Abstract—This work presents image segmentation technique based on colour features with K-means clustering algorithm. In this we did not used any training data. In this paper, we present a simple and efficient implementation of k-means clustering algorithm. The regions are grouped into a set of classes using K-means clustering algorithm. Results are grouped into clusters so avoiding feature calculation for every pixel in the image. Although the colour is not frequently used for image segmentation, it gives a high discriminative power of regions present in the image. Here clusters are grouped & segmentation is obtained in form of colors through which important objects are segmented, extracted or recognized.
Index Terms—color Image segmentation, K-means, clusters, unsupervised classification.
I. INTRODUCTION
he process of image segmentation is defined as: “the search for homogenous regions in an image and later the classification of these regions”. It also means the partitioning of an image into meaningful regions based on homogeneity or heterogeneity criteria (Haralick et al; 1992). Image segmentation techniques can be differentiated into the following basic concepts: pixel oriented, Contour-oriented, region-oriented, model- oriented, colour oriented and hybrid. Colour segmentation of image is a crucial operation in image analysis and in many computer vision, image interpretation, and pattern recognition system, with applications in scientific and industrial field(s) such as medicine, Remote Sensing, Microscopy, content- based image and video retrieval, document analysis, industrial automation and quality control (Ricardo Dutra, et al;2008). The performance of colour segmentation may significantly affect the quality of an image understanding system (H.S.Chen et al; 2006).The most common features used in image segmentation include texture, shape, grey level intensity, and colour. The constitution of the right data space is a common problem in connection with segmentation/classification. In order to construct realistic classifiers, the features that are sufficiently representative of the physical process must be searched. In
Manuscript received June 19, 2012.
Patel Janakkumar Baldevbhai is with the Image and Signal Processing Lab., Electrical Engineering Department, Research Scholar, EED, Indian Institute of Technology Roorkee, Uttarakhand, India on duty leave under QIP scheme of AICTE from the L.D.R.P. Institute of Technology & Research, Gandhinagar, and Gujarat, India. (Corresponding author phone:
09458121095; 079-23221371(R) e-mail: [email protected]).
R.S. Anand is with the Electrical Engineering Department, Professor, EED, Indian Institute of Technology Roorkee, Uttarakhand, India
the literature, it is observed that different transforms are used to extract desired information from remote-sensing images or biomedical images (Mehmet Nadir Kurnaz et al; 2005). Segmentation evaluation techniques can be generally divided into two categories (supervised and unsupervised). The first category is not applicable to remote sensing because an optimum segmentation (ground truth segmentation) is difficult to obtain. Moreover, available segmentation evaluation techniques have not been thoroughly tested for remotely sensed data. Therefore, for comparison purposes, it is possible to proceed with the classification process and then indirectly assess the segmentation process through the produced classification accuracies. (Ahmed Darwish, et al; 2003).Clustering is a mathematical tool that attempts to discover structures or certain patterns in a data set, where the objects inside each cluster show a certain degree of similarity.
For image segment based classification, the images that need to be classified are segmented into many homogeneous areas with similar spectrum information firstly, and the image segments‟ features are extracted based on the specific requirements of ground features classification. The colour homogeneity is based on the standard deviation of the spectral colours, while the shape homogeneity is based on the compactness and smoothness of shape. There are two principles in the iteration of parameters:1) In addition to necessary fineness, we should choose a scale value as large as possible to distinguish different regions; 2) we should use the colour criterion where possible. Because the spectral information is the most important in imagery data, the quality of segmentation would be reduced in high weightiness of shape criterion.
This work presents a novel image segmentation based on colour features from the images. In this we did not used any training data and the work is divided into two stages. First enhancing color separation of satellite image using decor relation stretching is carried out and then the regions are grouped into a set of five classes using K-means clustering algorithm. Using this two-step process, it is possible to reduce the computational cost avoiding feature calculation for every pixel in the image. Although the colour is not frequently used for image segmentation, it gives a high discriminative power of regions present in the image.
Colour segmentation is an essential issue with regard to vision applications, such as object detection and navigation (Bosch et al., 2007; Lin, 2007). The process of color segmentation consists of color representation, color feature extraction, similarity measurement and classification. In
Color Image Segmentationusing Clustering
Technique
Patel Janak kumar Baldevbhai, R.S. Anand
color representation, the RGB (Red, Green and Blue) model, which expresses color as a mixture of red, green and blue three color components, is often used to depict the color information of an image (Bascle et al., 2007; Weng et al., 2007). By using a transformation, the secondary colors, which are CMY (Cyan, Magenta and Yellow) or RG–GB–BR, can be obtained and used as an alternative color model (Wang et al., 2007). The HSI model, which transforms RGB into Hue, Saturation and Intensity, is also a popular color model at present, and its good performance has been shown in many works (Kim et al., 2007, 2008; Wangenheim et al., 2007). HSV (Value) and HSL (Luminance) are very similar to the HSI model due to the transformation formulas applied. Using the HSI color model, a specific color is able to be recognized regardless of variations in saturation and intensity. CIE Luv, CIE Lab and YCbCr (Wang and Huang, 2006; He et al., 2007) are color spaces which represent a color by its lightness (L), luminance (Y) and chromaticity (uv, ab and CbCr). The idea of color ratio was first introduced by Barnard and Finlayson in 2000 to identify the „„shadow‟‟ and „„non-shadow‟‟ regions to be robust under changes in luminance. In 2002, the RGB ratio of the pixel value to the local sum (R/Rsum, G/Gsum, B/Bsum) was proposed by Finlayson et al. to deal with the influences of shadows produced by variations in illumination. In addition, Finlayson et al. (2005) presented an alternative RGB ratio definition, which is the ratio of the intensity of a pixel to the local average (R/Rave, G/Gave, B/Bave), and this formula is used due to its invariance to luminance and device changes. In this paper, we propose a new RGB ratio model, which is based on the fact that a change in the intensity of a reference color will lead to a change in the RGB color components, but their ratios to the reference color (R/Rref, G/Gref, B/Bref) will be linear to an intensity change (Benedek and Sziranyi, 2007; Mikic et al., 2000). With this property, a specific color, such as the reference Colour, can be described as a linear color model, so that it is invariant to intensity variation. Moreover, information about the three color components (RGB) is used to describe the chromaticity by the proposed RGB ratio space. Therefore, while inheriting the characteristics of HSI and RGB models, the RGB ratio has several advantages with regard object recognition under variations in intensity.
There exist many complex and state-of-the-art techniques for colour segmentation which are excellent at partitioning an input image. For example, the global color statistics can be represented by a set of overlapping regions and modeled by a mixture of Gaussians (GMM), and a local mixture model is described by Markov Random Fields (Kato, 2008). By optimizing parameters of the global and local models, the maximum likelihood is estimated and then a pixel can be classified. Although this approach has good segmentation results, a large number of iterations are necessary to determine the optimal parameters. As a result, 16 s of computation time is needed for an image with a 256X256 resolution (Tai, 2007). Hill manipulation of the colour histogram is another widely used approach to achieve colour segmentation. A three-dimensional histogram can be obtained by accumulating three colour components of pixels. Dominant hill detection and minor hill dismantling are then
used to estimate the clustering index (Al Aghbari and Al-Haj, 2006). The idea of a „histon‟, which is an encrustation of a histogram such that the elements in the histon are the set of all the pixels that can be classified as possibly belonging to the same segment, was introduced for color segmentation by Murshrif and Ray (2008), and the total computation time this approach requires for a 179X122 image is 2.41 s. Neural networks (Bascle et al., 2007) have recently been used as a clustering kernel for color segmentation, where components of the RGB space and the intensity are used as inputs and three calibrated colour components are considered as outputs of the modified multi-layer perceptron (MLP). After the training procedure, good segmentation performance is achieved. Furthermore, the look-up tables (LUT) of the modified MLP can be applied for real-time applications, so that the execution time for a 320X 240 image is only 0.00375 s. However, a huge database needs to be created for this system to work, and if an input image is very different from those in the database, the network should be re-trained to improve the results. The well-known K-means method (Lloyd) is one of the most commonly used techniques in the clustering-based segmentation field for industrial applications and machine learning (Berkhin, 2002; Mignotte, 2008). The fuzzy c-means theory (the fuzzy version of K-means) is applied as the clustering method (Kuo et al., 2008), and similarity measurement is based on Euclidean distance (Luis-Garcia et al., 2008). Bosch et al. (2007) presented an approach that can recognize grass, sky, snow and road using fuzzy logic with predefined classes, for which the average processing time for an image size of 180X120 to 250X250 is 60 s. Efficient fuzzy c-means clustering (qFCM) is also applied to speed up the clustering process by splitting a target image into several small sub-images (Chen et al., 2005). The computation time that qFCM requires for a 128X128 gray-level image is 0.1–1.2 s. The use of a template image is another fast segmentation method. For instance, an image database of eyes can be established, and a skin colour database can be obtained from a colour conversion matrix with color of the sclera. Consequently, fixed thresholds of the HSV space are introduced to detect the skin area in an input image (Do et al., 2007). However, the use of template images is restricted to specific objects, and may require a large image database. In this paper, a dynamic fuzzy variable range is proposed to achieve a high quality segmentation result. Firstly, the linearity between the RGB ratio and intensity is estimated by a linear progressive method and parameter estimation. Secondly, upper and lower boundaries are obtained statistically for each colour ratio. These boundaries are used to define the fuzzy membership functions ofcolor ratio clusters, which dynamically vary corresponding to intensity changes. The proposed fuzzy system‟s parameter optimization, undertaken using a back propagation neural network, makes the fuzzy decision more adaptive and more effective. Meijer (1992) used sine-wave sounds to transform image information without any image pre-processing, while a multi-resolution approach was introduced to image-to-sound mapping by Capelle et al. (1998).
decor relation stretching. Section 4 describes the K-means clustering method. In section 5 the proposed method of segmentation of image based on colour with K-means clustering is presented and discussed. Experimental results obtained with suggested method are shown in section 6. Finally, section 7 concludes with some final remarks.
Mean shift-based clustering
A clustering algorithm based on mean shift was proposed in [13]. Unfortunately, it becomes impractical in the context of texture segmentation due to the expensive computation required in order to find the nearest neighbours of a point in a highdimensional space. Hence, in this work, an approximate version has been utilized. It starts by initializing the mean shift procedure on a given point and then iterates as usual until a stationary point is reached. However, at each iteration, all points involved in the mean shift computation are marked as “already visited”. Therefore, they are not taken as initial points anymore. These points are also assigned a vote regarding their membership to the cluster associated with the mode being detected. The algorithm repeats this procedure with the remaining “not visited” points.
Once all mode candidates have been found, mode merging is performed by means of the same approximate mean shift algorithm by considering the found modes as data points. If two modes are merged, their membership votes are also merged, thus keeping track of the new cluster structure. The mode merging step is repeated until no modes are merged. Membership of each point is finally determined by majority voting.
Graph clustering based on the normalized cut
The graph clustering algorithm based on the normalized cut proposed in [14] has become popular in the last years. However, the main drawback of this approach is that the computational technique for minimizing the normalized cut is based on eigenvectors. Thus, it suffers from scalability problems, since in cases where the number of data points is very large, eigenvector computation becomes prohibitive. Recently, Dhillon et al. [15] proposed a more efficient technique referred to as GRACLUS, which embeds a weighted kernel k-means algorithm into a multilevel approach in order to optimize locally the normalized cut.
However, before applying GRACLUS to the pattern discovery stage, the problem of specifying the number of clusters must be addressed such as with k-means. Usually, the alternative is to first bipartition the whole graph and then repartitions the already segmented parts if the normalized cut is below a specified value [14].
II. K-MEANS CLUSTERING
There are many methods of clustering developed for a wide variety of purposes. Clustering algorithms used for unsupervised classification of remote sensing image data vary according to the efficiency with which clustering takes place (John R Jenson, 1986).K-means is the clustering algorithm used to determine the natural spectral groupings present in a data set. This accepts from analyst the number of
clusters to be located in the data. The algorithm then arbitrarily seeds or locates, that number of cluster centers in multidimensional measurement space. Each pixel in the image is then assigned to the cluster whose arbitrary mean vector is closest. The procedure continues until there is no significant change in the location of class mean vectors between successive iterations of the algorithms (Lille sand and Keiffer, 2000). As K-means approach is iterative, it is computationally intensive and hence applied only to image subareas rather than to full scenes and can be treated as unsupervised training areas (Lillesand & Keiffer, 2000). K-means-based clustering
Due to its simplicity and good convergence properties, the iterative k-means algorithm is probably the most widely used clustering algorithm. However, it suffers from important drawbacks, such as the requirement of specifying the number of clusters and the non-deterministic results produced if random initialization is used (which is often the case). In order to overcome the aforementioned problems, a wrapper for k-means, which is a variation of the resolution-driven clustering algorithm proposed in [11], has been applied. It has two main stages: split and refinement. Regarding the split stage, let us assume that the data points have been split into
C disjoint clusters (initially C = 1). The mean distance between the centroid and its associated points (intra-cluster mean distance) is computed for each cluster and the global mean distance (mean of intra-cluster mean distances) is obtained for the whole partition. If this global mean distance exceeds a threshold, the largest cluster in terms of intra-cluster mean distance is split into two. The split is done by finding the main principal component ρ of the cluster and initializing two new child centroids at c ±d, where c is the centroid of the cluster to be split and d = ρ√2λ/π, with λ being the eigenvalue associated with the main principal component ρ. After the split stage, the refinement stage consists of applying k-means using the (C + 1) available centroids as initial seeds. Both split and refinement are iterated until no new clusters are generated.
The proposed wrapper has two main advantages over the classical k-means. First, instead of the desired number of clusters, the mean distance threshold controls the output of the algorithm.Such a threshold is more intuitive and closely related to perceptual properties than the number of clusters. Second, the algorithm always behaves in the same way given the same input. Therefore, there is no need for running different trials and keeping the best set of clusters according to some criterion as it is the case when the initialization step of k-means has a random component.
Colour-Based Segmentation Using K-Means Clustering
Thebasicaimistosegmentcolorsinanautomatedfashionusingth eL*a*b*colorspaceandK-means
clustering.Theentireprocesscanbesummarizedinfollowingste ps.
Step1:Readtheimage
Readtheimagefrommother source whichisin.JPEGformat.
Step2:ForcolorseparationofanimageapplytheDecor
relationstretching.
Howmanycolorsdoweseeintheimage ifweignorevariations inbrightness? Therearethree colors:white,blue,andpink. Wecaneasilyvisuallydistinguish thesecolorsfromoneanother. TheL*a*b*colorspace(alsoknownasCIELAB
orCIEL*a*b*)enablesustoquantifythese visualdifferences. The L*a*b*colorspaceisderivedfromtheCIEXYZtristimulusvalues. The
L*a*b*spaceconsistsofaluminositylayer'L*',chromaticity-layer 'a*'indicatingwherecolor falls alongthered-greenaxis,and chromaticity-layer'b*'indicatingwherethecolorfallsalongthe
blue-yellow axis.Allofthecolorinformation
isinthe'a*'and'b*'layers.Wecanmeasurethe difference
betweentwocolorsusingtheEuclideandistancemetric.Convertthe imagetoL*a*b* colorspace.
Step4:ClassifytheColorsin'a*b*'SpaceUsingK-MeansClustering
.
Clusteringisa way
toseparategroupsofobjects.K-meansclusteringtreatseach
objectashaving alocationinspace. Itfindspartitions
suchthatobjectswithineachclusterareasclosetoeach
otheraspossible,andas farfromobjectsinotherclustersas possible.K-meansclusteringrequires
thatyouspecifythenumberofclusters tobepartitioned
andadistancemetrictoquantifyhow
closetwoobjectsaretoeachother.Sincethecolorinformation existsinthe'a*b*'space,your
objectsarepixelswith'a*'and'b*'values. UseK-meanstocluster theobjectsintothreeclusters usingtheEuclideandistancemetric. Step5:LabelEveryPixelinthe
ImageUsingtheResultsfromK-MEANS
Foreveryobjectinourinput,K-meansreturnsanindexcorrespon
ding toacluster. Labelevery pixelin
theimagewithitsclusterindex.
Step6:CreateImagesthatSegmenttheImagebyColor.
Usingpixellabels,wehavetoseparateobjectsinimagebycolor, whichwillresultinfiveimages.
Step 7: Segment the Nuclei into a Separate Image
Then programmatically determine the index of the cluster containing the blue objects because K means will not return the same cluster idx value every time. We can do this using the cluster center value, which contains the mean 'a*' and 'b*' value for each cluster.
1. Select
k
-seeds s.t.d
(k
i,k
j) >d
min2. Assign points to clusters by min dist.
Cluster (
p
i) =Arg
min
(d
(p
i,s
j))j
s
{s
1,…,s
k}3. Compute new cluster centroids:
1
th i p j cluster
j
n
i
C
p
4. Reassign points to clusters (as in 2 above) 5. Iterate until no points change clusters
Supervised pixel-based classification
At this stage, the set of texture patterns found by the previous stage are used as texture models for a supervised pixel based classifier, thus effectively transforming the original
unsupervised problem into a supervised one.
As its name suggests, a pixel-based classifier aims at determining the class to which every pixel of an input image belongs, which leads to the segmentation of the image as a collateral effect.
In order to achieve this objective, several measures are computed for each image pixel by applying a number of texture feature extraction methods as described in Section 3.1.
Classification with multiple evaluation window sizes Although previous works on supervised pixel-based classification have already shown the benefits of utilizing multiple evaluation window sizes [10, 11], which approach is the best for combining these different sources of information is still an open issue.
For instance, in [10], different window sizes were integrated by assigning a weight to their corresponding probabilities according to how well each window size separates a given training pattern from the others. However, since the training patterns are single-textured images, the assigned weight is not representative of the structure of the test image, which in turn is composed of multiple texture patterns. Furthermore, this method may be biased to the largest window, as it captures more information and, hence, has better capabilities of distinguishing between texture classes. Later, in [11], improved classification rates were obtained by directly fusing the outcome of multiple evaluation window sizes using the KNN rule. The main problem with this approach is that it does not guarantee that the most appropriate window size will always receive the majority of votes.
Ideally, the strategy for classifying a test image using multiple evaluation window sizes should apply large windows inside regions of homogeneous texture in order to avoid noisy classified pixels and small windows near the boundaries between those regions in order to define them precisely. Unfortunately, that kind of knowledge about the structure of the image is only available after it has been segmented. Notwithstanding, an a priori approximation of that strategy can be devised through the following steps:
Step 1: Select the largest available evaluation window and classify the test image pixels labelled as unknown (initially, all pixels are labelled as unknown).
Step 2: In the classified image, locate the pixels that belong to boundaries between regions of different textures and mark them as unknown, as well as their neighbourhoods.
The size of the neighbourhood corresponds to the size of the window used to classify the image.
Step 3: Discard the current evaluation window.
Step4: Repeat steps 1 to 3 until the smallest evaluation window has been utilized.
This scheme, which can be thought of as a top-down approach, has been used during the supervised classification stage of the proposed segmentation technique. In addition to closely approximating the previously described ideal strategy for using multiple evaluation window sizes, this approach avoids the classification of every image pixel with all the available windows. Hence, it leads to a lower computation time than previous approaches.
III. RESULTS AND DISCUSSION
software. We have used Peppers, Planet, Lena images from Mat Lab software as a standard image. Addition to these images we have implemented this proposed algorithm on heart image also & obtain segmentation results. Figure 1(a) shows original image of Peppers.png image and figure 1(b)-1(g) show various segmented objects from original image. Here various color clusters and segmented objects are clearly visible. Table 1 shows parameter values of Peppers.png image like Min, Max, mean, median, mode, Standard Deviation and range. Figure 1(h) shows the scatter plot of original image Peppers.png. Figure 1 (i) shows Scatter plot with Bar and values of Peppers.png image. Figure 1 (j) shows Graph of Parameter values of Peppers.png image. Figure 1 (k) shows Radar Graph of Parameter values of Peppers.png image. Figure 2 (a) shows the second image of our test data image of original Planets standard image from mat lab software. Figure 2(b) and 2(c) shows Object Segmentation from Planets image. Table 2 shows Parameter Values of Planets image. Figure 2(d) shows Scatter plot of Planets image. Figure 2 (e) represents Graph of Parameter values of Planets image and Figure 2 (f) represents Radar Graph of Parameter values of Planets image. Similarly Figure 3 shows results for Lena Image. Figure 4 shows segmentation results of Heart image.
Figure 1 (a) Original Peppers standard image from matlab
Figure 1 (b) Object Segmentation from Peppers image having orange color
Figure 1 (c) Object Segmentation from Peppers image having light green color
Figure 1 (d) Object Segmentation from Peppers image having red color
[image:5.595.305.557.66.252.2] [image:5.595.304.557.290.474.2] [image:5.595.48.300.355.544.2] [image:5.595.301.555.514.702.2] [image:5.595.46.294.570.759.2]Figure 1 (f) Object Segmentation from Peppers image
Figure 1 (g) Object Segmentation from Peppers image
100 120 140 160 180 200 220
80 100 120 140 160 180 200 220
Scatterplot of the segmented pixels in 'a*b*' space
'a*' values
'b
*'
v
a
lu
e
s
Figure 1 (h) Scatter plot of Peppers.png image
100 120 140 160 180 200 220 80
100 120 140 160 180 200 220
10 20 30 40 50 60
Black 105 x min 150 x max 127 x mean 128 x median 136 x mode 9.173x std Red Green Violet Magenta Yellow
Figure 1 (i) Scatter plot with Bar and values of Peppers.png image
Figure 1 (j) Graph of Parameter values of Peppers.png image
0 50 100 150 200
250Black X Black Y
Red X
Red Y
Green X
Green Y Violet X
Violet Y Magenta
X Magenta
Y Yellow X
Yellow Y
Min
Max
mean
median
mode
std
range
[image:6.595.46.299.51.241.2] [image:6.595.322.536.62.234.2] [image:6.595.47.316.275.519.2] [image:6.595.46.534.283.689.2]Table 1: Parameter Values of Peppers.png image
[image:7.595.5.298.96.453.2]Figure 2 (a) Original Planets standard image from matlab
Table 2: Parameter Values of Planets image
Figure 2(b) Object Segmentation from Planets image
Figure 2(c) Object Segmentation from Planets image
120 130 140 150 160 170 180 190 200 60
80 100 120 140 160 180 200
Scatterplot of the segmented pixels in 'a*b*' space
'a*' values
'b
*'
v
a
lu
e
s
Figure 2(d) Scatter plot of Planets image
0 50 100 150 200 250
Red X
Red Y
Violet X
Violet Y
Figure 2 (e) Graph of Parameter values of Planets image Peppers.png Min Max mean med mode STD ran
ge Black X 105 150 127.21 128 136 9.1729 45
Black Y 126 160 147.13 148 148 6.6843 34
Red X 106 156 122.8 120 115 9.467 50
Red Y 152 176 165 165 167 4.936 24
Green X 156 201 183.05 185 187 7.9559 45
Green Y 133 201 169.10 168 173 12.5043 68
Violet X 128 179 155.5 156 168 12.93 51 Violet Y 176 214 202.4 204 204 7.818 38 Magenta X 110 156 126.3 123 121 9.347 37 Magenta Y 163 200 181.3 182 185 6.347 37
Yellow X 126 184 147.6 147 147 4.66 58
Yellow Y 92 153 115.5 115 115 6.838 61
Planets.jpg Min Max mean med mode std range
Red X 120 161 134.2 133 132 4.6 41
Red Y 61 121 97.84 96 95 11.52 60
[image:7.595.15.299.96.455.2] [image:7.595.327.548.287.466.2] [image:7.595.308.562.510.685.2] [image:7.595.25.297.522.750.2]0 50 100 150
200Min
Max
mean
median mode
std range
Red X
Red Y
Violet X
Violet Y
Figure 2 (f) Radar Graph of Parameter values of Planets image
Figure 3 (a) Original standard image of Lena from matlab
Figure 3(b) Object Segmentation from Lena image
[image:8.595.67.275.310.575.2]Figure 3(d) Object Segmentation from Lena image
[image:9.595.46.244.62.312.2]Figure 3(e) Object Segmentation from Lena image
Table 3: Parameter Values of Lena image
120 130 140 150 160 170 180 190 80
100 120 140 160 180 200 220
Scatterplot of the segmented pixels in 'a*b*' space
'a*' values
'b
*'
v
a
lu
e
s
Figure 3(f) Scatter plot of Lena image Lena.tif
f
Min Ma x
mea n
media n
mod e
std rang e Black X 168 190 173.7 174 174 2.91
3
22
Black Y 140 187 151.3 151 149 3.99 1
47
Red X 166 182 171 171 172 2.40
2
16
Red Y 127 148 142 142 143 3.29
5
21
Green X 147 176 161 163 165 5.96
2
29
Green Y 124 143 133.6 134 141 5.32 2
19
Violet X 132 178 161.1 162 163 4.51 9
46
Violet Y 90 125 116.2 117 120 5.71 4
35
Magenta X
125 148 139.5 139 138 3.63 23
Magenta Y
109 182 143.4 141 139 8.52 3
73
Yellow X
133 169 157.9 158 156 6.08 1
36
Yellow Y
142 210 152.1 151 146 6.85 5
[image:9.595.272.556.84.407.2] [image:9.595.46.242.335.597.2] [image:9.595.316.538.442.629.2]0 50 100 150 200 250 B la ck X B la ck Y R e d X R e d Y G re en X G re en Y V io le t X V io le t Y M a ge n ta X M a ge n ta Y Y ello w X Y ell o w Y Min Max mean median mode std range
Figure 3(g) Graph of Lena image parameter values
0 50 100 150 200
250Black X Black Y
Red X
Red Y
Green X Green Y Violet X
Violet Y Magenta X Magenta Y Yellow X Yellow Y Min Max mean median mode std range
Figure 3(h) Radar Graph plot of Lena image parameter values
Figure 4 (a) Original image of Heart
Figure 4 (b) Segmented object1 of Heart image
Figure 4 (c) Segmented object2 of Heart image
Figure 4 (d) Segmented object3 of Heart image
110 120 130 140 150 160 170 180 190 200 60 80 100 120 140 160 180 200
Scatterplot of the segmented pixels in 'a*b*' space
'a*' values 'b *' v a lu e s
[image:10.595.50.323.56.267.2] [image:10.595.318.538.407.709.2] [image:10.595.45.219.603.743.2]Figure 5 Quantitative Comparison of Segmentation Methods
[image:11.595.50.556.60.583.2]Figure 6 Quantitative Comparisons of Segmentation Methods
IV. CONCLUSION
We have presented an efficient implementation of k-means clustering algorithm. The algorithm has been implemented on standard images from mat lab software. Results are plotted in scatter plots showing the clusters & Radar plot showing the data analysis of clusters. Various segmentation methods are given in form of chart. The plot of segmentation method shows unsupervised k means cluster Method is better as compared to supervised classification segmentation methods. And the more well separated the clusters, the faster the algorithm runs. This algorithm is significantly more efficient than the other methods.
REFERENCES
[1] Ahmed Darwish, et al, Image Segmentation for the Purpose Of Object-Based Classification,, 2003 IEEE pp. 2039-2041
[2] Darren MacDonald, et al; Evaluation of colour image segmentation hierarchies, proceeding of the 3rd Canadian conference on
Computer and Robot Vision, IEEE, 2006.
[3] H C Chen et al, Visible color difference-based quantitative evaluation of colour segmentation, IEEE proceeding, Vis image signal process vol.153 No.5 Oct 2006 pp 598-609.
[4] Jean-Christophe Devaux et al; Aerial colour image segmentation by Karhunen-Loeve transform, 0-7695-0750-6, IEEE 2000, pp 309- 312.
[5] Jun Tang, A color image segmentation algorithm based on region growing, 978-1-4244-6349-7, IEEE, vol 6, 2010, pp. 634-637 [6] Mehmet Nadir Kurnaz,et al; Segmentation of remote-sensing images
by incremental neural network, Pattern Recognition Letters 26 (2005) 1096–1104, pp 1096-1103.
[7] N Bartneck et al; Colour segmentation with polynomial classification,0-8186-2915-0/92, 1992, pp. 635-638.
[8] Nae-Joung Kwak et al; color image segmentation using edge and adaptive threshold value based on the image characteristics,IEEE proceeding0-7803-8639-6,2004, pp 555-558.
[9] Lingkui Meng, et al, Study on Image Segment Based Land Use Classification and Mapping, 2009 IEEE, pp
[10] Pal N.R.et al; A review on image segmentation techniques, Pattern Recognition 26(9), 1993, pp1277-1294.
[11] Ricardo Dutra da Silva et al; Satellite image segmentation using wavelet transforms based on color and texture features, ISVC 2008, part II, LNCS 5359, 2008, pp 113-122
[12] Robert A Schowengerdt, Remote sensing- models and Methods for Image Processing, IIIrd edition, Elsevier Inc.
[13] T. W. Chen, Y. L. Chen and S. Y. Chien. “Fast image segmentation
based on K-Means clustering with histograms in HSV color space”. In
Proceedings of 10th IEEE Workshop on Multimedia Signal Processing, 2008.
[14] C.W. Chen, J. Luo, K.J. Parker, “Image segmentation via adaptive K-mean clustering and knowledge based morphological operations with biomedical applications”, IEEE Transactions on Image Processing, Vol.7 (12), 1998, pp 1673-1683.
[15] B. Sowmya, B. Sheelarani, “Colour Image Segmentation Using Soft
Computing Techniques”.
International Journal of Soft Computing Applications, 4:69-80, 2009. [16] M. Mirmehdi, M. Petrou, Segmentation of color textures, IEEE Trans.
Pattern Anal. 22 (2000) 142-159.
[17] S.C. Kim, T.J. Kang, Texture classification and segmentation using wavelet packet frame and Gaussian mixture model, Pattern Recogn. 40 (2007) 1207-1221.
[18] D. Puig, M.A. Garcia, Automatic texture feature selection for image pixel Classification, Pattern Recogn. 39 (2006) 1996-2009. [19] J. Melendez, M.A. Garcia, D. Puig, Efficient distance-based
per-pixeltexture classification with Gabor wavelet filters, Pattern Anal. Appl. 11
(2008) 365-372.
[20] M. Omran, A. Engelbrecht, A. Salman, An overview of clustering methods, Intell. Data Anal. 11 (2007) 583-605.
[21] D. Comaniciu, P. Meer, Mean shift: A robust approach toward feature Space analysis, IEEE Trans. Pattern Anal. 24 (2002) 603-619. [22] J. Shi, J. Malik, Normalized cuts and image segmentation, IEEE
Trans.
Pattern Anal. 22 (2000) 888-905.
[23] I.S. Dhillon, Y. Guan, B. Kulis, Weighted graph cuts without eigenvectors: A multilevel approach, IEEE Trans. Pattern Anal. 29 (2007) 1944-1957.
[24] D. Tsujinishi, Y. Koshiba, S. Abe, Why pairwise is better than oneagainst-all or all-at-once, in: Proceedings of the IEEE IJCNN, 2004, pp.693-698.
[25] W.-Y. Ma, B.S. Manjunath, Edge Flow: A technique for boundary detection and image segmentation, IEEE Trans. Image Process. 9 (2000) 1375-1388.
[26] A.Y. Yang et al., Unsupervised segmentation of natural images via lossydata compression, Comput. Vis. Image Und. 110 (2008) 212-225.
Janak B. Patel (born in 1971) received B.E. (Electronics & Communication Engg from L.D. College of Engg. Ahmedabad, and M.E. (Electronics Communication & System Engg.) in 2000 from DDIT. He is Asst. Prof. & H.O.D. at L.D.R.P.I.T.R., Gujarat. He is pursuing Ph.D. at Indian Institute of Technology, Roorkee.
[image:12.595.63.427.48.540.2]