SEARCHING AND FITTING STRATEGIES IN ACTIVE SHAPE

MODELS

Jianhua Zhang

College of Information Engineering, Zhejiang University of Technology, Hangzhou, China

S. Y. Chen

Department of Informatics, University of Hamburg, Germany

Sheng Liu, Qiu Guan, Haihong Wu

College of Information Engineering, Zhejiang University of Technology, Hangzhou, China

Keywords: Active Shape Model, shape parameter, self-adjustment, active searching, image fitting, outlying points.

Abstract: The Active Shape Model (ASM) is an ever-increasingly important method for object modelling, shape

recognition, and image localization. However, when the target is not clarity, or the initial model placed very

far from the target, ASM may have problem to locate on an acceptable result. In this paper, a new strategy is

proposed on the ASM searching and fitting procedure, which forms an active searching method. Using this

new strategy, the influence of the clarity of the target and initial position of the model is reduced and the

result of the fitting is more accuracy. Experiments and results show that the new strategy are effective for

improving the performance of the image fitting.

1 INTRODUCTION

The Active Shape Model (ASM), firstly proposed by

Cootes et.al.

(Cootes et.al, 1995), is an important

method for modelling of a deformable model, image

fitting, shape recognition, and shape localization.

For the traditional ASM, it performs successfully

when the target image is clear and the initial position

of the model closes to the target. However, when the

target is not clarity, for example, the beard in a facial

image or a weak boundary in medical images, or the

initial model placed very far from the target, ASM

may have problem to locate on an acceptable result.

Fig. 1 illustrates a few situations of these scenes.

Figure 1: Problems occurred in ASM fitting to an image

(the left is the initial pose and the right is the fitting result).

Some improvements were put forward to ASM in

recent years. ( Bruijne et al. 2001,Wan et al.

2005,Wang et al. 2002). However, ASMs still need

an appropriate initial position and clear target

images. Li and Chutatape presented the

self-adjusting weight in the transformation from

389

Zhang J., Y. Chen S., Liu S., Guan Q. and Wu H. (2007).

SEARCHING AND FITTING STRATEGIES IN ACTIVE SHAPE MODELS.

In Proceedings of the Fourth International Conference on Informatics in Control, Automation and Robotics, pages 389-392

DOI: 10.5220/0001651403890392

 SciTePress

shape space to image space and the exclusion of

outlying points in obtaining shape parameters (Li

and Chutatape 2003). This modification is more

robust and converge faster than the conventional

ASM in the case where the edge of optic disk is

weak or occluded by blood vessels, but it is uneasy

to extend to other cases.

In this paper, a novel method is presented that

avoid the influence of the targe and initial position of

model by using the minimize square error(MSE)

obtained by an equation as defined like (Cootes et

al.1995):

))(())(()(

ydhCydhdf

−−=

−

(1)

If the MSEs of some points are too large, we

ignore these points when the shape and pose

parameters are attained. The organization of the

remainder paper is as follows. Section 2 outlines the

standard ASM procedure. Some new strategies are

developed and described in Section 3. Practical

experimental and results are given in Section 4 and a

conclusion is followed in Section 5.

2 THE ACTIVE SHAPE MODEL

ASM have been helpful for image fitting, shape

recognition, and shape localization. Because models

are built by the training set, instance of an ASM can

only deform in the ways found in its training set. The

local structures are also considered by ASM through

modelling the gray-level information of each

landmark.

2.1 The Point Distribution Model

(PDM)

PDM is built to describe both typical shape and

typical variability by disposing a training set which

we have chosen. Firstly, landmarks are labelled on

each image in the training set by hand.

After all images in the training set have been

labeled, they must be aligned with respect to a set of

axes. The required alignment is achieved by scaling,

rotating, and translating the training shapes in order

that they agree as closely as possible.

And then we capture the statistics of the set of

aligned shape. In previous steps, labeling and

aligning shapes, the mean shape is obtained by the

equation which defined as following:

∑

(2)

where the N is the number of the shapes in the

training set. The X

is a vector which represents the

coordinates of landmarks in the i-th shape:

],,...,,,,[

2211 imimiiiii

yxyxyxX =

(3)

In Eq.(3), m is the number of the points on one

shape. Afterwards, we calculate the covariance of

the training set as following:

XXXX

S )()(

−−

−

∑

(4)

And then the eigenvectors

and

corresponding eigenvalues

are computed. Now

we can approximate any shapes in the training set, X,

by using

bXX Φ+≈

(5)

where

)|...|(

21 n

and b is a N

dimensions vector which can be given by

)( XXb

−Φ=

(6)

For reducing the dimensions of the model and

variations, the Principal Components Analysis

(PCA) is employed. Thus we get the t dimensions

vector b. Eq (5) can be modified as following:

PbXXX +=

′

≈

(7)

where b is the t dimensions vector, and P is the

corresponding eigenvectors. And now we can

present a new shape that is similar to the shape in the

training set as the following:

dXXX +=

(8)

where dX

= Pb. And

)( XXPb

−=

(9)

2.2 Extract Gray-level Information and

Image Fitting

After the PDM is built, the new points are found by

modelling gray-level appearance on the target image

which present the object and transform the model

into a new better location.

ICINCO 2007 - International Conference on Informatics in Control, Automation and Robotics

390

We consider the gray-level values along a line

passing through the landmark in perpendicular to the

boundary formed by the landmark and its

neighbours. Gray-level profile g

is extracted from

pixels that are centred at the landmark for each

landmark point j in the image i of the training set.

We get the gray-level profile g

as following:

ijijijij

gggg ],...,,[

110 −

(10)

ijijijijijij

ggggggdg ],...,,[

211201 −−

−−−=

(11)

Here, the mean values are calculated as

following :

∑

ijij

(12)

Now the gray-level information has been modelled,

and for each landmark, there is a certain profile

. To

transform the mean shape into target object , the

target points responded to the points in the model

must be found, according to the modelled grey-level

information, then the model transforms to the new

model that form by target points, but this

transformation is restricted by the shape parameters

b which is defined in Eq(9). In this way, the new

shape will not bring too large distortion to represent

the object shape in almost situation. However, in the

standard ASM, there are some instances that the new

shape will occur the too large distortion and this

distortion will lead the fitting process to a failure

result (Ghassan Hamarneh).

3 SELECTION CRITERION

For fitting the images into a good shape model, we

must have a good strategy to exclude some outlying

points and select good target points. This is done

with a MSE criterion. When we calculate the shape

and pose parameters, such as the scalars

dtdds ,,

, we

need to move our current estimate X

as close as

possible to X

+dx

. Within this process, however, if

some points do not match the target object well or

their movements keep away from the target object, it

will lead to the bad direction in the image fitting,

when current estimate X

closes to X

+dx

. For

avoiding this situation, we consider to exclude those

points which are denoted as the outlying points. The

initial shape outlying these dissociated points is

denoted as

′

, and the target shape is denoted as (X

+dx

)t. In this paper, we implement such a strategy

as following:

(1) Firstly, the MSE of each point is calculated

by eq(1). And the msei is defined as:

))(())(()(

ydhCydhdfmse

−−==

−

(13)

(2) When the MSE of each point,

mse

, is

obtained, we sign the point that the

mse

value is

large than h (e.g. h=1.5) times of the mean of all the

mse

. And then we exclude these points and get the

new initial shape

′

and the new target shape

)(

′

dxX

(3) Then

′

is aligned to (X

+dx

)t and obtain

the shape and pose parameters,

dtdds ,,

(4) The shape parameters are calculated without

the influence of those dissociated points and they can

be used to transform X

into new shape.

Fig. 2 illustrates that the new shape is affected by

excluding the outlying points. It shows that the new

shape (green line) with outlying points excluded is

obviously improved since the dissociated points do

not involved in shape formation anymore.

Figure 2: The red line marked by ‘model shape’ is the

initial shape. The cyan line marked by ‘target shape’ is the

target shape. The green line marked by ‘new shape 2’ is

the new shape fitted from no outlying points. The blue line

marked by ‘new shape 1’ is the new shapefitted with the

outlying points. And the point marked by ‘point 1’ is an

example of outlying points that should be excluded.

4 EXPERIMENTS

4.1 Data Set

To evaluate our method, 400 facial images are used

to build the PDM in experiments. On facial images,

we labelled the lip with eight landmarks for each

image. Fig. 4 illustrates these landmarks.

SEARCHING AND FITTING STRATEGIES IN ACTIVE SHAPE MODELS

391

4.2 Experimental Result

In the experiments, we adopt the leave-one-out

strategy in order to evaluate the performance more

accurately and sufficiently. When each facial image

is been fitting, the remaining 399 facial images are

utilized to establish the PDM. And the same way is

performed in anklebone images. Fig. 4 illustrates the

search result of the new strategies and the traditional

ASM.

Figure 3: landmarks of the facial image.

(a) (b) (c)

Figure 4: Comparison of the searching results. Column (a)

is the standard model and its initial place. (b) Fitting

results with the standard ASM. (c) Fitting results with new

strategy.

5 CONCLUSION

In this paper, to enhance the robustness and accuracy

of image fitting, we propose a new strategy on the

Active Shape Model (ASM) method. The main

advantages are obvious from observation of practical

experiments. For example, according to the MSE

that is obtained at the process of the image fitting,

the outlying points whose corresponding MSE are

too large is excluded for forming a new shape. These

outlying points are brought by those target images

that are not clarity with some interferential object

and the new strategy can avoid effectively the

influence of outlying points. By comparison with

practical implementation, the proposed strategy

works satisfactorily.

ACKNOWLEDGEMENTS

This work is supported by the National Natural

Science Foundation of China [NSFC-60405009,

60605013], [ZJNSF-Y105101, Y104185], and a

grant for Key Research Items from the Dept of

Science and Technology of Zhejiang Province

[2006C21002]. S. Y. Chen is a research fellow of the

Alexander von Humboldt Foundation, Germany.

REFERENCES

T. F. Cootes, C.J.Taylor, D. Cooper, and J. Graham.:

Active shape models--their training and application.

Computer vision and image understanding, 61(1):

pp38-59, 1995.

Marleen de Bruijne, Bram van Ginneken, Wiro J. Niessen,

and Max A. Viergever: Active shape model

segmentation using a non-linear appearance model:

application to 3D AAA segmentation. IEEE

Transactions on Medical Imaging, 2001

Kwok-Wai Wan, Kin-Man Lam, Kit-Chong Ng: An

accurate active shape model for facial feature

extraction. Pattern Recognition Letters 26 (2005) 2409

–2423

Wei Wang, Shiguang Shan, Wen Gao, Bo Cao, Baocai

Yin: An Improved Active Shape Model for Face

Alignment. Multimodal Interfaces, 2002. Proceedings.

Fourth IEEE International Conference on

Huiqi Li, Opas Chutatape: Boundary detection of optic

disk by a modified ASM method. Pattern Recognition

36 (2003) 2093 – 2104

Lu Huchuan, Shi Wengang: Accurate Active shape model

for face alignment. Proceedings of the 17th IEEE

International Conference on Tools with Artificial

Intelligence (ICTAI’05)

Chai Xiujuan, Shan Shiguang, Gao Wen, Chen Xilin:

Example-Based Learning for Automatic Face

Alignment. Journal of Software(2005) Vol.16,No.5

A. Hill, T.F. Cootes, C.J. Taylor: Active Shape Models

and the shape approximation problem. Image and

Vision Computing 14 (1996) 601-607

Ghassan Hamarneh.Active shape models, modeling shape

variations and gray level information and an

application to image search and classification

[EB/OL].http://www.ae.chalmers.se/～jessi/

ICINCO 2007 - International Conference on Informatics in Control, Automation and Robotics

392