ladabu

Embed Size (px)

Citation preview

  • 8/12/2019 ladabu

    1/6

    Active contours with shape priors

    50

    Once this minimum is achieved, the mean computed over the transformed shapes is

    called the empirical mean. An empirical covariance is then computed from it. A

    Principal Component Analysis (PCA) is applied to generate the characteristic Eigen

    modes of deformations.

    The shape prior term is modeled as the search for the optimum rigid transformations

    that minimize a distance between evolving contour and the characteristic deformation

    mode. The authors have suggested that the empirical mean can also be used as prior.

    In that case the shape prior term will try to reduce the distance between the evolving

    contour and the empirical mean by finding the rigid transforms that generates the best

    fit. The difference between the two formulations is that, using empirical mean as shape

    statistics will result in more rigid priors, while by using the characteristic Eigen modes

    statistics the shape priors will allow for variations and will result in more flexible

    priors.

    x Fourier-based shape descriptors provide quite an efficient and powerful way of

    contour representation. Such a representation can be particularly useful in the context

    of explicit active contours with shape priors. In this regard, one method (Staib &

    Duncan 1992) proposes the contour representation using elliptical Fourier descriptors.

    The prior shapes are considered as a set of outline boundaries corresponding to each

    shape prior class. For each class, the shapes are defined in terms of the elliptical

    Fourier descriptor parameters. Using the mean and variance of these parameters, for

    each class, a Gaussian probability density function is defined. To embed the shape

    prior information in the curve evolution, first the shape of the curve has to be matched

    to one of the existing classes of shapes. This is done by expressing the evolving curve

    also as elliptical Fourier descriptor parameters and estimating the maximum a

    posteriori (minimum error). This is accomplished by comparing the parameters of

    evolving curve with the Gaussian approximation of shape distribution for existing

    shape classes. The class which increases the maximum a posteriori will be chosen and

    the shape constraints of the particular class will be enforced by using the elliptical

    descriptors of that class and adjusting them to the pose and scale of evolving curve. To

    attract the evolving curve to the image information, the image gradient is used. This

    method is very sensitive to initial shape of the deformable curve.

  • 8/12/2019 ladabu

    2/6

    Active contours with shape priors

    51

    x Recently, Fourier based geometric shape priors have been used with the variational

    setup for snakes (Charmi et al. 2008). Both the template and the deformable model are

    represented by a set of Fourier descriptors. A force based approach is then applied to

    guide the deformable contour towards the template by minimizing the difference

    between the reconstructions of the template and the deformable curve. However, this

    method is sensitive to the starting point and needs the reconstruction of the template

    and deformable contours in the spatial domain in order to compute the force that

    guides the deformable model towards the template.

    x /HJHQGUHV PRPHQWV are used as shape priors in region-based active contours

    (Foulonneau et al. 2006). This is achieved by minimizing a distance between shape

    descriptors GHILQHG E\ /HJHQGUHV PRPHQWV DQG WKDW RI HYROYLQJ FRQWRXU . One main

    limitation of this method is that the shape priors are not invariant to rotation changes.

    The method can be theoretically extended for rotational invariance, but this leads to

    the increase in complexity by many folds.

    x In one recent technical report (Park 2010), invariant shape priors have been added in

    the polygonal implementation of Mumford Shah formulation. To embed the shape

    prior information, the reference shape is represented in the form of the inter-vertex

    distance. The inter-vertex distance of a polygon is a matrix in which each row

    corresponds to the distance of a vertex from all the other vertices. In its crude form,

    the inter-vertex distance description is invariant to the translation.

    The prior information is added in the curve evolution process by computing the inter-

    vertex distance for the deforming curve and comparing it to that of the reference

    shape. The shape prior term is then modeled as the minimization of the distance

    between inter-vertex representation of the reference shape and the evolving curve.

    Both the reference shape and deforming curve must have the same number of points in

    order for this formulation to work. The best scale and rotation for prior shape is

    estimated during the evolution. The scale is derived from the scale of the evolving

    contour. To achieve the rotation invariance and estimate the correct orientation of the

    object to be segmented, the reference shape is dynamically rotated during each step to

    find the best fit between the orientation of the evolving contour and that of the prior

    shape. The best fit is the orientation that minimizes the prior energy term.

  • 8/12/2019 ladabu

    3/6

    Active contours with shape priors

    52

    To adjust the influence of the shape prior term, a weight parameter is associated with

    it. For the evolution under shape prior, the authors, propose two schemes to balance

    the influence of the prior shape term with respect to the other terms of the Mumford

    Shah model. As a first strategy, they suggest that an initial segmentation should be

    carried out without using the prior energy terms. This segmentation can be then used

    as an initialization step for the model with shape prior energy. As a second strategy,

    the authors propose gradual increase in the parameter of shape energy term during the

    curve evolution.

    One limitation of this method is that during the evolution, the evolving contour cannot

    be resampled (i.e. vertices cannot be added or deleted). This limitation comes from the

    fact that inter-vertex distance matrices of reference shape and evolving contour needs

    to be of the same size for prior term calculation. Therefore, the number of vertices of

    reference and evolving contour needs stay the same during the entire process. The

    other limitation is that the inter-vertex distance based description can only be applied

    to convex polygonal shapes.

    In this context we propose an improved version of the greedy algorithm of explicit active

    contour models with shape priors. These priors enable the deformable model to converge

    towards the desired shape even in the presence of occlusion and context noise. We introduce

    these shape priors through the use of stable and complete Fourier descriptors (Bartolini et al.

    2005). These priors are invariant to the translation, scaling, rotation factors and starting point.

    We will present this contribution in its full details in chapter 2 of this thesis.

    Here, we will present a brief discussion regarding our choice of active contour model and

    shape descriptors. The greedy algorithm was chosen because it is fast and it achieves better

    segmentation performance when compared to the variational and dynamic programming

    approaches. Moreover, it is more stable (Kim & Lee 2003). Another motivation for the use of

    the greedy algorithm came from the fact that its energy based setup is quite intuitive for

    adding more energy terms. Thus, we can introduce shape based information in greedy

    algorithm through one of such energy terms. The shape prior energy term tries to minimize

    the distance between the shape of the evolving contour and a reference contour. For the

    description of these contours, a stable set of Fourier descriptors (Bartolini et al. 2005) is used.

    The choice of Fourier descriptors was made because they are very effective for representing

    contours and their efficiency in contour-based shape matching is high when compared to the

  • 8/12/2019 ladabu

    4/6

  • 8/12/2019 ladabu

    5/6

    2 $Q DFWLYH FRQWRXU PRGHO ZLWK LPSURYHG VKDSH S

    XVLQJ )RXULHU GHVFULSWRUV

    The segmentation results by active contour methods suffer seriously when there is occlusion,

    context noise and object overlapping or missing parts of the object. The main objective of this

    thesis is to improve the segmentation results of active contours in case of such problems. Our

    hypothesis is that if some prior information about the shape of the object to be segmented is

    available then embedding such information in the active contour model can improve the

    segmentation results. We base our hypothesis on the notion that any missing or occluded parts

    RI WKH REMHFWV ERXQGDU\ FDQ EH UHFRYHUHG XVLQJ VXFK SULRU LQIRUPDWLRQ ,Q WKH F

    work, we address some difficulties that arise when introducing the shape priors in the active

    contour models. These include: which shape descriptors should be used? How and when to

    integrate the shape priors in the energy equation? How to achieve the segmentation with

    shape priors in a computationally efficient manner? How to balance the influence of the prior

    based energy in the presence of other energy terms?

    In this chapter, we will be presenting our main contribution by adding such shape priors to the

    segmentation process of active contours to improve the segmentation results.

    We propose to add shape prior based energy term in the greedy algorithm of active contours.

    Such an energy term should constraint the deformation of active contour with respect to a

    prior shape but should at the same time balance the influence of the shape prior term in the

    presence of other energy terms. This energy term should be able to guide the segmentation

    with respect to a prior shape by minimizing a distance between the prior shape and the

    evolving contour. The greedy algorithm was chosen because of its speed, stability and its

    adaptability for permitting the integration of additional energy terms quite intuitively.

    To introduce any such energy term, we need a robust and compact shape description method

    that is invariant to at least the basic transformations such as translation, rotation and scaling

    factors. In this regard, many shape descriptors have been discussed in chapter 1, based on

    which Fourier based shape descriptors have been chosen as the shape description method.

    This choice was made on multiple criteria. Firstly, the Fourier based descriptors are naturally

  • 8/12/2019 ladabu

    6/6