Color edges extraction using statistical features and automatic threshold technique: application to the breast cancer cells
 Salim Ben Chaabane^{1}Email author and
 Farhat Fnaiech^{1}
https://doi.org/10.1186/1475925X134
© Ben Chaabane and Fnaiech; licensee BioMed Central Ltd. 2014
Received: 13 September 2013
Accepted: 4 January 2014
Published: 23 January 2014
Abstract
Background
Color image segmentation has been so far applied in many areas; hence, recently many different techniques have been developed and proposed. In the medical imaging area, the image segmentation may be helpful to provide assistance to doctor in order to followup the disease of a certain patient from the breast cancer processed images. The main objective of this work is to rebuild and also to enhance each cell from the three component images provided by an input image. Indeed, from an initial segmentation obtained using the statistical features and histogram threshold techniques, the resulting segmentation may represent accurately the non complete and pasted cells and enhance them. This allows real help to doctors, and consequently, these cells become clear and easy to be counted.
Methods
A novel method for color edges extraction based on statistical features and automatic threshold is presented. The traditional edge detector, based on the first and the second order neighborhood, describing the relationship between the current pixel and its neighbors, is extended to the statistical domain. Hence, color edges in an image are obtained by combining the statistical features and the automatic threshold techniques. Finally, on the obtained color edges with specific primitive color, a combination rule is used to integrate the edge results over the three color components.
Results
Breast cancer cell images were used to evaluate the performance of the proposed method both quantitatively and qualitatively. Hence, a visual and a numerical assessment based on the probability of correct classification (P_{ C }), the false classification (P_{ f }), and the classification accuracy (Sens(%)) are presented and compared with existing techniques. The proposed method shows its superiority in the detection of points which really belong to the cells, and also the facility of counting the number of the processed cells.
Conclusions
Computer simulations highlight that the proposed method substantially enhances the segmented image with smaller error rates better than other existing algorithms under the same settings (patterns and parameters). Moreover, it provides high classification accuracy, reaching the rate of 97.94%. Additionally, the segmentation method may be extended to other medical imaging types having similar properties.
Keywords
Introduction
Image segmentation is considered as a critical and an important basic operation for meaningful analysis and interpretation of image acquired [1]. It is one of the most difficult tasks in image processing, which determines the quality of the final results of analysis [2]. The goal of image segmentation is the decomposition of an image into meaningful or spatially coherent regions sharing similar features [3].
Color image segmentation attracts more and more attention in many pattern recognition and computer vision applications [4]. The additional information provided by color may be of great help in the image analysis process and yield better results than other approaches using only gray scale information. This situation may occur when the information cannot be extracted using only gray scale, then it becomes mandatory to use color information [3, 4].
For instance, typical color image segmentation methods are mainly extended from monochrome segmentation approaches in different color spaces. In other words, gray level segmentation methods are directly applied to each component of a color space, and then the results are combined together to obtain the final segmentation results.
Different color image segmentation techniques have been, so far developed and detailed in the literature [3, 4]. Generally, color image segmentation approaches are based on either discontinuity and/or homogeneity of information values in a region. The approach based on discontinuity tends to segment an image by detecting isolated points, lines and edges based on abrupt changes in gray levels. Typically, the approaches are usually based on homogeneity including threshold, clustering, region growing, region splitting and merging, etc. Consequently, the general color segmentation problem consists in choosing an adopted color model and the segmentation technique for a specific problem [4, 5].
In the last decades; panoply of techniques have been proposed and studied solving the color edge detection problem [5–7]. Basically, these techniques detect the fast change between two regions [8]. Edge detectors used in these techniques turn out simple canny operator [8, 9]. The output of most existing edge detectors can only provide candidates for the regions boundaries, because of the obtained color edges normally are discontinuous or overdetected. Whereas, in principle; the actual region boundaries should be closed curves. Consequently, some postprocedures, such as edge tracking, gap filling, smoothing, and thinning, should be performed in order to obtain the closed region boundaries. All these postprocedures are timeconsuming; and converting the edge candidates to the region boundaries is thus not an easy task.
For instance, these timeconsuming postprocedures may be avoided by integrating the results of the boundarybased approach over those + of the region based approach.
In a digital color image, a typical edge characteristic is defined when two brightness neighbouring pixels values change suddenly. Flowingly, in the literature many robust and complex coloredge detection techniques have been proposed [8, 10, 11]. These sophisticated techniques may provide highly accurate results.
These techniques are divided into two categories of edge detectors: 1) gradient operators and 2) second derivative operators.
Gradient operators, such as the Roberts, Prewitt, and Sobel operators [8], detect the edges by looking for the maximum and minimum in the first derivative of the luminance of an image.
The second derivative operators, such as the MarrHildreth and MarrPoggio operators [11], search for zerocrossings in the second derivative of the luminance of an image to find the edges. However, most existing edge detectors consider only the horizontal and vertical edge patterns. In some applications of these methods, they suffer of drawbacks such time consuming, the obtained edges normally are discontinuous or over detected, and used only for gray level images.
The proposed color edge detection approach, in this paper, is conceptually different and explores a new strategy. In fact, instead of considering an elaborating a better designed color edge extraction model, the proposed technique rather explores the possible alternative of combining statistical features and the automatic threshold technique. After the determination of the features of each component image, the threshold technique is then used to extract the edge of each image to be combined. The feature images are computed on overlapping small windows centered on the pixel to be classified. Then, edge results for the three color components are integrated through a fusion rule in order to get a final reliable and accurate segmentation result. Hence, this work may be seen as a straightforward additional improvement of the issues proposed by Fan et al. [5].
For instance, this method is applied to color medical image segmentation, where our main purpose is to provide some additional help to the doctor to followup the state of patient disease breast cancer by enhancing and rebuilding some distorted and pasted cells to be easily counted. The objective is to rebuild each cell from a series of three component images (R, G and B). From an initial segmentation obtained by using statistical features and automatic threshold technique, one seeks a resulting segmentation which may represents as better as possible the cells, in order to show to the doctors an appropriate schema of the set of points really forming part of the cells, as well as a suitable representation of the cells to be easily counted and checked. Section 2 introduces the proposed method for color edge extraction. The experimental results are discussed in Section 3, and the conclusion is given in Section 4.
Materials and methods
Materials
In the medical domain, the doctors follow the diseases of the breast cancer taking into account the number of tumor cells expressing hormone receptor (HR). The assessment of labelled cells expressing the HR can be done on cells in suspension or histological overpart. This latter method is used in the pathology laboratory of cancer center Salah Azaiez in Tunis, Tunisia. The sample preparation is intended to highlight the different cells to be visible.

The fixing is used for the structures conservation and the parts hardening. It must be done immediately after sampling; by immersing the material in an adequate volume of fixative (the fixer used in Salah Institute Azaiez is the formalin at (10%)).

The inclusion allows realizing the fine and regular cuts. The most commonly used inclusion medium is the paraffin.

The cuts of the paraffin block are made with a microtome. They allow producing the section slices (thin slices) of 2 μm to 5 μm thick.

The coloring performed on slides; accentuates the contrasts for the better recognition of the different structure elements. Other colors are intended to highlight specific molecules such as the immunohistochemistry. In our study, the immunohistochemistry coloring is applied to highlighting the hormone receptors.

The installation: the colored cuts are placed between the blade and the cover slip with a synthetic refractive resin index close to that of glass. Then, a microscopic preparation is obtained ready to be observed with a microscope.

The acquisition is made by a digital camera NIKON Coolpix 995 mounted on a Nikon E600 microscope that provides a color image in jpg format, with a variable size depending on the desired view field. Although the jpg format is a loss compressed format, but the morphology is preserved and what is of interest to us is the intensity of the dye which has not been significantly changed.
Methods
Image segmentation based on discontinuity tends to partition an image by detecting isolated points based on an abrupt change in gray levels [9, 12]. In the framework of our application, we are interested in segmenting the color images represented by several space colors. In fact, the problem is to detect the actual region boundaries with closed curves.
The output of most existing edge detectors such as canny operator [8, 9], can only provide candidates for the regions boundaries, because of the obtained color edges normally discontinuous or overdetected. In this context, color edge detection using statistical features and the threshold technique appears to be an interesting method.
The segmentation method, proposed in this paper, is conceptually different and explores a new strategy. In fact, instead of considering an existing segmentation procedure, the proposed technique rather explores the benefit of combining several approaches.
This method is a hybrid image segmentation technique which integrates both the results of the statistical method and the automatic threshold technique, in which the statistical features are used as the initial seed for the threshold procedures.
In this paper, instead of using the simple edge detectors, a statistical feature extracted from the cooccurrence matrix is used. The cooccurrence matrix is computed starting from a sliding window centered on the pixels of the original image. Note that the concept of cooccurrence matrix [13] was defined to express the image properties related to the second order statistics.
Meaningfully, the proposed method is divided into two stages. At the first stage, the statistical features are extracted from the cooccurrence matrix to construct a new image called the attribute image [13], and then, in the second stage, the threshold technique is then used to obtain the final segmentation results.
Edge detection
In this application, instead of implementing a simple edge detector, the statistical features extracted from the cooccurrence matrix for identifying the relationship between the pixels of an image are used. An improved coloredge detector based on the statistical features and an automatic threshold technique is developed.
The graylevel cooccurrence matrix (GLCM) [13], is largely related to the appearance frequency of couples of pixels from an image. It contains significant information which improves the discrimination classes of an image. Hence, it plays an important role in image segmentation. Here, the cooccurrence matrix is defined as the appearance frequency of pixel couples separated by a distance d in a particular direction θ and it is also called the spatial information dependence method.
Assume g _{ xy } is the intensity of a pixel p _{ xy } at the location (x, y) in an (M × N) image, w _{ xy } is a size (t × t) window centred at (x, y) for the computation of cooccurrence matrix.
The window (w _{ xy }) is the local regions where the spatial information dependence method is calculated. Thus, the size of the sliding window affects the computation of the cooccurrence matrix, in other words, affects the computation of the statistical features.
In principle, the window should be big enough to allow sufficient information to compute the statistical features extracted from the cooccurrence matrix. On the other side, a large window is time consuming to be processed. Empirically, a window is chosen for computing the cooccurrence matrix.
Each element of Cooc (i, j, R) represents the number of pixels couple (i, j), which denotes how often a pixel with graylevel (gray scale intensity) value i occurs horizontally adjacent to a pixel with the value j in the image. R(d, θ) is the space relation of the two pixels with d the distance between the two pixels and θ = {0°, 45°, 90°, 135°} the orientation of the two pixels versus the horizontal.
Here, the offset (d _{ x } , d _{ y } ), is specifying the distance between the pixelofinterest and its neighbour. Note that the offset (d _{ x } , d _{ y } ) parameterization makes the cooccurrence matrix sensitive to rotation. Choosing an offset vector, such as the rotation of the image is not equal to 180 degrees, will result in a different cooccurrence matrix for the same (rotated) image.
This can be avoided by forming the cooccurrence matrix using a set of offsets sweeping through 180 degrees at the same distance parameter d to achieve a degree of rotational invariance (i.e., [0, d] for 0°: Cooc horizontal, [d,d] for 45°: Cooc right diagonal, [d, 0] for 90°: Cooc vertical, and [d,d] for 135°: Cooc left diagonal).
In fact, the features are generated by calculating the features for each one of the cooccurrence matrices obtained by using the directions 0°, 45°, 90°, and 135°, then averaging these four values. The symbol d, represents the parameter distance, it can be assigned the value one or higher. In general, d value is set to 1 as the parameter distance. Consequently, the cooccurrence matrix allows evaluating locally the region contents of the image, this allows the detection of changes in the local statistics of the image.
where card denotes the number of elements in the set.
Consequently, several attributes may be extracted from the cooccurrence matrix such as: Mean (Mean), diagonal moment (DM), contrast (Cont), energy (Ener), directivity (Direc), entropy (Entr), opposite differential moment (ODM), variance (Var), etc.
Let Cooc(i, j) represents the occurring frequency of each pixels couple. It is obtained by calculating how often a pixel with graylevel (gray scale intensity) value i occurs horizontally adjacent to a pixel with the value j in the w _{ xy } window, N _{ c } is the maximal gray level pixel value of the sliding window w _{ xy }.
where (N _{ c } × N _{ c }) is the size of the cooccurrence matrix. The size of the cooccurrence matrix is variable according to the maximum value of the intensity in each window.
Each component of a color image may be taken as a twodimensional (2D) lightintensity function I(x, y), which contains (M × N) pixels, each with a value of brightness, i.e., grey level, from 0 to Ng. Grey level 0 is the darkest and grey level Ng is the brightest. In our study, the task of edge extraction is to classify the pixels into two opposite classes namely edge and non edge classes.
The discontinuity is a measure of abrupt changes in gray levels of pixels p _{ xy }, i.e. the discontinuity is described by its edge value, and could be obtained by applying edge detectors to the corresponding region. In the literature, there exist many different edge operators: Sobel, Canny, Derish, Laplace, etc. [8, 9], but their functions and performances are not the same. However, most existing edge detectors consider only the horizontal and vertical edge patterns and none of the proposed operators are fully satisfactory in real world use.
These Edge Strengths (ES1, ES2, ES3, ES4, ES5, ES6, ES7 and ES8) are computed as a weighted sum of the pixel values in the neighbourhood N(x, y) of the current pixel (x, y), as shown in Figure 4(a), where the weight coefficients are given in Figure 4(b).
The optimal thresholds T _{ R } , T _{ G } and T _{ B }, are automatically determined by the Otsu’s threshold technique, as described in Section 2.2. The pixel (x,y) of each primitive color is classified as an edge pixel if its local maximum edge strength of the attribute image (LMES) is higher than the optimal threshold determined automatically by the Otsu’s threshold technique, in which case is set to 1. Otherwise, it is classified as a nonedge pixel and is set to 0.
Otsu’s Threshold method
Generally, it is quite important to effectively determine a threshold value for a graylevel image when extracting the edges, i.e., to extract objects of interest from their background.
In this study, the optimal threshold can be readily selected.
by the Otsu threshold technique [14, 15]. The technique is shown to be highly efficient for the twoclass data classification problem.
To illustrate this, let the local maximum edge strength of pixels have range [0, L]. In an input image, let’s assume that there are f_{i} pixels whose local maximum edge strength has the value i, with i ∈ [0, L].
If an image can be divided into two classes, edge and nonedge classes, by a threshold at level t, where edge class consists of the local maximum edge strength features from 0 to t, and nonedge contains the other local maximum edge strength features with t + 1 to L, then the probability distributions for the edge and nonedge pixel classes can be defined, respectively.
Where $\mathrm{N}={\displaystyle {\sum}_{\mathrm{k}=0}^{\mathrm{L}}}{\mathrm{f}}_{\mathrm{k}}$ indicates the total number of pixels.
where ${\mathrm{\mu}}_{\mathrm{T}}={\displaystyle {\sum}_{\mathrm{i}=0}^{\mathrm{L}}}{\mathrm{ip}}_{\mathrm{i}}$ is the total mean of the local maximum edge strength features.
As a result, an optimal bilevel threshold can be readily selected by the Otsu’s threshold method by maximizing the betweenclass variance of the two classes. The major steps of the proposed detector are shown in Figure 1.
Results and discussion
Dataset
The images originally are stored in RGB format. Each of the primitive color (red, green and blue) takes 8 bits and has the intensity range from 0 to 255. The labeling of the original image is generated by the user based on the image used for segmentation.
Comparisons
In fact, to evaluate the efficiency and accuracy of the proposed edge extraction method, the results are compared versus existing methods, as described earlier. The efficiency evaluations for these different methods are carried out on the Matlab software 7.1.
First the segmentation results in RGB color space by applying the proposed method to red, green, and blue color features, respectively, are presented. Figure 5(a) gives the original image. The results of R, G and B components by the proposed method are shown in Figure 5(b),(c) and (d) respectively. In this case, some obvious edges in the color image are missed if the edge detector is performed on each component image (R, G and B) independently.
This shows the lack of information when using only one information source and may be explained by the high degree of correlation among of the three components of the RGB color space. Hence, it demonstrates the necessity of the merging process.
Also, to evaluate the performance of the proposed coloredge detector, we tested many color synthetic images.
Figure 6(a) gives the (N × M) synthetic image where a “salt and pepper” noise of D density was added. This affects approximately (D × (N × M)) pixels. Figure 6(b) shows the variance image.
The results obtained by the Prewitt, Canny, Sobel and Laplacian edge detectors are given in Figure 6(c),(d),(e), respectively. Figure 6(f) shows the extracted edges by performing our edge detector on the variance image.
In fact, some obvious edges in the image are missed and some incorrectly classified pixels are presented if only the simple edges detectors are performed on the noisy image.
Note that our edge detector can provide more potential edges as compared with the other detectors under the same conditions.
Accordingly, the experimental results indicate that the proposed edge detector can provide more accurate results, see Figure 6(f).
Segmentation sensitivity from Prewitt, Canny, Sobel operator and the proposed method for the data set shown in Figure 9
P rewitt operator  Canny operator  Sobel operator  Proposed technique  

Sensitivity segmentation (%)  
Image 1  85.32  87.44  86.15  92.37 
Image 2  64.52  88.96  85.54  96.46 
Image 3  74.98  87.27  86.37  95.76 
Image 4  88.53  95.60  94.31  95.45 
Image 5  76.79  91.17  77.28  97.94 
Image 6  74.28  87.72  93.46  96.48 
Image 7  76.46  96.46  85.47  97.24 
Image 8  64.18  88.25  91.45  94.35 
Image 9  74.73  87.44  93.73  96.43 
Image 10  66.17  79.76  92.01  95.76 
Image 11  47.46  69.28  77.19  88.25 
Image 12  56.49  67.16  81.72  92.43 
with: Sen(%), N _{ pcc }, N × M denote respectively the segmentation sensitivity (%), the number of correctly classified pixels and the dimension of the image.
Referring to the segmentation sensitivity given in Table 1, one can observe that 25.02, 12.73, 13.63 and 04.24 of pixels were missegmented in Figure 6(c),(d),(e) and (f), respectively. Comparing Figure 6(c),(d) and (e) with (f), it is clear that the resulting image by the proposed edge detector is much better than the one by the Prewitt, Canny and Sobel operators. In fact, the different regions are correctly classified in Figure 6(f) which is not the case in Figure 6(c), (d) and (e).
Another evaluation criterion is used to measure quantitatively the detection quality of the signal corresponding to the edge of the objects. This criterion is based on the determination of the probability of correct detection and false detection. In this case, reference edge pixels are necessary (see Figure 7(d)). This criterion resides in the comparison between labels of the obtained edge pixels and the reference edge pixels.
where N_{1r} and the N_{1r} is the number of pixel in the reference R_{1} and R_{2}, respectively. N_{1C} is the number of correctly classified pixels relative with R_{1}. N_{1f} is the number of pixels which belong to R_{2} and incorrectly classified in R_{1}. These criteria were employed by Dou et al. [18], to evaluate the effectiveness of the segmentation method in the case of abnormal tissue. In our application, these criteria are used to evaluate the proposed segmentation method in the case of color images of breast cancer cells.
Suppose R_{1} the cells pixels and R_{2} the background, thus (P_{f}) represents the probability that a background pixel is marked as a cell pixel. (P_{C}) represents the probability that a cell pixel is actually marked as a cell pixel.
The process for obtaining the correct classified pixels is not a manual process; thus a software based reference image is run. It consists in a small program which compares the labels of the obtained pixels and the reference pixels as shown in Figure 7(d). The correctly classified pixel denotes a pixel with a label equal to its corresponding pixel in the reference image.
Since the proposed method is based on the second order statistical feature and the threshold technique, we perform a comparison with several other wellknown first order statistical features combined with threshold technique (FSFT) [5].
Figure 7 shows a comparison of the results obtained by the proposed method using the first and the second order statistical features, applied to color cell images (which is a challenging problem in this field).
They correspond, respectively, to Figure 7(b) and (c). In fact, the cells are correctly localized in Figure 7(c) which is not the case in Figure 7(b).
This demonstrates the superiority of combining the second order statistical features and the threshold technique for edges detection.
The probability of correct classification and false classification from the FSFT method and the proposed method for the data set shown in Figure 9
Probability of correct classification (P_{ C }) and false classification (P _{ f } )  

FSFT method  Proposed technique  
(P _{ C } )  (P _{ f } )  (P _{ C } )  (P _{ f } )  
Image 1  0.5656  0.27  0.6196  0.13 
Image 2  0.6187  0.39  0.6778  0.21 
Image 3  0.6434  0.19  0.7081  0.15 
Image 4  0.6074  0.35  0.7117  0.11 
Image 5  0.6197  0.33  0.7786  0.13 
Image 6  0.6595  0.29  0.8141  0.17 
Image 7  0.7735  0.17  0.8973  0.14 
Image 8  0.6855  0.26  0.8464  0.21 
Image 9  0.5790  0.38  0.7081  0.17 
Image 10  0.7592  0.18  0.8979  0.15 
Image 11  0.6009  0.35  0.7441  0.19 
Image 12  0.6892  0.26  0.8027  0.21 
Its indicator corresponding to the probability of false classification (17%) is also lower than the FSFT method (29%).
Moreover, this demonstrates that the simple edge detectors are instable in noisy environments and provides edges normally discontinuous or badly detected.
In fact, errors were largely reduced when exploiting simultaneously the second order statistical features and threshold techniques for color edges extraction.
Consequently, the proposed method allows to obtain a binary mask (see Figure 7(c)), which can be used to locate the edge of the cells.
Also let us compare the performance of the proposed algorithm to those in other published reports that have been applied to color images. These include Trahanias et al. [19, 20] and Carron et al. [21]. The segmentation results are shown in Figure 8.
Figures 8(c), (d), (e) and (f) show the final segmentation results obtained from the CED_VS algorithm [19], the CED_HSI algorithm [21], the VS_CED method [20] and the proposed algorithm, respectively. Comparing Figure 8(c), (d), (e) and (f), one can observe that the two regions are correctly segmented in Figure 8(f), e.g. the cell is much better segmented in (f) than those in (c), (d) and (e), also the first, the second and the third images contain some holes in the cell and some pixels were incorrectly segmented. These do not exist in the correctly segmented image, but using the proposed algorithms, only a few singularity points are left in the final image as shown in Figure 8(f).
Conclusion
In this paper, we have proposed a new method for color edges detection based on statistical features and threshold technique. In the first phase, the edges are identified in each attribute image of the three primitive colors via a threshold operation. Then, edge results for the three color components are integrated through the fusion rule.
Instead of considering and elaborating and better designed segmentation model of natural and textured images, the proposed technique rather explores the possible alternative of combining two segmentation techniques in order to get a good consistency edges.
The results obtained show the generic and robust character of the method in the sense that the spatial neighborhood information is introduced to take into account the spatial correlation between a pixel couples from an image that exist in any physical images. Also, this technique can accurately detect the edges and suppress the impact of the noise on the results, while the edge has a good consistency. Extensive testing results have shown great potential on the proposed method. It can be useful for color edges detection.
Declarations
Acknowledgements
The authors would like to thank the EditorinChief, Prof. Kenneth R. Foster, and the anonymous reviewers. They also address their thanks to Professor Khaled Ben Romdhane, the head of the cancer service of Hospital Salah Azaiez Bab Saadoun, Tunis, Tunisia for providing the diverse images and for his helpful collaboration.
Authors’ Affiliations
References
 Yan Chai H, Jih Bao T, Khin Wee L, Tian Swee T, Hussain Salleh S: Performance metrics for active contour models in image segmentation. Int J Phys Sci 2011,6(27):6329–6341.Google Scholar
 Ben Chaabane S, Sayadi M, Fnaiech F, Brassart E: DempsterShafer evidence theory for image segmentation: application in cells images. Int J Signal Process 2009,5(2):126–132.Google Scholar
 Ben Chaabane S, Sayadi M, Fnaiech F, Brassart E: A New method for the mass functions estimation in the dempsterShafer’s evidence theory: application to color image segmentation. Circ Syst Signal Process CSSP 2011, (30):55–71.Google Scholar
 Ben Chaabane S, Sayadi M, Fnaiech F, Brassart E: Color image segmentation using homogeneity approach and data fusion techniques. Eurasip J Adv Signal Process 2010,2010(1):1–11.View ArticleGoogle Scholar
 Jianping F, David KY, Ahmed KE, Walid GA: Automatic image segmentation by integrating coloredge extraction and seeded region growing. IEEE Trans Image Process 2001,10(10):1454–1466. 10.1109/83.951532View ArticleGoogle Scholar
 Davis L: Survey of edge detection techniques. Comp Vision Graph Image Process 1975,1975(4):248–270.View ArticleGoogle Scholar
 Nalwa VS, Binford TO: On detecting edges. IEEE Trans Pattern Anal Machine Intell 1986,1986(8):699–714.View ArticleGoogle Scholar
 Qian RJ, Huang TS: Optimal edge detection in twodimensional images. IEEE Trans Image Process 1996,5(7):1215–1220. 10.1109/83.502412View ArticleGoogle Scholar
 Liu G, Haralick RM: Assignment problem in edge detection performance evaluation. Proc IEEE Comp Soc Conf Comput Vision Pattern Recog 2000, 1: 26–31.Google Scholar
 Colot O: A color image processing method for melanoma detection. Lecture notes in computer science. Berlin: Springer; 1998:562–569.Google Scholar
 Chao SM, Tsai DM: An anisotropic diffusionbased defect detection for lowcontrast glass substrates. Image Vis Comput 2008,2008(26):187–200.View ArticleGoogle Scholar
 Mery D, Berti MA: Automatic flaw detection in aluminum castings based on the tracking of potential defects in a radioscopic image sequence. IEEE Trans Robot Autom 2002,18(6):890–901. 10.1109/TRA.2002.805646View ArticleGoogle Scholar
 Haralick RM, Shanmugan K, Distein I: Textural features for image classification. IEEE Trans Syst 1973, 3: 610–621.Google Scholar
 Otsu N: A threshold selection method from graylevel histograms. IEEE Trans Syst Man Cybern SMC 1979,9(1):62–66.MathSciNetView ArticleGoogle Scholar
 Fan J, Zhang L, Gan F: Spatiotemporal segmentation based on spatiotemporal entropic threshold. Opt Eng 1997, 36: 2845–2851. 10.1117/1.601512View ArticleGoogle Scholar
 Grau V, Mewes AU, Alca˜niz M, Kikinis R, Warfield SK: Improved watershed transform for medical image segmentation using prior information. IEEE Trans Med Imaging 2004,23(4):447–458. 10.1109/TMI.2004.824224View ArticleGoogle Scholar
 Duda RO, Hart PE, Sork DG: Pattern classification. New York, NY, USA: WileyInterscience; 2000.Google Scholar
 Dou W, Ruan S, Bloyet D, Constans JM, Chen Y: Segmentation based on information fusion applied to brain tissue on IRM. SPIEIST Electronic Imaging 2004,5298(3):492–503.Google Scholar
 Trahanias PE, Venetsanopoulos AN: Color edge detection using vector order statistics. IEEE Trans Image Process 1993,2(2):259–264. 10.1109/83.217230View ArticleGoogle Scholar
 Trahanias PE, Venetsanopoulos AN: Vector order statistics operators as color edge detectors. IEEE Trans Syst Man Cybern B Cybern 1996,26(1):135–143. 10.1109/3477.484445View ArticleGoogle Scholar
 Carron T, Lambert P: Color edge detector using jointly hue, saturation and intensity. Proc ICIP 1994, 3: 977–981.Google Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.