an image. Then, an image transformation is implemented via the above ranges. The transformation should simplify an original image and improve segmentation performance. Finally, three existing criteria are applied to a transformed image for threshold determination. The performance of range-constrained methods is compared with their conventional counterparts by testing a variety of images. Experimental results show their superiority over the counterparts. The remainder of this paper is organized as follows: Section 2 introduces properties of human visual perception. Three unsupervised range-constrained methods are proposed in Section 3. Their performance is detailed on a variety of images and compared with their counterparts in Section 4. Conclusions appear in Section 5. 2. Human visual perception Human visual perception has the following properties as described in the literature (Arora et al., 2008). (1) Human eye is insensitive to features present at the both extremes of pixel intensity, whereas sensitive to distinguishing features at the mid-range intensities. This characteristic suggests a focus upon mid-region of a gray scale image, i.e., around the mean, when segmenting images. (2) A lot of images may have either histograms with high intensity values or more structures near a certain value (usually the mean) than that farther from the mean. A rough estimation of such a histogram exhibits a Gaussian distribution. 3. Unsupervised range-constrained thresholding methods Conventional nonparametric thresholding approaches nd the optimal threshold via optimizing some criteria. The process neglects the properties of human visual perception, resulting in unsatisfactory segmentation for some real world images. In an effort to eliminate the limitation, the authors have tried a new scheme in the light of human visual perception. The scheme rst nds gray level ranges of object and background via human visual perception and image statistical characteristics in an unsupervised

way, then simplies an original image by image transformation based on the ranges, and nally applies the criteria of three existing methods to the transformed image, thus forming rangeconstrained approaches with better segmentation performance. 3.1. Gray level ranges of object and background For the sake of description convenience, we assume that object (foreground) pixels have higher gray levels than background ones. In the light of human visual perception, two gray levels should be chosen as the lower bound for object and the upper bound for background by using statistical characteristics of an image. After nding the lower and upper bounds, gray level ranges of object and background are accordingly determined, respectively. The detailed process is as follows: (1) Without losing generality, Let I be a gray scale image with L levels [0, 1, . . . , L 1]. The number of pixels with gray level i is denoted by ni and the total number of pixels by N = n0 + n1 + + nL1. The mean and standard deviation of the image are dened as

L1 1X ini ; N i0

1 !1 2 : 2

L1 1 X i l2 ni N 1 i0

(2) Compute two gray levels

T u l b r; T l l b r;

3 4

where b is a parameter and its value can be automatically determined by optimizing the proposed criterion in Eq. (13). Tu and Tl are the upper and lower bounds for background and object, respectively. (3) Determine gray level ranges of object and background via the following way:

RF T l L 1; RB 0 T u :

5 6

Finding the upper and lower bounds needs choosing a reasonable value for parameter b, which in turn involves a statistical criterion to be dened. Assuming that the image I is divided into three portions via two gray levels t1 and t2, where t1 < t2, with three parts denoted by Cb, Cf and Cm, where Cb is the background class with gray levels [0, . . . , t1 1], Cf the foreground class with levels [t2 + 1, . . . , L 1], and Cm the middle class with levels [t1, . . . , t2]. Cm is the middle transition region between Cb and Cf as described in (Hu et al., 2008). The mean of each class is dened as

lb lf

t1 1 1 X ini ; Nb i0 L1 1 X ini ; Nf it 1
2

lm

t2 1 X ini ; Nm it
1

where Nb, Nf and Nm are the numbers of pixels in Cb, Cf and Cm, respectively. And their respective standard deviation can be given by

394
t 1 1 X 1 i lb 2 ni Nb 1 i0

Z. Li et al. / Pattern Recognition Letters 32 (2011) 392402

!1 2 ; !1 2 ; !1 2 : 12 11 10

rb

(3) Calculate the value of our statistical criterion rS by Eq. (13). If rS < MVS, then MVS = rS, b = 0.1 i and i = i + 1; else, i = i + 1. Subsequently, return to step 2. Once b is determined, the upper and lower bounds can be calculated by Eqs. (3) and (4). Take cell image in Fig. 1(a) as an example. Its foreground, background and middle classes are shown in Fig. 1(b)(d), where bright pixels are our focuses. The upper and lower bounds are 222 and 227, and b is automatically set to 0.1. Accordingly, gray level ranges of the background and object are [0 222) and (227 L 1], respectively. 3.2. Image transformation In order to implement image transformation, gray level ranges of object and background must rst be obtained as described in Section 3.1. The ranges can be treated as gray level constrains on object and background. For the image I, the process of image transformation is as follows: (1) Calculate the upper and lower bounds by Eqs. (3) and (4). (2) Obtain gray level ranges of object and background, i.e., RF and RB, by Eqs. (5) and (6). (3) Implement image transformation via the following way:

rf

L1 X 1 i lf 2 ni Nf 1 it 1
2

rm

t2 X 1 i lm 2 ni Nm 1 it
1

Based on the above standard deviations, the statistical criterion could be dened as

rS a rb rf 1 a rm ;

13

where a is a parameter between 0 and 1. The criterion consists of two terms, the rst term standing for the sum of standard deviations corresponding to Cb and Cf. Standard deviation is a common statistical measure reecting degree of deviations between mean and individuals. Hence, the term could represent intra-class similarities of the background and foreground to some extent. The smaller the term is, the higher the similarities. But it is worth mentioning that both background and foreground themselves of a practical image usually have some deviations on pixels gray levels, especially for background. So it is unreasonable to determine the value of b by minimizing only the rst term. To solve the problem, rm, standard deviation of the transitional class, Cm, is introduced into the criterion as a penalty term. The parameter a in Eq. (13) is a weight balancing the contributions of the two terms. For the image I, following steps describe the detail about automatic selection of b: (1) Initialize MVS to be innite, and i = 1, where MVS is the minimum value of rS, i is the temporal number of iterations. (2) Compute two gray levels t1 and t2 via the following equations:

8 if f i; j 2 RB ; > Tu < ftr i; j T l if f i; j 2 RF ; > : f i; j otherwise;

16

where f(i, j) and ftr(i, j) are gray levels at pixel (i, j) of the original image and the transformed form, respectively. The transformation weakens gray level changes in both object and background simultaneously, thus simplifying the original image. The weakening effect is favorable to subsequent image segmentation. Take transformed form Fig. 2(b) for the cell image as an example. Their respective histogram is displayed in Fig. 2(c) and (d). From Fig. 2, one can conclude that gray level changes of object and background have been weakened dramatically, and the transformed image becomes much simpler than the original. Furthermore, one can observe that the transformation turns a histogram of unimodal distribution into an apparent bimodal one, the latter being preferable when detecting the valley in a histogram, for valley is usually regarded as the appropriate location of the optimal threshold. 3.3. Method 1: Range-Constrained Rameshs method (RCramesh) RCramesh recursively approximates the histogram of a given transformed image with a bilevel function, and nds the optimal threshold by minimizing approximation error. Here, the error is represented by the variance of the approximated histogram. It is worth mentioning that gray level range of the transformed image

t1 l 0:1 i r; t2 l 0:1 i r;

14 15

where l and r are dened in Eqs. (1) and (2), respectively. If t1 < 0 or t2 > L 1, then break the process; else, continue step 3. Here, the reason using a constant 0.1 in Eqs. (14) and (15) is as follows: image thresholding includes an implicit assumption that object and background have distinctive gray levels, that is, their difference should be at least several gray levels. Standard deviation of an image is usually between 20 and 100, with practical values for our sample images in experiments being as: Cell (24.532), Tile (20.829), PCB (54.341), Gearwheel (100.05), Potatoes (83.056), Block (73.122), Lena (47.51), Peppers (53.179), Flower (41.912) and Corn (62.083). Therefore, 0.1 is basically the minimum coefcient for standard deviation r in Eqs. (14) and (15).

Fig. 1. Cell image and its classes: (a) original, (b) foreground class, (c) background class, (d) middle class.

Z. Li et al. / Pattern Recognition Letters 32 (2011) 392402

395

Fig. 2. Images and histograms: (a) original cell image, (b) the transformed image, (c) histogram of (a), (d) histogram of (b).

should be [Tu Tl], where Tu and Tl are the upper and lower bounds. For a given gray level Tu 6 t 6 Tl, the approximation error in RCramesh can be formulated as

Et
where

Tl t X X 1 1 i l1 2 i l2 2 ; N1 1 iT N 2 1 it1
u

17

total number of pixels by N N T u N T u 1 N T l . Suppose that the pixels in the transformed image are divided into two classes, A and B, by a level t (Tu 6 t 6 Tl). A is the set of pixels with levels [Tu, . . . , t], and B the set of pixels with levels [t + 1, . . . , Tl]. The two sets constitute the corresponding probability spaces, and their probability density function, pA(x, y) and pB(x, y), can be formulated as follows by using Parzen window estimation:

l1

t 1 X ini ; N1 iT
u

18

pA x; y

Ni t 1 XX 1 ux; y; xj ; yj ; h2 i ; N N iT j1 h2 Ni u Tl Ni 1 X X 1 ux; y; xj ; yj ; h2 i ; N N it1 j1 h2 Ni

26

l2

Tl 1 X ini ; N2 it1

19

pB x; y
where

27

where N1 is the number of pixels with levels [Tu, . . . , t], and N2 the number of pixels with levels [t + 1, . . . , Tl]. The optimal threshold t* can be determined by RCramesh,

ux; y; xj ; yj ; h2 i N

! x xj 2 y yj 2 1 : exp 2 2p 2hN
i

28

T Arg min fEtg:


T u 6t6T l

20

3.4. Method 2: Range-Constrained Tsallis method (RCtsallis) For a transformed image, let pi be the probability of gray level i appeared in the image, where Tu 6 i 6 Tl. Assuming that the pixels in the image are classied into two classes, A and B, by a gray level t, where Tu 6 t 6 Tl. Class A corresponds to the foreground and class B to the background, or vice versa. Cumulative probability of each class can be dened as

In Eqs. (26)(28), hNi denotes the window width of the set composed of those pixels with gray level i, (xj, yj) is the coordinate of the jth pixel in the set. And the optimal threshold t* can be determined by RCwang,

ZZ ZZ t Arg min p2 x; ydx dy A T u 6t6T i X  ZZ pA x; ypB x; ydx dy ; 2


X

p2 x; ydx dy B 29

xA

t X iT u

where X = [1, m] [1, n], m and n are the height and width of the image, respectively. And the following equations hold:

pi ;

21

ZZ
X

p2 x; ydx dy A

Ni Nk t t 1 XXXX

N2

Gxj ; yj ; xr ; yr ; hNi hNk ; 30

iT u j1 kT u r1

xB

Tl X it1

pi :

22 ZZ
X
iT u pi =

A priori Tsallis entropy for each class is dened as

p2 x; ydx dy B

Tl Tl Nk Ni 1 X X X X

SA t q

Pt

N2

Gxj ; yj ; xr ; yr ; hNi hNk ; 31

it1 j1 kt1 r1

xA q xB

q1 1 PT l
it1 pi =

23
ZZ pA x;ypB x;ydxdy
X
Tl Nk Ni t 1 XX X X 2

SB t q

q1

24

N2 iT u

Gxj ;yj ;xr ;yr ;hNi hNk ; 32

j1 kt1 r1

And the optimal threshold t* can be determined by RCtsallis,

where

T Arg max fSA t SB t 1 qSA tSB tg: q q q q


T u 6t6T l

25

Gxj ; yj ; xr ; yr ; hNi hNk

1
2 hNi

hNk

uxj ; yj ; xr ; yr ; h2 i h2 k ; N N

33

3.5. Method 3: Range-Constrained Wangs method (RCwang) For a transformed image with gray levels [Tu, Tu + 1, . . . , Tl], the number of pixels with level i (Tu 6 i 6 Tl) is denoted by ni and the

and the denition of function u can refer to Eq. (28). The advantages of RCramesh, RCtsallis and RCwang over their counterparts are as follows: (1) a transformed image is used instead of the original one during thresholding, which transfers

396

Z. Li et al. / Pattern Recognition Letters 32 (2011) 392402 Table 1 Thresholds, numbers of misclassied pixels, ME values and running times obtained by applying various methods to the NDT images. NDT images Thresholding methods Rameshs method Cell Threshold Misclassied pixels ME Running times (s) Tile Threshold Misclassied pixels ME Running times (s) PCB Threshold Misclassied pixels ME Running times (s) 241 53,861 0.82185 1.641 213 46,719 0.71288 1.375 173 15,707 0.23967 0.359 RCramesh 222 2634 0.040192 1.844 197 1711 0.026108 1.563 79 2655 0.040512 0.625 Tsallis 171 7375 0.11253 3.469 149 10229 0.15608 3.406 161 14,426 0.22012 0.375 RCtsallis 226 4172 0.06366 3.672 200 2703 0.041245 3.594 80 2501 0.038162 0.641 Wangs method 171 7375 0.11253 281.27 149 10229 0.15608 265.69 161 14,426 0.22012 203.59 RCwang 225 3780 0.057678 281.47 199 2257 0.034439 265.88 80 2501 0.038162 203.86

Fig. 3. Thresholding results of cell image: (a) original, (b) ground truth image, (c) Rameshs method (t = 241), (d) RCramesh (t = 222), (e) Tsallis (t = 171), (f) RCtsallis (t = 226), (g) Wangs method (t = 171), (h) RCwang (t = 225).

Fig. 4. Thresholding results of tile image: (a) original, (b) ground truth image, (c) Rameshs method (t = 213), (d) RCramesh (t = 197), (e) Tsallis (t = 149), (f) RCtsallis (t = 200), (g) Wangs method (t = 149), (h) RCwang (t = 199).

Z. Li et al. / Pattern Recognition Letters 32 (2011) 392402

397

Fig. 5. Thresholding results of PCB image: (a) original, (b) ground truth image, (c) Rameshs method (t = 173), (d) RCramesh (t = 79), (e) Tsallis (t = 161), (f) RCtsallis (t = 80), (g) Wangs method (t = 161), (h) RCwang (t = 80).

Table 2 Thresholds, numbers of misclassied pixels, ME values and running times obtained by applying various methods to the simple real world images. Simple images Thresholding methods Rameshs method Gearwheel Threshold Misclassied pixels ME Running times (s) Potatoes Threshold Misclassied pixels ME Running times (s) Block Threshold Misclassied pixels ME Running times (s) 228 26,829 0.40938 0.5 143 1805 0.027542 0.609 126 12,554 0.19156 0.437 RCramesh 84 106 0.0016174 0.672 97 0 0 0.812 52 428 0.0065308 0.593 Tsallis 13 2247 0.034286 1.094 64 1823 0.027817 0.734 17 2589 0.039505 0.781 RCtsallis 85 131 0.0019989 1.266 98 38 0.00057983 0.937 53 378 0.0057678 0.937 Wangs method 202 12,664 0.19324 262.34 65 1593 0.024307 189.13 128 13,910 0.21225 232.56 RCwang 85 131 0.0019989 262.52 98 38 0.00057983 189.33 53 378 0.0057678 232.72

Fig. 6. Thresholding results of gearwheel image: (a) original, (b) ground truth image, (c) Rameshs method (t = 228), (d) RCramesh (t = 84), (e) Tsallis (t = 13), (f) RCtsallis (t = 85), (g) Wangs method (t = 202), (h) RCwang (t = 85).

398

Z. Li et al. / Pattern Recognition Letters 32 (2011) 392402

range of threshold selection from [0 L 1] to [Tu Tl]. This coincides with the properties of human visual perception and improves segmentation performance, (2) image transformation in the proposed approaches weakens gray level changes in object and background, thus simplifying the original image, which is helpful for subsequent image thresholding.

signed to background. For a two-class segmentation, ME can be simply formulated as

ME 1

jBO \ BT j jF O \ F T j ; jBO j jF O j

34

4. Experimental results To evaluate the performance of our range-constrained methods, a variety of images have been chosen as testing samples. The results yielded by our methods were compared with those obtained by their counterparts, i.e., Rameshs method (1995), Tsallis (Albuquerque et al., 2004) and Wangs method (2008). The quality of segmentation result is quantitatively evaluated via misclassication error (ME) measure (Yasnoff et al., 1977), which regards image segmentation as a pixel classication process. The measure reects the percentage of background pixels erroneously classied into foreground, and conversely, foreground pixels erroneously as-

where BO and FO are the background and foreground of the ground truth image, BT and FT the background and foreground pixels in the thresholded image, and || the cardinality of a set. The value of ME varies between 0 for a perfectly classied image and 1 for a totally erroneously classied one. A lower value of ME means better quality. In Tsallis and RCtsallis, q = 3. In our methods, a = 0.4. All experiments are performed on a notebook PC with 2.13G Intel Core 2 Duo CPU and 3G RAM. All the images used in the experiments are of 256 256 pixels and 8-bit (i.e., 256 gray levels). 4.1. Experiments on NDT images The range-constrained methods were rst applied to three NDT images and compared with their counterparts. NDT means to detect an object and quantify its possible defects without harmful

Fig. 7. Thresholding results of potatoes image: (a) original, (b) ground truth image, (c) Rameshs method (t = 143), (d) RCramesh (t = 97), (e) Tsallis (t = 64), (f) RCtsallis (t = 98), (g) Wangs method (t = 65), (h) RCwang (t = 98).

Fig. 8. Thresholding results of block image: (a) original, (b) ground truth image, (c) Rameshs method (t = 126), (d) RCramesh (t = 52), (e) Tsallis (t = 17), (f) RCtsallis (t = 53), (g) Wangs method (t = 128), (h) RCwang (t = 53).

Z. Li et al. / Pattern Recognition Letters 32 (2011) 392402

399

effects on itself by special equipments and methods. It is used in a broad variety of applications, such as aeronautics and astronautics, nuclear industry, chemistry and civil constructions. The results in terms of thresholds, numbers of misclassied pixels, ME values and running times for various approaches are listed in Table 1. The table shows that segmentation results obtained by authors methods have less misclassied pixels and lower ME values, implying better performance. This could be attributed to the utilization of human visual perception in nding ranges of object and background for subsequent image transformation. The transformation dramatically simplies the original image by weakening gray level changes of object and background. The coincidence with human visual perception and the image simplication are helpful for improving segmentation performance. In addition, one can observe that conventional methods are only slightly faster than our approaches with regard to the speed of segmentation. The rangeconstrained methods reduce search space during thresholding from the whole gray levels of an original image to a much smaller range [Tu Tl] and save running time, as compared with conventional approaches. Nevertheless, the new methods need extra time to

estimate ranges of object and background, with implementing image transformation. Performance judgment over various methods can also be evidenced by visual segmentation results shown in Figs. 35. From the gures, one can conclude that the proposed methods enjoy better visual effects, as they segment the objects more accurately. 4.2. Experiments on other images In this section, seven general real world images were chosen as test specimen. These images fall into two groups. The rst is for simple images, and the second for complex ones. Table 2 lists quantitative comparison of segmentation results for the rst group. The data show that the proposed methods achieve better segmentations with less misclassied pixels and lower ME values. Visual thresholding results are displayed in Figs. 68. The gures indicate, in addition to a completely segmenting for objects, the new approaches exhibit less background noise. Segmentation results for the complex images are in Figs. 912. The rst two are classic gray level types, and the rest come from

Fig. 9. Thresholding results of Lena image: (a) original, (b) histogram, (c) Rameshs method (t = 107), (d) RCramesh (t = 119), (e) Tsallis (t = 161), (f) RCtsallis (t = 127), (g) Wangs method (t = 122), (h) RCwang (t = 126).

Fig. 10. Thresholding results of peppers image: (a) original, (b) histogram, (c) Rameshs method (t = 131), (d) RCramesh (t = 124), (e) Tsallis (t = 70), (f) RCtsallis (t = 128), (g) Wangs method (t = 95), (h) RCwang (t = 128).

400

Z. Li et al. / Pattern Recognition Letters 32 (2011) 392402

famous Berkeley image database. Due to their complex structures, quantitative measurement of segmentation quality experiences serious difculty at present. Only visual perception of subjective nature is applicable to the judgment of segmentation quality. Figs. 912 show range-constrained methods preserving more details of the objects, implying better results. More details relating ranges of object and background for all the images in our experiments are provided in Table 3.

4.3. Parameter selection The range-constrained methods involve two parameters a and b, and b is determined by our proposed statistical criterion with a. Hence, only one parameter a is left uncertain. The parameter is used to balance the contributions of the two terms in our criterion, and smaller value implies larger contribution of the transitional class. To nd the reasonable value for a, a series of

Fig. 11. Thresholding results of ower image: (a) original, (b) histogram, (c) Rameshs method (t = 225), (d) RCramesh (t = 54), (e) Tsallis (t = 123), (f) RCtsallis (t = 55), (g) Wangs method (t = 122), (h) RCwang (t = 55).

Fig. 12. Thresholding results of corn image: (a) original, (b) histogram, (c) Rameshs method (t = 155), (d) RCramesh (t = 100), (e) Tsallis (t = 138), (f) RCtsallis (t = 101), (g) Wangs method (t = 131), (h) RCwang (t = 101).

Table 3 Ranges of object and background obtained by the proposed methods for image transformation under the assumption that object pixels have higher gray levels than background ones. Images Cell Tile PCB Gearwheel Potatoes Range of object (227 255] (201 255] (90 255] (104 255] (113 255] Range of background [0 222) [0 197) [0 79) [0 84) [0 97) Images Block Lena Peppers Flower Corn Range of object (67 255] (128 255] (135 255] (63 255] (112 255] Range of background [0 52) [0 119) [0 124) [0 54) [0 100)

Z. Li et al. / Pattern Recognition Letters 32 (2011) 392402 Table 4 ME and MME (Mean Misclassication Error) values obtained by using three range-constrained methods with different a. Images RCramesh Cell Tile PCB Gearwheel Potatoes Block MME RCtsallis Cell Tile PCB Gearwheel Potatoes Block MME RCwang Cell Tile PCB Gearwheel Potatoes Block MME 0.1 0.04019 0.02611 0.04051 0.00162 0.00000 0.00653 0.01916 0.06366 0.04125 0.03816 0.00200 0.00058 0.00577 0.02524 0.05768 0.03444 0.03816 0.00200 0.00058 0.00577 0.02310 0.2 0.04019 0.02611 0.04051 0.00162 0.00000 0.00653 0.01916 0.06366 0.04125 0.03816 0.00200 0.00058 0.00577 0.02524 0.05768 0.03444 0.03816 0.00200 0.00058 0.00577 0.02310 0.02250 0.3 0.04019 0.02611 0.04051 0.00162 0.00000 0.00653 0.01916 0.06366 0.04125 0.03816 0.00200 0.00058 0.00577 0.02524 0.05768 0.03444 0.03816 0.00200 0.00058 0.00577 0.02310 0.02250 0.4 0.04019 0.02611 0.04051 0.00162 0.00000 0.00653 0.01916 0.06366 0.04125 0.03816 0.00200 0.00058 0.00577 0.02524 0.05768 0.03444 0.03816 0.00200 0.00058 0.00577 0.02310 0.02250 0.5 0.02969 0.02611 0.04051 0.00162 0.00000 0.00653 0.01741 0.81273 0.04125 0.03816 0.00200 0.00058 0.00577 0.15008 0.12642 0.03444 0.03816 0.00200 0.00058 0.00577 0.03456 0.06735 0.6 0.02969 0.05739 0.66911 0.00162 0.54103 0.00653 0.21756 0.81273 0.35802 0.66072 0.00200 0.16295 0.00577 0.33370 0.12642 0.12059 0.11636 0.00200 0.33977 0.00577 0.11848 0.22325 0.7 0.02969 0.05739 0.66911 0.03326 0.60266 0.02322 0.23589 0.81273 0.35802 0.66072 0.03253 0.35788 0.02237 0.37404 0.12642 0.12059 0.11636 0.03253 0.38225 0.02237 0.13342 0.24778 0.8 0.02969 0.05739 0.66911 0.04979 0.62694 0.56946 0.33373 0.81273 0.35802 0.66072 0.03536 0.38225 0.43695 0.44767 0.12642 0.12059 0.11636 0.03844 0.38225 0.09178 0.14597 0.30912 0.9 0.02969 0.05739 0.66911 0.04979 0.62694 0.56946 0.33373 0.81273 0.35802 0.66072 0.03536 0.38225 0.43695 0.44767 0.12642 0.12059 0.11636 0.03844 0.38225 0.09178 0.14597 0.30912 1

401

0.02969 0.05739 0.66911 0.04979 0.62694 0.56946 0.33373 0.81273 0.35802 0.66072 0.03536 0.38225 0.43695 0.44767 0.12642 0.12059 0.11636 0.03844 0.38225 0.09178 0.14597 0.30912

Average MME for three methods 0.02250

experiments on six images with different a have been carried out. Experimental results are listed in Table 4. The table indicates that: (1) ME values of images usually vary from small to big. The reason is that a big a determines an unreasonable value for b, which in turn leads to inaccurate gray level ranges of object and background. Such inaccuracy in ranges will signicantly affect image transformation and subsequent image thresholding, (2) with reference to MME values of six images, three range-constrained methods have different optimal values for a. For example, in RCtsallis and RCwang, the optimal value for positions somewhere between 0.1 and 0.4, whereas in RCramesh, the optimal value is 0.5. However, it is worth mentioning that the difference of MME values obtained by RCramesh is small when a varies between 0.1 and 0.5. Therefore, as a general guide, the parameter ranges between 0.1 and 0.4, a 0.4 being selected for our proposed methods. Experimental data reveal a = 0.4 usually yield lowest average MME value for the proposed methods, while a varying between 0.1 and 0.4 leads to the same b as well as transformed image, which brings about segmentation result with the same average MME value. When a higher than 0.4, however, average MME value becomes larger, thus indicating a = 0.4 acting as a turning point.

with human visual perception and simplies an original image, which benets image segmentation. In addition, experimental results show that running times of the proposed methods are comparative with those of their counterparts, and the reason has been interpreted in Section 4.1. Experimental results on a variety of real world images including NDT images show the effectiveness of the proposed range-constrained methods. Acknowledgements This work is supported by National Natural Science Foundation of China (Grant Nos. 60472061, 60632050, 90820004, 60875010), National 863 Project (Grant Nos. 2006AA04Z238, 2006AA01Z119), Technology Project of Provincial University of Fujian Province (JK2010046) and Technology Project of Education Department of Fujian Province (JA10226). References
Albuquerque, M.P., Esquef, I.A., Mello, A.R.G., 2004. Image thresholding using Tsallis entropy. Pattern Recognition Lett. 25 (9), 10591065. Arora, S., Acharya, J., Verma, A., Panigrahi, P.K., 2008. Multilevel thresholding for image segmentation through a fast statistical recursive algorithm. Pattern Recognition Lett. 29 (2), 119125. Bazi, Y., Bruzzone, L., Melgani, F., 2007. Image thresholding based on the EM algorithm and the generalized Gaussian distribution. Pattern Recognition 40 (2), 619634. Coudray, N., Buessler, J.L., Urban, J.P., 2010. Robust threshold estimation for images with unimodal histograms. Pattern Recognition Lett. 31 (9), 10101019. Horng, M.H., 2010. A multilevel image thresholding using the honey bee mating optimization. Appl. Math. Comput. 215 (9), 33023310. Hou, Z., Hu, Q., Nowinski, W.L., 2006. On minimum variance thresholding. Pattern Recognition Lett. 27 (14), 17321743. Hu, Q., Hou, Z., Nowinski, W.L., 2006. Supervised range-constrained thresholding. IEEE Trans. Image Process. 15 (1), 228240. Hu, Q., Luo, S., Qiao, Y., Qian, G., 2008. Supervised grayscale thresholding based on transition regions. Image Vision Comput. 26 (12), 16771684. Huang, S., Ahmadi, M., Ahmed, M.A.S., 2008. A hidden Markov model-based character extraction method. Pattern Recognition 41 (9), 28902900. Huang, D.Y., Wang, C.H., 2009. Optimal multi-level thresholding using a two-stage Otsu optimization approach. Pattern Recognition Lett. 30 (3), 275284.

5. Conclusions Three novel range-constrained thresholding methods are proposed in this paper. In the light of human visual perception, the methods rst use statistical characteristics of an image to estimate gray level ranges of object and background. At this stage, a new statistical criterion for automatic selection of the parameter b is developed. Subsequently, the gray level ranges are accordingly determined. Automatic choice of b enhances the universality of our methods. Then, range-constrained methods utilize the above ranges to implement image transformation. Finally, criteria of three existing thresholding methods are applied to the transformed image for threshold selection. Image transformation by conning gray level ranges of object and background coincides

402

Z. Li et al. / Pattern Recognition Letters 32 (2011) 392402 Ramesh, N., Yoo, J.H., Sethi, I.K., 1995. Thresholding based on histogram approximation. IEE Proc. Vision Image Signal Process. 142 (5), 271279. Sahoo, P.K., Arora, G., 2004. A thresholding method based on two-dimensional Renyis entropy. Pattern Recognition 37 (6), 11491161. Sahoo, P.K., Soltani, S., Wong, A.K.C., 1988. A survey of thresholding techniques. Comput. Vision Graphics Image Process. 41 (2), 233260. Sen, D., Pal, S.K., 2010. Gradient histogram: thresholding in a region of interest for edge detection. Image Vision Comput. 28 (4), 677695. Sezgin, M., Sankur, B., 2004. Survey over image thresholding techniques and quantitative performance evaluation. J. Electron. Imaging 13 (1), 146165. Tao, W., Jin, H., Zhang, Y., Liu, L., Wang, D., 2008. Image thresholding using graph cuts. IEEE Trans. Systems Man Cybernet., Part A 38 (5), 11811195. Tizhoosh, H.R., 2005. Image thresholding using type II fuzzy sets. Pattern Recognition 38 (12), 23632372. Wang, S., Chung, F., Xiong, F., 2008. A novel image thresholding method based on parzen window estimate. Pattern Recognition 41 (1), 117129. Wang, N., Li, X., Chen, X., 2010. Fast three-dimensional Otsu thresholding with shufed frog-leaping algorithm. Pattern Recognition Lett. 31 (13), 1809 1815. Yasnoff, W.A., Mui, J.K., Bacus, J.W., 1977. Error measures for scene segmentation. Pattern Recognition 9 (4), 217231.

Kapur, J.N., Sahoo, P.K., Wong, A.K.C., 1985. A new method for grey-level picture thresholding using the entropy of the histogram. Comput. Vision Graphics Image Process. 29 (3), 273285. Kwon, S.H., 2004. Threshold selection based on cluster analysis. Pattern Recognition Lett. 25 (9), 10451050. Malyszko, D., Stepaniuk, J., 2010. Adaptive multilevel rough entropy evolutionary thresholding. Inf. Sci. 180 (7), 11381158. Min, T.H., Park, R.H., 2009. Eyelid and eyelash detection method in the normalized iris image using the parabolic Hough model and Otsus thresholding method. Pattern Recognition Lett. 30 (12), 11381143. Ng, H.F., 2006. Automatic thresholding for defect detection. Pattern Recognition Lett. 27, 16441649. Nomura, S., Yamanaka, K., Shiose, T., Kawakami, H., Katai, O., 2009. Morphological preprocessing method to thresholding degraded word images. Pattern Recognition Lett. 30 (8), 729744. Otsu, N., 1979. A threshold selection method from gray-level histograms. IEEE Trans. Systems Man Cybernet. 9 (1), 6266. Pai, Y.T., Chang, Y.F., Ruan, S.J., 2010. Adaptive thresholding algorithm: efcient computation technique based on intelligent block detection for degraded document images. Pattern Recognition 43 (9), 31773187. Pun, T., 1980. A new method for grey-level picture thresholding using the entropy of histogram. Signal Process. 2 (3), 223227.

You might also like