Image Segmentation using K-means clustering and Thresholding

Transcription

1 Image Segmentation using Kmeans clustering an Thresholing Preeti Panwar 1, Girhar Gopal 2, Rakesh Kumar 3 1M.Tech Stuent, Department of Computer Science & Applications, Kurukshetra University, Kurukshetra, Haryana 2Assistant Professor, Department of Computer Science & Applications, Kurukshetra University, Kurukshetra, Haryana 3Professor, Department of Computer Science & Applications, Kurukshetra University, Kurukshetra, Haryana *** Abstract Image segmentation is the ivision or separation of an image into regions i.e. set of pixels, pixels in a region are similar accoring to some criterion such as colour, intensity or texture. This paper compares the colorbase segmentation with kmeans clustering an thresholing functions. The kmeans use partition cluster metho. The kmeans clustering algorithm is use to partition an image into k clusters. Kmeans clustering an thresholing are use in this research for the comparison. The comparisons of both techniques are base on segmentation parameters such as mean square error, peak signaltonoise ratio an signaltonoise ratio. MSR an are wiely use to measure the egree of image istortion because they can represent the overall graylevel error containe in the entire image. Keywors: Image segmentation, kmeans clustering, thresholing, MSR, 1. INTRODUCTION Image segmentation is one of the most important techniques in image processing. It is a preprocessing step in the area of image analysis, computer vision, an pattern recognition [1]. The process of iviing a igital image into multiple regions (sets of pixels) is calle image segmentation. Image segmentation is commonly use to etermine objects an bounaries (lines, curves, etc.) in images. The result of image segmentation is a set of segments that inclue the entire image, or a set of contours extracte from the image (ege etection). All pixels in a region is relate with respect to some features or compute property, such as color, intensity or texture [2]. Ajacent regions are significantly ifferent with respect to the same characteristics. Some of applications of image segmentation are: igital libraries, imageprocessing, meical imaging, computer vision, face recognition, image an vieo retrieval etc[3]. Image segmentation is a lowlevel image processing task that aims at iviing an image into homogenous regions. Segmentation algorithms are base on one of the two basic properties of intensity, iscontinuity an similarity. First category is to partition an image base on small changes in intensity, like eges in an image. Secon category is base on partitioning an image into regions that are similar accoring to some preefine criterion. Threshol approach comes uner this category [4]. Image segmentation methos fall into ifferent categories: Region base segmentation, Ege base segmentation, an Clustering base segmentation, Thresholing, Artificial neural network, featurebase segmentation. Clustering of an image is one of the goo techniques, which is use for segmentation of images. After extraction of features, these features are put together into wellseparate clusters base on each class of an image. The clustering algorithm aim is to evelop the partitioning ecisions base on initial set of clusters that is upate after each iteration[5]. This paper focuse on kmeans clustering, thresholing an their proceures. Section 2 escribes the Kmeans clustering metho in etail. In section 3, thresholing function is escribe. Section 4 etails about the comparison formation. Section 5 illustrates the experiment setup an results. The analysis of results is provie in section 6, an finally section 7 presents the conclusions of the stuy. 2. KMEANS CLUSTERING Currently the clustering metho often use for segmenting largescale images. Clustering is one of the unsupervise learning metho in which a set of essentials is separate into uniform groups. There are ifferent types of clustering: hierarchical clustering, Fuzzy Cmeans clustering, Kmeans clustering. The K means metho is one of the most generally use clustering techniques for various applications [6]. Kmeans clustering is a partitionbase cluster analysis metho. The Kmeans clustering technique is a wiely use approach that has been applie to solve lowlevel image segmentation tasks. The choosing of initial cluster centers is very important since this prevents the clustering algorithm to proucing incorrect ecisions. The most common initialization proceure chooses the 2016, IRJET Impact Factor value: 4.45 ISO 9:2008 Certifie Journal Page 1787

2 initial cluster centres ranomly from input ata[5]. The proceure of Kmeans clustering is given below: 2.1 Proceure Kmeans is an efficient clustering technique. Base on initial centrois of cluster it is use to separate similar ata into groups. Accoring to this algorithm, firstly it chooses k ata value as initial cluster centers, then fins the istance between each cluster center an each ata value an assign it to the nearest cluster, upate the averages of every clusters, repeat this process until the criterion is not match. Kmeans clustering aims to ivie ata into k clusters in which each ata valuebelongs to the cluster with the closest mean[7]. Fig 1. Shows the process of basic kmeans. formula, base on the mean value of the objects in the cluster. 4. Upate the cluster means, i.e. etermine the mean value of the objects for each cluster 5. Until no change. Where is a selecte istance (intra) calculate between a ata point xi an the cluster centre cj, is an inicator of the istance of cluster center from their n ata points. The term intra is use to measure the compactness of the clusters. The inter term is the minimum istance between the cluster centrois. One of the main isavantages of kmeans is the fact that there is a nee to specify the number of clusters as an input to the algorithm. As esigne, the algorithm is not able of fining the appropriate number of clusters an epens upon the user to ientify this in avance[8]. 1: Kmeans Algorithm Process Fig 3. THRESHOLDING Image thresholing is an important technique for image processing an pattern recognition. Several methos have been propose to choose the threshols automatically. Thresholing is one of the most commonly use image segmentation technology[9].its characteristics are simple operation, an the segmentation results are of series of continuous regions. Thresholing base image segmentation requires fining a threshol value T that establishes the borer among graylevel image range corresponing to objects an a range equivalent to backgroun. After thresholing the graylevel image is change to binary. There exist algorithms that use more than one threshol value, which enables to assign pixels to one of a few classes instea of just two. Threshol value may be entere manually or automatically [10]. The proceure of thresholing is given below: KMeans Algorithm: The algorithm for kmeans, where each cluster s center is represente by mean value of objects in the cluster[8]. Input: k: the number of clusters. D: a ata set containing n objects. Output: A set of k clusters. Metho: 1. Ranomly select k objects from D as the initial cluster centers. 2. Repeat 3. (re) assign each object to the cluster to which the object is mainlyrelate using given below 3.1 Proceure The process of threshol segmentation is as follows: first, fin out a threshol T, for every pixel in the image, if the gray value is greater than T, then set it s the target point (the value is 1), otherwise locate it as the backgroun point (the value is 0), or vice versa, so the image is ivie into backgroun region an target region. Similarly, in programming, the target pixel can also be set as 255, backgroun pixel 0, or vice verso, so the image is partition into the target region an backgroun region. The formula can be represente as follows: [9] 2016, IRJET Impact Factor value: 4.45 ISO 9:2008 Certifie Journal Page 1788

3 4. PROPOSED WORK This paper compares the performance of various segmentation techniques for color images. Two techniques are use for the comparison i.e., kmeans clustering an thresholing. Segmentation by Kmeans clustering an thresholing techniques are compare by their performance in segmentation of color images. Segmentation of an image entails the ivision or separation of the image into regions of relatecharacteristic. In this four images are taken for the segmentation. These four images are: onion.png, peppers.png, hestain.png, fabric.tif. Fig 2(a) Peppers.png Fig 2() Onion.png Fig 2. Original Images We perform the kmeans clustering an aaptive thresholing to obtain the result. The performance of these techniques is measure using segmentation parameters peak signaltonoise ratio,mean square error, signaltonoise ratio. 5. EXPERIMENTAL RESULT Comparative evolution of various images has been one. The comparison of various images is one in MATLAB. Five ifferent images are use for this experiment because the images have ifferent color regions. The result of experiment is use to fin the MSE, an SNR value. The results that are obtaine by using k means clustering an thresholing shown in below figures. 5.1 Segmentation by kmeans 1. First image is peppers.png Fig 2(b) Fabric.png Fig 3.1 (a) segmentation using Fig 2(c) Hestain.png Fig 3.1 (b) segmentation using 2016, IRJET Impact Factor value: 4.45 ISO 9:2008 Certifie Journal Page 1789

4 3. Thir image is Hestain.png Fig 3.1(c) Segmentation using Fig 3 segmente image of peppers.png Fig 3.3 (a) Segmentation using 2. Secon image is Fabric.png Fig 3.3 (b) Segmentation using Fig 3.2 (a) Segmentation using Fig 3.3(c) Segmentation using Fig 3.2 (b) Segmentation using 4. Fourth image is Onion.png Fig 3.2(c) Segmentation using Fig 3.4 (a) Segmentation using 2016, IRJET Impact Factor value: 4.45 ISO 9:2008 Certifie Journal Page 1790

5 Fig 3.4 (b) Segmentation using Fig 4 (c) Hestain Fig 3.4(c) Segmentation using Fig 3. Shows the segmente image after clustering 5.2 Segmentation by thresholing: Fig 4() Onion Fig 4. Shows segmente image after thresholing 6. PERFORMANCE ANALYSIS The performance of these techniques is measure using segmentation parameters: Mean Square Error, Peakto SignalNoise, SignaltoNoise. 1. Mean square error is an average of the squares of the ifference between the preicate observations an actual. For an m*n image the MSE can be calculate as Fig 4(a) Peppers 2. Peak signal to noise ratio term for ratio between the maximum possible power of a signal an the power of corrupting noise that affects the fielity of its representation. Because many signals have an extremely large ynamic range. is commonly use as measure of quality reconstruction of image. High value of inicates the high quality of image. is usually expresse in terms of the logarithmic scale. Fig 4(b) Fabric 2016, IRJET Impact Factor value: 4.45 ISO 9:2008 Certifie Journal Page 1791

6 Table2: Results of peppers.png 3. Signal to noise ratio is a measure use to compare the level of a require signal to the level of backgroun noise. It is efine as the ratio of signal power to the noise power. Metho SNR MSR Cluster1 Cluster2 Cluster Table 2 shows the value of peppers.png image for three ifferent clusters. The values are represente using parameters, SNR an MSE. 2. Table of secon image Fabric.png Table3: Results of fabric.png Thresholing table Table1. Results of thresholing Met ho PSN R Peppers Fabric Hestain Onion SNR e+000 MSE e e+ooo e Table 1 represent the value of thresholing for four images using ifferent parameters. The three parameters, SNR an MSE show the value of four images. Kmeans clustering table 1. Table of first image Peppers.png Metho SNR MSE Cluster1 Cluster2 Cluster e e e e e e Table 3 shows the value of fabric.png image for three ifferent clusters. The values are represente using parameters, SNR an MSE. 3. Table of thir image Hestain.png Table4: Results of hestain.png Metho SNR MSE Cluster1 Cluster2 Cluster e e e Table 3 shows the value of hestain.png image for three ifferent clusters. The values are represente using parameters, SNR an MSE. 4. Table of fourth image Onion.png 2016, IRJET Impact Factor value: 4.45 ISO 9:2008 Certifie Journal Page 1792

7 management an research. Table5: Results of onion.png Metho Cluster1 Cluster2 Cluster SNR e e e+0 00 MSR Table 5 shows the value of onion.png image for three ifferent clusters. The values are represente using parameters, SNR an MSE. 7. CONCLUSION A comparative stuy of two segmentation techniques has been performe in this stuy. The Kmeans clustering an thresholing techniques were chosen for segmentation. Using these two techniques, the performance for ifferent images were segmente by using the parameters like MSE,, an SNR. From observations shown in this paper, one can conclue that the thresholing technique gives an output of two segments. However, in kmeans techniques, output is of various segments accoring to cluster size. Performance improves accoring to cluster sizes. More is the cluster size more is the accuracy percentage. MSR an are use to measure the quality of reconstruction. value of four images in kmeans clustering is higher than thresholing an MSE value is lower. [4] O. Singh, New Metho of Image Segmentation for Stanar Images, IJCST, vol. 2, no. 3, september [5] S. pana, Color Image Segmentation Using Kmeans Clustering an Thresholing Technique, IJESC, march [6] L. H. a. J. Y. Lihua Tian, Research on Image Segmentation base on Clustering Algorithm, International Journal of Signal Processing, Image Processing an Pattern Recognition, vol. 9, pp. 112, [7] B. T. Sachin Shine, Improve Kmeans Algorithm for Searching Research Papers, International Journal of Computer Science & Communication Networks, vol. 4, pp [8] A. S. B. M. a. H. K. S, Dynamic Clustering of Data with Moifie KMeans Algorithm, International Conference on Information an Computer Networks, vol. 27, [9] L. H. a. L. Shengpu, An Algorithm an Implementation for Image Segmentation, International Journal of Signal Processing, Image Processing an Pattern Recognition, vol. 9, pp , [10] W. B. a. S. Grabowski, Multipass approach to aaptive thresholing base image segmentation, 26 feb REFERENCES [1] R. R. N. Senthilkumaran, A Stuy on Rough Set Theory for Meical Image Segmentation, International Journal of Recent Trens in Engineering, vol. 2, november [2] S. k. A. P. Prasa Dakhole, Fabric Fault Detection Using Image Processing Matlab, International Journal For Emerging Trens in Engineering an Management Research (IJETEMR), vol. 2, no. 1, 21 january [3] C. R. Tippana, Homogeneous Regions for Image Segmentation Base on Fuzzy, international journal & magazine of engineering, technology, 2016, IRJET Impact Factor value: 4.45 ISO 9:2008 Certifie Journal Page 1793