## Region Competition and its Analysis: A Unified Theory for Image Segmentation (1995)

Citations: | 2 - 0 self |

### BibTeX

@TECHREPORT{Zhu95regioncompetition,

author = {Song Chun Zhu and A. L. Yuille},

title = {Region Competition and its Analysis: A Unified Theory for Image Segmentation},

institution = {},

year = {1995}

}

### OpenURL

### Abstract

We present a novel statistical and variational approach to image segmentation based on a new algorithm named region competition. This algorithm is derived by minimizing a generalized Bayes/MDL criterion using the variational principle. The algorithm is guaranteed to converge to a local minimum and combines aspects of snakes/balloons and region growing. Indeed the classic snakes/balloons and region growing algorithms can be directly derived from our approach. We provide theoretical analysis of region competition including accuracy of boundary location, criteria for initial conditions, and the relationship to edge detection using filters. It is straightforward to generalize the algorithm to multiband segmentation and we demonstrate it on grey level images, color images and texture images. The novel color model allows us to eliminate intensity gradients and shadows, thereby obtaining segmentation based on the albedos of objects, and it also helps detect highlight regions. A short version ...

### Citations

3723 |
Stochastic relaxation, Gibbs distributions, and Bayesian restoration of images
- Geman, Geman
- 1984
(Show Context)
Citation Context ...(ii) Snake [20] and Balloon methods [6], [7], [43], [38] (iii) Region growing and merging techniques [2], [1], [28], and (iv) Global optimization approaches based on energy functions [32] or Bayesian =-=[11]-=-, [3], [12] and MDL (Minimum Description Length) criteria [24] [21] [19]. A common property of these approaches is that they all make hypotheses about the image, test features and make decisions by ap... |

3098 | A computational approach to edge detection
- Canny
- 1986
(Show Context)
Citation Context ...roblem of early vision and it has been intensively studied. Approaches to image segmentation can be roughly classified into four groups: (i) Local filtering approaches such as the Canny edge detector =-=[4]-=-, (ii) Snake [20] and Balloon methods [6], [7], [43], [38] (iii) Region growing and merging techniques [2], [1], [28], and (iv) Global optimization approaches based on energy functions [32] or Bayesia... |

3082 | Snakes: active contour models
- Kass, Witkin, et al.
- 1988
(Show Context)
Citation Context ...vision and it has been intensively studied. Approaches to image segmentation can be roughly classified into four groups: (i) Local filtering approaches such as the Canny edge detector [4], (ii) Snake =-=[20]-=- and Balloon methods [6], [7], [43], [38] (iii) Region growing and merging techniques [2], [1], [28], and (iv) Global optimization approaches based on energy functions [32] or Bayesian [11], [3], [12]... |

2650 |
Introduction to Statistical Pattern Recognition
- Fukunaga
- 1990
(Show Context)
Citation Context ...e. n 2 = 1 and n 1 is very large (say n 1 ? 100). In this case we can treat �� =s�� 1 ; oe 2 =soe 2 1 , and �� 2 = I (x;y) (the intensity at point (x; y)), and approximate the squared Fish=-=er distance [9] by:-=- (I \Gamma��) 2 oe 2 . A variant of region growing [28] is to fit the intensity within each region to a parameterized model, such as a plane or a quadratic form. Then tests like Equation (4) can b... |

737 | Visual Reconstruction
- Blake, Zisserman
- 1987
(Show Context)
Citation Context ...nake [20] and Balloon methods [6], [7], [43], [38] (iii) Region growing and merging techniques [2], [1], [28], and (iv) Global optimization approaches based on energy functions [32] or Bayesian [11], =-=[3]-=-, [12] and MDL (Minimum Description Length) criteria [24] [21] [19]. A common property of these approaches is that they all make hypotheses about the image, test features and make decisions by applyin... |

459 | On active contour models and balloons
- Cohen
- 1991
(Show Context)
Citation Context ...vely studied. Approaches to image segmentation can be roughly classified into four groups: (i) Local filtering approaches such as the Canny edge detector [4], (ii) Snake [20] and Balloon methods [6], =-=[7]-=-, [43], [38] (iii) Region growing and merging techniques [2], [1], [28], and (iv) Global optimization approaches based on energy functions [32] or Bayesian [11], [3], [12] and MDL (Minimum Description... |

362 |
Seeded Region Growing
- Adams, Leanne
- 1994
(Show Context)
Citation Context ...ssified into four groups: (i) Local filtering approaches such as the Canny edge detector [4], (ii) Snake [20] and Balloon methods [6], [7], [43], [38] (iii) Region growing and merging techniques [2], =-=[1]-=-, [28], and (iv) Global optimization approaches based on energy functions [32] or Bayesian [11], [3], [12] and MDL (Minimum Description Length) criteria [24] [21] [19]. A common property of these appr... |

222 | Constructing Simple Stable Descriptions for Image Partitioning
- Leclerc
- 1989
(Show Context)
Citation Context ... Region growing and merging techniques [2], [1], [28], and (iv) Global optimization approaches based on energy functions [32] or Bayesian [11], [3], [12] and MDL (Minimum Description Length) criteria =-=[24]-=- [21] [19]. A common property of these approaches is that they all make hypotheses about the image, test features and make decisions by applying thresholds explicitly or implicitly. As shown by the sh... |

170 | A Physical Approach to Color Image Understanding
- Klinker, Shafer, et al.
- 1990
(Show Context)
Citation Context ...sity properties. But the image intensity is a function both of the material properties of the object, its albedo, and the surface orientation and lighting. We assume the dichromatic reflectance model =-=[22]-=- 11 : I(; ~r i ) = ae(; ~r s )F (~v; ~n; ~s)E(; ~n; ~s) + noise(~r i ) (27) where I is the image as function of the image position ~r i and wavelength , ae is the surface albedo as a function ofsand s... |

68 | A multiscale algorithm for image segmentation by variational method
- Koepfler, Lopez, et al.
- 1994
(Show Context)
Citation Context ...between regions. By contrast, our method uses region growing as a way of minimizing the MDL criterion for the entire image directly. Our work also has some similarities to ideas discussed in [32] and =-=[23]-=- [31]. These works differ from ours by not using a statistical criterion. This paper is organized as follows. Section (2) sets the scene by briefly describing Snakes and Balloons, Region Growing and E... |

57 |
Segmenting images using normalized color
- Healey
- 1992
(Show Context)
Citation Context ...tractable when we have color images as input, and it can be assumed that the geometric reflectance F is the same for all color bands. Color will allow us to segment regions with constant albedo [22], =-=[16]-=-, [40]. 6.1 The Color Model If we include specularity in the image formation process then the color dichromatic reflection model [22] is: I(; ~r i ) = ae(; ~r s )F b (~v; ~n; ~s)E(; ~n; ~s) + F s (~v;... |

55 |
Another interpretation of the EM algorithm for mixture distributions
- Hathaway
- 1986
(Show Context)
Citation Context ...e the probability P (I (~v) jff 1 ) is calculated and compared with 6 This two step iteration process is similar to the well-known EM algorithm which can be reexpressed as two steepest descent stages =-=[15]. th-=-e absolute threshold P 0 . Fisher's test [39] and the �� 2 test [28] correspond to particular choices of P (I (~v) jff 1 ). Secondly, the underlying statistical assumption for the balloon/snake mo... |

53 |
A finite element method applied to new active contour models and 3d reconstruction from cross sections
- Cohen, Cohen
- 1990
(Show Context)
Citation Context ...tensively studied. Approaches to image segmentation can be roughly classified into four groups: (i) Local filtering approaches such as the Canny edge detector [4], (ii) Snake [20] and Balloon methods =-=[6]-=-, [7], [43], [38] (iii) Region growing and merging techniques [2], [1], [28], and (iv) Global optimization approaches based on energy functions [32] or Bayesian [11], [3], [12] and MDL (Minimum Descri... |

32 | et.al. Deformable boundary finding influenced by region homogeneity. Available at the web site of the Yale - Chakraborty - 1994 |

25 |
Compact Region Extraction Using Weighted Pixel Linking in a Pyramid
- Hong, Rosenfeld
- 1984
(Show Context)
Citation Context ...n tries to maximize its area while smoothing its bounding contour and maximizing the intensity gradient along the contour. 2.2 Region Growing and Merging The goal of region merging and region growing =-=[17], [28-=-] is to divide the domain R of the image I into regions fR i : i = 1; :::; Mg so that R = [ M i=1 R i , R i " R j = OE if 1 The contour can also be represented parametrically, for example by B-sp... |

12 |
Foundations of Applied Mathematics
- Greenberg
- 1978
(Show Context)
Citation Context ... @y \Gamma d ds @L @sy = (\Gamma @Q @x + @P @y )sx = \Gammaf (x; y)sx (44) 13 Green's thereom is a special case of Stokes theorem and should not be confused with Green's identities. See, for example, =-=[14]-=-. Since ~ n (x;y) = (sy; \Gammasx) is the normal along the contour \Gamma, let ~v = (x; y), thus we have: ffiE ffiv = f(x; y)~ n (x;y) (45) To compute the functional derivative of an integral along th... |

8 |
Map Representations and Coding-Based Priors for Segmentation. CVPR
- Keeler
- 1991
(Show Context)
Citation Context ...on growing and merging techniques [2], [1], [28], and (iv) Global optimization approaches based on energy functions [32] or Bayesian [11], [3], [12] and MDL (Minimum Description Length) criteria [24] =-=[21]-=- [19]. A common property of these approaches is that they all make hypotheses about the image, test features and make decisions by applying thresholds explicitly or implicitly. As shown by the shadowe... |

6 |
Modeling light reflection for color computer vision
- Lee, Breneman, et al.
- 1990
(Show Context)
Citation Context ...oes not change the spectral composition of the illumination. 11 This equation is incorrect for some materials such as metals and some dielectrics. This model has been tested over a range of materials =-=[26]-=- and is shown to be a good approximation in many cases, though it fails for some dielectrics. If we assume red, green, and blue color bands, then we can simplify Equation (28) at each point (x; y) to:... |

2 | A.P.Pentland, "Cooperative Robust Estimation Using Layers of Support - Darrell - 1994 |

2 |
A.L.Yuille. "A common framework for image segmentation
- Geiger
- 1991
(Show Context)
Citation Context ...region. It is usually very difficult to minimize the energy functions resulting from Bayes or MDL. Algorithms such as simulated annealing [11], graduated non-convexity [3] and deterministic annealing =-=[10]-=- are perhaps the most successful. 3 Region Competition for Grey Level Images In this section, we first derive our region competition algorithm from a global optimization criterion. Then we show how th... |

2 |
P.Dong. "Boundary Detection by Constrained Optimization
- Geman
- 1990
(Show Context)
Citation Context ...[20] and Balloon methods [6], [7], [43], [38] (iii) Region growing and merging techniques [2], [1], [28], and (iv) Global optimization approaches based on energy functions [32] or Bayesian [11], [3], =-=[12]-=- and MDL (Minimum Description Length) criteria [24] [21] [19]. A common property of these approaches is that they all make hypotheses about the image, test features and make decisions by applying thre... |

2 |
The Heat Equation Shrinking Convex Plane Curves
- Grayson
- 1987
(Show Context)
Citation Context ...it is equal to the following heat diffusion equation. d\Gamma dt = @ 2 \Gamma @s 2 (14) where s is the arc length of the curve \Gamma. Detailed discussions of properties of this equation are given in =-=[13]-=- and [35]. Besides the smoothing term, the motion of ~v is determined by the likelihood ratio test. If P (I (~v) jff i ) ? P (I (~v) jff j ) -- i.e. if the intensity at ~v fits better to the distribut... |

2 |
Applied Multivariate Statsitical Analysis
- Johnson, Wichem
- 1982
(Show Context)
Citation Context ... 2 2 = noe 2 n 1soe 2 1 + n 2soe 2 2 \Gamma 1; (4) where n = n 1 + n 2 andsoe 2 is the sample variance of the mixture region (a generalization to the multi-dimensional case is called Hotelling's test =-=[18]-=-). If this statistic is below a certain threshold then the regions are merged. Region growing can be considered as a special case of region merging, where R 1 is the growing region and R 2 is a single... |

2 |
B.Dom, W.Niblack and D.Steele. "A fast algorithm for MDL-based multi-band image segmentation
- Kanungo
- 1994
(Show Context)
Citation Context ...owing and merging techniques [2], [1], [28], and (iv) Global optimization approaches based on energy functions [32] or Bayesian [11], [3], [12] and MDL (Minimum Description Length) criteria [24] [21] =-=[19]-=-. A common property of these approaches is that they all make hypotheses about the image, test features and make decisions by applying thresholds explicitly or implicitly. As shown by the shadowed are... |