1. K.Gomathy, D.Arun Shunmugam M.E. / International Journal of Engineering Research and
Applications (IJERA) ISSN: 2248-9622 www.ijera.com
Vol. 3, Issue 1, January -February 2013, pp.498-502
Constrained Active Contour For Interactive Image Segmentation
K.Gomathy#1, D.Arun Shunmugam M.E.,#2
#
P.G.Scholar, Department Of Cse, P.S.R.Engineering College,Sivakasi.
#
Assistant Professor, Department Of Cse, P.S.R.Engineering College,Sivakasi.
Abstract
Interactive image segmentation categories: boundary-based approaches and region-
algortihms incorporates small amount of user based approaches.
interaction to define the desired content to be Boundary-based approaches, the user is often asked
extracted, has received much attention in the to specify an initial area that is “close” to the
recent years. We propose a robust and accurate desirable boundary.
interactive method based on the recently The "active contours" start with an
developed continuous-domain convex active initialized contour and actively deform themselves to
contour model. The proposed method exhibits the desired border while reducing the defined energy
many desirable properties of an effective in each iteration until convergence.The active
interactive image segmentation algorithm, contours/Snake method [2] proposes the top down
including robustness to user inputs and different approach instead of using previous bottom up
initializations with an efficient and light-weight approaches and it attempts to evolve an initial
solution for rendering smooth shadow boundaries contour toward the object boundary. To find a path
that do not reveal the tessellation of the shadow- between boundary seed points specified by the user
casting geometry. Our algorithm reconstructs the Dijkstra’s shortest path algorithm is applied in the
smooth contours of the underlying mesh and then methods based upon intelligent scissors [3], [4]
extrudes shadow volumes from the smooth Boundary-based approaches requires great
silhouettes to render the shadows. For this care to specify the boundary area or the boundary
purpose we propose an improved silhouette points, especially for complex shapes, most recent
reconstruction using the vertex normal of the interactive image segmentation algorithms take the
underlying smooth mesh. Then our method regional information as the input. In region-based
subdivides the silhouette loops until the contours approaches, the user is often asked to draw two types
are sufficiently smooth and project to smooth of strokes to label some pixels as foreground or
shadow boundaries. Here we solve the two background, after which the algorithm completes the
problems in a unified framework. Gradient labeling of all other pixels. The region based
controlled partial differential equation (PDE) interactive segmentation algorithms, includes
surfaces to express terrain surfaces, in which the RandomWalks based Methods [7]–[9], Graph Cut-
surface shapes can be globally determined by the based methods [5], [6], and Geodesic methods [10],
contours, their locations, and height and gradient [11]. All the above methods basically treat an image
values. The surface generated by this method is as a weighted graph with the nodes corresponding
accurate in the sense of exactly coinciding with to pixels in the image and edges being placed
the original contours and smooth with C1 between neighboring pixels, and to minimize a
(contour active convex region) continuity certain energy function on the graph for producing
everywhere. The method can reveal smooth segmentation.
saddle shapes caused by surface branching of one In the problem of interactive image
to more and can make rational interpolated sub- segmentation with the input of foreground and
contours between two or more neighbouring background strokes, which requires only a small
contours. amount of interaction from the user. First the
region-based methods are overly sensitive to small
Keywords— contour, interpolation, mesh. variations in the interactions provided by the user As
silhouette, shadow-casting, gradient in [12], the Graph Cut algorithm is sensitive to the
number of seeds, while the Random Walk and
Geodesic algorithms are sensitive to locations of
I) INTRODUCTION seeds is mainly due to the different behaviors of the
Interactive image segmentation, incorporates different energy functions.
small amount of user interaction to define the The boundaries generated by the region-
desired content to be extracted, has received much based approaches, especially those generated by RW
attention in the recent years. Already many and approaches based on the geodesic, are often
interactive image segmentation algorithms have jaggy and they do not adhere to the geometry
been proposed. In general, interactive image features present in the image additional refinement
segmentation algorithms can be classified into two step is often needed to improve the segmentation
498 | P a g e
2. K.Gomathy, D.Arun Shunmugam M.E. / International Journal of Engineering Research and
Applications (IJERA) ISSN: 2248-9622 www.ijera.com
Vol. 3, Issue 1, January -February 2013, pp.498-502
performance of the region-based methods. Most of normalizing the intensity of the individual particles
the state-of-the-art interactive image segmentation images, removing reflections, and masking portions
methods [6],[7], [9], [10] rely on additional user of images. Image preprocessing is the technique of
inputs to either globally or locally refine the enhancing data images prior to computational
boundary. However, when dealing with complex processing. The median filter is a nonlinear digital
images, the user is often required to provide a lot of filtering technique, often used to remove noise. Such
additional strokes or boundary points and thus noise reduction is a typical pre-processing step to
struggles with laborious refinement/editing. Another improve the results of later processing .Median
way for boundary refinement is to use the active filtering is very widely used in digital image
contours/Snakes model [1] to refine the initial processing because, under certain conditions, it
boundary contour produced by a region-based preserves edges while removing noise.
segmentation approach. The refinement based on
Snakes is only able to change the contour locally for B. Initialization of Contour
smoothness but the approach is incapable of Initialization of contour is done using the
evolving the entire contour to snap to geometry region based method. Based on the initialized
features/edges, and also incapable of handling contour pre segmented output is obtained. In this
topology changes of the evolving contour. using the region based method value similarity and
The mathematical tool of the new method is spatial similarity is calculated and pre segmented
the continuous-domain convex active contour model output is obtained by initializing the contour.
[14], which makes use of both the boundary and the
regional information to find a global “optimal” C. Convex Active Contour Model
solution. Continuous-domain convex methods have The convex active contour model consists
started to receive attention since they avoid the of two terms: a regional term formulation and
inherent grid bias in all discrete graph-based boundary term formulation.
methods, and also have fast and global numerical
solvers through convex optimization [15], [16]. 1) Regional Term Formulation:
However, the convex active contour model so far The foreground and background seeds give
has mainly been applied for automatic image an excellent description about the color distributions
segmentation, which often results in over- of the foreground and background regions.
segmentation with trivial solutions for complex Foreground/background Gaussian mixture models
images [14], [16]. (GMMs) introduced in [20] are estimated from
foreground/background seeds, and used to represent
The major contribution of this includes the the color distributions of the foreground and
following. background regions. Let Pr(x|F) and Pr(x|B) denote
The powerful continuous-domain convex the probabilities that pixel x denotes the foreground
active contour with one of the region based and background GMMs, respectively.
methods, Geodesic/random walk where the region- PF (x) = −log Pr(x|F)/−log Pr(x|F) − log Pr(x|B)
based method is used in the first step to generate an and
initial contour, and the convex active contour is then PB(x) = −log Pr(x|B)/−log Pr(x|F) − log Pr(x|B).
applied to optimize the contour. Such integration (2)
utilizes the seed propagation and the location
features introduced by Geodesic/Random Walk to We incorporate this regional information derived
reduce the possible “small cut” problem in the from foreground/background strokes into the
convex active contour, and also the powerful contour regional term of the
evolving capability provided by the convex active convex active contour model as hr (x) = PB(x) − PF
contour model to absorb the non robustness of the (x). (3)
region-based approaches. Then our method This definition of hr ensures that the active
subdivides the silhouette loops until the contours are contour evolves toward the one complying with the
sufficiently smooth and project to smooth shadow known GMM models. For instance, for a pixel x, if
boundaries. PB(x) > PF (x) (respectively, PB(x) < PF (x)) and
PB(x)− PF (x) is positive (respectively, negative),
II) CONSTRAINED ACTIVE CONTOUR u(x) tends to decrease (respectively,increase) during
MODEL the contour evolution in order to minimize (1),which
In this we describe the continuous domain can lead to u(x) ≤ T (respectively, u(x) > T) and the
convex active contour model which extends the classification of the pixel belonging to the
convex active contour model. background (respectively, the foreground).
The hr definition of (3)fails in the case that
A. Preprocessing the foreground and background color models are not
Preprocessing images commonly involves well separated. Thus, to avoid this problem and also
removing low-frequency background noise, to make use of the segmentation result obtained by
499 | P a g e
3. K.Gomathy, D.Arun Shunmugam M.E. / International Journal of Engineering Research and
Applications (IJERA) ISSN: 2248-9622 www.ijera.com
Vol. 3, Issue 1, January -February 2013, pp.498-502
the Geodesic algorithm in step 1, we further propose matte alpha is computed. With these components we
to incorporate the probability map P(x) into the can paste the object onto a new Background if
region term hr as desired with no noticeable visual artifacts by the
hr (x) = α(PB(x) − PF (x)) + (1 − α)(1 − 2P(x)) (4) simple matting equation.
where α, α ∈ [0, 1], is a tradeoff factor. The E. Silhouette Reconstruction
second term (1−2P(x)) in (10) prevents the refined A silhouette is the image of a person, an
contour drifting too far apart from the initial object or scene represented as a solid shape of a
segmentation. Specifically, when P(x) > 0.5 and (1 − single color, usually black, its edges matching the
2P(x)) are negative, u(x) tends to increase in order to outline of the subject. The interior of a silhouette is
minimize (1), which favors classifying the pixel as a basically featureless, and the whole is typically
foreground pixel, and vice versa. presented on a light background, usually white, or
none at all.
In addition, it can be observed that when hr The silhouette differs from an outline which
(x) → +∞(respectively, hr (x) → −∞), the regional depicts the edge of an object in a linear form, while a
term forces u(x) = 0 (respectively, u(x) = 1) to silhouette appears as a solid shape. Silhouette
minimize This observation allows us to enforce images may be created in any visual artistic media.
some hard constraints in the contour evolution Our algorithm reconstructs the smooth
process. In particular, for those pixels that have no contours of the underlying mesh and then extrudes
ambiguity in classification, including the pixels shadow volumes from the smooth silhouettes to
lying on the foreground/background strokes and the render the shadows. For this purpose we propose an
pixels having very large or very small P(x) values improved silhouette reconstruction using the vertex
(P(x) > 0.9 or P(x) < 0.1), we treat them as hard normal of the underlying smooth mesh. Then our
constraints in the contour evolution process. We method subdivides the silhouette loops until the
directly assign a negative hr value and a positive hr contours are sufficiently smooth and project to
value, both with extremely large magnitude, to these smooth shadow boundaries. Here we solve the two
confirmed foreground and background pixels, problems in a unified framework. Gradient
respectively. In this way, we guarantee that the controlled partial differential equation (PDE)
refined result complies with the user input and also surfaces to express terrain surfaces, in which the
exploit more information from the initial surface shapes can be globally determined by the
segmentation result. contours, their locations, and height and gradient
values. The surface generated by this method is
2) Boundary Term Formulation: accurate in the sense of exactly coinciding with the
The boundary term of ∫Ω gb(x)|∇ u| dx in original contours and smooth with C1 (contour
(1) is essentially a weighed total variation of active convex region) continuity everywhere. The
function u, where the weight gb plays an important method can reveal smooth saddle shapes caused by
role.The definition of gb in is effective in the sense surface branching of one to more and can make
that it encourages the segmentation along the curves rational interpolated sub-contours between two or
where the edge detection function is minimal. The more neighboring contours.
problem with is that at locations with weak edges the
boundary is likely to be smoothed out. Thus, in this III) CONCLUSION
paper, we propose to incorporate the GMM In this paper, we have proposed a robust
probability map PF (x) to enhance the edge and accurate interactive image segmentation method
detection. Particularly, we define gb as based on the continuous domain convex active
contour model. We have demonstrated that our
gb = β · gc + (1 − β) · ge (6) method outperforms the state-of-the-art interactive
segmentation methods. It exhibits many desirable
where gc and ge are the results of applying properties for a good segmentation tool, including
the edge detection to the GMM probability map PF the robustness to user inputs and different
(x) and the original image, respectively, and β, β ∈ initializations, the ability to produce a smooth and
[0, 1], is a tradeoff factor computed in a similar way accurate boundary contour, and the ability to handle
as α given above formula. topology changes. In this we also proposed
improved silhouette reconstruction for handling
D. Constrained Active Contour Model sophisticated shapes
Based on the above calculated values the
constrained active contour model is constructed, IV) FUTURE WORK
based on the calculated values Compute the alpha This paper can be extended in a few ways.
channel inside the band, once distance is obtained. For example, it might be beneficial to apply the
Estimate Foreground and Background components continuous-domain convex active contour model for
in Luv space for each pixel inside the band, after
500 | P a g e
4. K.Gomathy, D.Arun Shunmugam M.E. / International Journal of Engineering Research and
Applications (IJERA) ISSN: 2248-9622 www.ijera.com
Vol. 3, Issue 1, January -February 2013, pp.498-502
other segmentation problems, such as image matting
or video segmentation.
V) RESULTS
501 | P a g e
5. K.Gomathy, D.Arun Shunmugam M.E. / International Journal of Engineering Research and
Applications (IJERA) ISSN: 2248-9622 www.ijera.com
Vol. 3, Issue 1, January -February 2013, pp.498-502
REFERENCES 13. F. Chan, F. Lam, P. Poon, H. Zhu, and K.
1. Thi Nhat Anh Nguyen, Jianfei Cai, Senior Chan, “Object boundary location by region
Member, IEEE, Juyong Zhang, and Jianmin and contour deformation,” IEE Proc. Vis.,
Zheng “Robust Interactive Image Image Signal Process., vol. 143, no. 6, pp.
Segmentation Using Convex Active 353–360, Dec. 2002.
Contours” IEEE trans.,image process, vol. 14. X. Bresson, S. Esedo glu, P.
21, no. 8, august 2012 Vandergheynst, J.-P. Thiran, and S. Osher,
2. M. Kass, A. Witkin, and D. Terzopoulos, “Fast global minimization of the active
“Snakes: Active contour models,” Int. J. contour/snake model,” J. Math. Imaging
Comput. Vis., vol. 1, no. 4, pp. 321–331, Vis., vol. 28, no. 2, pp. 151–167, 2007.
1988. 15. Nguyen Thi Nhat Anh, Jianfei Cai, Juyong
3. E. Mortensen and W. Barrett, “Interactive Zhang, Jianmin Zheng “Constrained Active
segmentation with intelligent scissors,” Contours for Boundary Refinement in
Graph. Models Image Process., vol. 60, no. Interactive Image Segmentation” in IEEE
5, pp. 349–384,Sep. 1998. International Symposium on Circuits And
4. A. Falcao, J. Udupa, and F. Miyazawa, “An Systems (ICSAS) may2012,PP-870-873.
ultrafast user-steered image segmentation 16. T. Chan, S. Esedoglu, and M. Nikolova,
paradigm: Live wire on the fly,” IEEE “Algorithms for finding global minimizers
Trans. Med. Imag., vol. 19, no. 1, pp. 55– of image segmentation and denoising
62, Jan. 2002. models,” SIAM J. Appl. Math., vol. 66, no.
5. Y. Boykov and M. Jolly, “Interactive graph 5, pp. 1632–1648, 2006.
cuts for optimal boundary and region 17. T. Goldstein, X. Bresson, and S. Osher,
segmentation of objects in N-D images,” in “Geometric applications of the split
Proc. IEEE Int. Conf. Comput. Vis., vol. 1. Bregman method: Segmentation and
Vancouver, BC, Canada, Jul. 2001, pp. surface reconstruction,” J. Sci. Comput.,
105– 112. vol. 45, nos. 1–3, pp. 272–293, 2010.
6. C. Rother, V. Kolmogorov, and A. Blake, 18. V. Caselles, R. Kimmel, and G. Sapiro,
“Grabcut: Interactive foreground extraction “Geodesic active contours,” Int.J. Comput.
using iterated graph cuts,” ACM Vis., vol. 22, no. 1, pp. 61–79, 1997.
SIGGRAPH, vol. 23, no. 3, pp. 1–6, 2004. 19. T. Chan and L. Vese, “Active contours
7. L. Grady, “Random walks for image without edges,” IEEE Trans. Image
segmentation,” IEEE Trans. Pattern Anal. Process., vol. 10, no. 2, pp. 266–277, Feb.
Mach. Intell., vol. 28, no. 11, pp. 1768– 2001.
1783, Nov. 2006. 20. Y. Wang, W. Yin, and Y. Zhang, “A fast
8. J. Zhang, J. Zheng, and J. Cai, “A diffusion algorithm for image deblurring with total
approach to seeded image segmentation,” in variation regularization,” Dept. Comput.
Proc. IEEE Comput. Vis. Pattern Recognit., Appl. Math., Rice Univ., Houston, TX,
San Francisco, CA, Jun. 2010, pp. 2125– Tech. Rep. TR07-10, 2007.
2132. 21. C. Yang, R. Duraiswami, N. Gumerov, and
9. W. Yang, J. Cai, J. Zheng, and J. Luo, L. Davis, “Improved fast Gauss transform
“User-friendly interactive image and efficient kernel density estimation,” in
segmentation through unified combinatorial Proc. IEEE Int. Conf. Comput. Vis., vol. 1.
user inputs,” IEEE Trans. Image Process., Nice, France, 2003, pp. 664–671.
vol. 19, no. 9, pp. 2470–2479, Sep. 2010. 22. A. Blake, C. Rother, M. Brown, P. Pérez,
10. X. Bai and G. Sapiro, “A geodesic H. Philip, and S. Torr,“Interactive image
framework for fast interactive image and segmentation using an adaptive GMMRF
video segmentation and matting,” in Proc. model,” in Proc. Eur. Conf. Comput. Vis.,
IEEE Int. Conf. Comput. Vis., Rio de Cambridge, U.K., 2004, pp. 428–441.
Janeiro, Brazil, Oct. 2007, pp. 1–8. 23. O. Duchenne, J.-Y. Audibert, and R.
11. A. Criminisi, T. Sharp, and A. Blake, Keriven,
“Geos: Geodesic image segmentation,” in 24. “Segmentation by transduction,”in Proc.
Proc. Eur. Conf. Comput. Vis., Cambridge, IEEE Comput. Vis. Pattern Recog.,
U.K., 2008, pp. 99–112. Anchorage, AK, Jun. 2008, pp. 1–8.
12. A. Sinop and L. Grady, “A seeded image 25. K. McGuinness and N. E. O’Connor,
segmentation framework unifying graph “Toward automated evaluation of
cuts and random walker which yields a new interactive segmentation,” Comput. Vis.
algorithm,” in Proc. IEEE Int. Conf. Image Understand., vol. 115,no. 6, pp.
Comput. Vis., Rio de Janeiro, Brazil, Oct. 868–884, 2011.
2007, pp. 1–8.
502 | P a g e