International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)

Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com Volume 2, Issue 2, March April 2013 ISSN 2278-6856

Evolution of Contours in Deformable Models using Single and Grid Mask Initialization
Neeru Gulati
Department of Computer Science & Engineering Ambala College of Engineering & Applied Research, Devsthali, Ambala

Abstract: The task of segmentation which partitions an

image into different regions that contains different objects is still remain challenging. Deformable models, also known as active contour models, begin with a contour in an image plane and then use different methods to evolve the contour according to some evolution equation. The Chan - Vese algorithm evolve the contour via a level set method which is a powerful tool for performing contour evolution. The way to initialize the contour is a very important step and directly affects the quality and performance of segmentation. In this paper, we have initialized the contour as single and grid mask and applied it on teeth images useful in dental application. We have also studied the optimal no of contours in a grid that can give the best and acceptable segmentation results. We segment the images using deformable model (Chan -Vese) based on curve evolution and Mumford-Shah functional based on the level set. The results have been compared on the basis of convergence time.

Keywords: Active contours, energy minimization, segmentation, energy function, deformable models.

1. INTRODUCTION
Image segmentation is the part of every image processing application. Segmentation partitions an image into different parts or objects. It refers to the process of partitioning a digital image into multiple segments i.e. set of pixels, pixels in a region are similar according to some homogeneity criteria such as colour, intensity or texture, so as to locate and identify objects and boundaries in an image [1]. More precisely, image segmentation is the process of assigning a label to every pixel in an image such that pixels with the same label share certain visual characteristics. We can segment the image by using a PDE based method & solving the PDE equation by a numerical scheme. Image segmentation based on PDEs is mainly carried out by active contour model or snakes. This method was first introduced by Kass et al in 1987 [2]. Kass developed this method to find familiar objects in presence of noise and other ambiguities. . Snakes lock on to nearby minima in the potential energy generated by processing an image. The name deformable models or snakes first appeared in work by Terzopoulos and his collaborators. Active contours or snakes are computer generated curves that move within the image to find object boundaries under the influence of forces of curve and image itself. Active contours detect the objects in a given image, based on techniques of curve evolution, Volume 2, Issue 2 March April 2013

MumfordShah functional for segmentation and level sets. A formulation, to implement these active contours, was proposed by the level set theory [3]. One may visualize the active contour as a rubber band of arbitrary shape that is deforming with time trying to get as close as possible to the object contour. This model can detect objects whose boundaries are not necessarily defined by gradient, Active contours without edges, Active contour without edges for vector image, Active contours with multi-phases. Each one is having different speed and quality. All these methods require proper initialization of mask to evolve. Proper initialization of mask directly affects the performance and quality of segmentation. Different initial conditions can produces different results. This paper contains the discussion regarding the initialization of mask in any image followed by the segmentation using Chan-Vese based on curve evolution, Mumford Shah functional and level set [4]. In section II of this paper, Chan-Vese technique of active contours has been described, section III describes the method of initialization of mask as single and grid, the rest of the paper contains the comparison between the initialization as single mask and grid mask on different images of dental application.

2. IMAGE SEGMENTATION
Segmentation is the process of partitioning the image into set of distinct regions which are different in some important quantitative or qualitative way. Image segmentation based on PDEs is mainly carried out by active contour model or snakes. This method was first introduced by Kass et al in 1987 [2]. The original idea by Kass, Witkin and Terzopoulos [4] is to initialize a curve in an image and let this curve move until it adapts to the contour of a searched object.The motion of the curve is driven by the image itself. The driving force is obtained by defining a potential in the image that shall be small near objects contours. The objective is to minimize an energy which can be seen as a particular case of the minimal partition problem. But the classical snakes could detect the object with edges defined by gradient. In the level set formulation, the problem becomes a meancurvature flow-like evolving the active contour, which will stop on the desired boundary. Chan-Vese Energy function [4] is given by the following equation: Page 285

International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)

Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com Volume 2, Issue 2, March April 2013 ISSN 2278-6856
at fixed coordinates through time. Chan Vese algorithm evolves the contour via a level set method [4]. The Chan-Vese energy equation can be rewritten in terms of the level set function as equation 2 E = Energy of the curve + energy of the image. = given image. C = the boundary of the object to be detected. C = any closed curve. (2) C1 = average intensity of the image inside C. C2 = average intensity of the image outside C. where H denotes the Heaviside function and the Dirac mass. Chan-Vese algorithm consists of following steps [4]: 1. Initialize , the initial curve in the image. 2. Compute C1 and C2. , the average intensities inside the curve and outside of it. 3. Calculate the energy function from equation (1). 4. Check whether the solution is stationary, energy is minimum. If not, then repeat. Chan-Vese implemented active contours to find segmentation of the image through various methods: Active contours without edges, Active contour without edges for vector image, Active contours with multiphases. Each one is having different speed and quality. These methods could be implemented on gray scale images as well as on coloured images. The initialization of the mask plays an important role in segmentation. ChanVese has also been extended for segmentation based on Gabor texture cues [6]. A disadvantage of ChanVese is that it only provides segmentation into two phases. Extensions of ChanVese have been developed for nested segmentation curves [7] and multiphase segmentation [8], [9]. Tsai, Yezzi, and Willsky [10] developed a general level set curve evolution framework similar to ChanVese for MumfordShah segmentation and simultaneous segmentation, denoising and inpainting. We have analysed the properties of two methods of initializing masks i.e. single and grid initialization and found that they highly influence the performance of segmentation.

(1)

, , , are the parameters whose values lie between 0 and 1.The objective is to minimize the E, when E will be minimum C= C and image will be segmented. In ChanVese approach involves geometric active contour model (based upon Mumford Shah Functional) [4]. The model begins with a contour in the image plane defining an initial segmentation and then contour is evolved according to evolution equation. The goal of algorithm is to minimize this fitting energy for a given image and corresponding will define segmentation. 2.1 Level Set Function The level set, a continuous deformable model method with implicit representation, was developed by Osher and Sethian [5].It is a numerical technique for tracking interfaces and shapes of an object. It is very easy to track shapes with topological changes by the level set method. The level set methods have become very popular in the field of image segmentation due to their ability to represent boundaries of the object that changes with time or are ill defined. At any given time, the level set function simultaneously defines an edge contour and a segmentation of the image.

3. INITIALIZATION OF MASK
Figure 1 A curve C propagating in normal direction [4]. The segmentation is given by inside regions and outside region as shown in figure 1. The minimization problem requires minimizing overall set boundaries C. The level set function is normally a smooth well behaved function. Instead of tracking a curve through time, the level set method evolves a curve by updating the level set function Volume 2, Issue 2 March April 2013 An outstanding characteristic of level set methods is that contours can split or merge as the topology of the level set function changes. Therefore level set methods can detect more than one boundary simultaneously and we can also initialize the multiple contours as a grid . 3.1 Initialization as a Single Mask

Page 286

International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)

Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com Volume 2, Issue 2, March April 2013 ISSN 2278-6856
In case of single mask, we can initialize the mask as a single closed curve. It can be a circle, square and even rectangle. 3.2 Initialization as a Grid mask Masking grid as initialize mask can be created by either by the user or by built in mask creation function using mathematical morphology. The shapes of the cells on grids can be various, such as square, rectangle, diamond and disk. Mask with multiple contours with radius r has been initialized by following morphological operations: 1. Fix the value of r. 2. Initialize the structuring element. 3. Dilate the image with above structuring element. So, the initial contours in a grid are specified manually by changing value of r.. Figure 2 Block diagram for the steps of the evolution methodology.

4. THE EVOLUTION METHODOLOGY

First of all the input image is taken. The initial input image can be color or any grayscale image. We have used the intensity images to run the CHAN-VESE segmentation algorithm [4]. The next step is to initialize the initial contour in the image and then the values of CHAN-VESE parameters are set as per the requirement of the segmentation. We calculate the internal and external energy as per the energy equation and let the curve to evolve till the energy is minimized .At the point of convergence we get the segmented image. Figure. 2 shows the block diagram for evolution methodology of contour.

5. RESULTS AND ANALYSIS

We have tested the algorithm on a large number of natural images. In this paper we demonstrate results that represent our findings from these experiments. 5.1 Single vs. Grid Mask Segmentation In figures 3(a)-(c), we show three images, their segmentation results after initializing the mask as a single contour and grid contour respectively. In these experiments, the shape of the contour during grid mask initialization is disk, which is different from the circle in normal initializations .Extracting the contour of an individual tooth is a very important step in automated dental classification and identification [11]. So we have applied CHAN-VESE algorithm on dental images. We have compared the segmentation using single and grid mask initializations. As shown in figure 4(a)-(c), the original images of teeth x-ray has been taken. Figure 3 (a) Original image (b)Segmentation using single mask and (c) Segmentation using grid mask with 1=1.0, 2=1.0,=255*255*0.1,=0 The time of execution in grid mask is less as compared to single mask case .Segmentation quality is also better in grid mask initialization The results are better in case of grid mask. In grid initialization, the segmentation is almost complete in a few iterations where as in single mask initialization it takes more convergence time. We have calculated the time of convergence for mask initializations. We can also visualize the segmentation by keeping the iteration fixed. Table 1 shows the findings of these parameters. 5.2 Study of optimal number of contours in a Grid In a grid mask, multiple contours are placed on image. We have studied about the minimum number of contours in a grid that can give the best segmentation results. In figures 4(a)-(f), we show original image, their segmentation results after varying the number of contours. The results show that there is some minimum value of contours n for a particular image which gives Page 287

Volume 2, Issue 2 March April 2013

International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)

Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com Volume 2, Issue 2, March April 2013 ISSN 2278-6856
good segmentation results. We have found that further increasing the value of n gives the unstable results. Figure 5 shows that the convergence time for a particular image also varies with the variation of numbers of contours in a grid. We have fixed the number of iterations. But in some cases it converges before the specified number of iterations and we get the results in very fast manner. acceptable and reliable results. With that optimum number of contour highest performance has been shown on the basis of segmentation quality and convergence time. The decision of image segmentation quality in this paper is quite subjective.

Figure 5 Variation of Time of Execution with variation in number of contours in grid of segmentation on Teeth2 image. Table 1: Comparative analysis of Single Mask and Grid Mask Initialization for different images with the values of 1=1.0, 2=1.0, =255*255*0.2, =0

Figure 4 (a) Original image (b-f) Segmentation using 6, 9, 12, 15 and 20 contours respectively with 1=1.0, 2=1.0, =255*255*0.2, =0. For this image we get the best segmentation result in 9 contours in a grid .In this case we have kept the number of iteration fixed i.e. 4000.

6. CONCLUSION
We have analyzed the evolution of contour using single mask and grid mask initializations to assess the best approach and performed the Chan-Vese [4] segmentation on real images. The segmented images in both cases are compared on the basis of the convergence time. The evolution of contour is much faster in grid mask. It converges in a few iterations as compared to single mask initialization and gives the acceptable results which are approximate 4 times faster than the single mask initialization.. We have also studied the minimum number of contours in a grid mask that can give the best and acceptable segmentation results. For a particular image, there is some minimum number of contours i.e. optimum number of contours in a grid which gives the Volume 2, Issue 2 March April 2013

ACKNOWLEDGEMENT The author is thankful to Dr. ChanderMohan, Professor in Computer Science and Engineering Department at Ambala College of Engineering and Applied Research, Devsthali, Ambala for their valuable comments and suggestions. References [1] Rafael C. Gonzalez, Richard E. Woods, Digital Image Processing, 2nd ed., Beijing: Publishing House of Electronics Industry, 2007. [2] M. Kass, A. Witkin, and D. Terzopoulos, Snakes: active contour models, Intl J. Comp. Vis., vol. 1, no. 4, pp. 321331, 1987. [3] J. Sethian,Level Set Methods and Fast Marching Methods, Cambridge: Cambridge University Press, 1999. [4] Tony F. Chan and Luminita A Vese, Active Contours Without Edges, IEEE Transactions on Image Processing, Vol. 10 No.2, Feb 2001. Page 288

International Journal of Emerging Trends & Technology in Computer Science (IJETTCS)

Web Site: www.ijettcs.org Email: editor@ijettcs.org, editorijettcs@gmail.com Volume 2, Issue 2, March April 2013 ISSN 2278-6856
[5] S. Osher, J.A. Sethian, Fronts propagating with curvature-dependent speed: Algorithms based on Hamilton-Jacobi formulations, Journal of Computational Physics, vol. 79, no. 1, pp. 1249, 1988. [6] B. Sandberg, T.F. Chan, L.A. Vese, A Level-Set and Gabor-Based Active Contour Algorithm for Segmenting Textured Images, UCLA CAM Report 02-39, 2002. [7] G. Chung, L.A. Vese, Energy Minimization Based Segmentation and Denoising Using a Multilayer Level Set Approach, Energy Minimization Methods in Computer Vision and Pattern Recognition, vol. 3757/2005, pp. 439455, 2005. [8] L.A. Vese, T.F. Chan, A Multiphase Level Set Framework for Image Segmentation Using the Mumford and Shah Model, International Journal of Computer Vision, vol. 50, no. 3, pp. 271 293, 2002. [9] M. Keegan, B. Sandberg, T.F. Chan, A Logic Framework for Multiphase Multichannel Image Segmentation, UCLA CAM Report 10-60, 2010. [10] A.Tsai, A. Yezzi, A.S. Willsky, Curve Evolution Implementation of the Mumford Shah Functional for Image Segmentation, Denoising, Interpolation and Magnification, IEEE Transactions on Image Processing, vol. 10, no. 8, pp. 11691186, 2001. [11] Samir Shah, Ayman Abaza, Arun Ross and Hany Ammar Automatic Tooth Segmentation Using Active Contour Without Edges, Proceedings of IEEE on Biometric Consortium Conference, Baltimore MD, pp 1-6 , 2006.

Volume 2, Issue 2 March April 2013

Page 289