744 ieee journal of oceanic engineering, vol. 32...

744 IEEE JOURNAL OF OCEANIC ENGINEERING, VOL. 32, NO. 3, JULY 2007

Sidescan Sonar Segmentation Using TextureDescriptors and Active Contours

Maria Lianantonakis and Yvan R. Petillot

Abstract—This paper is concerned with the application of ac-tive contour methods to unsupervised binary segmentation of high-resolution sonar images. First, texture features are extracted froma sidescan image containing two distinct regions. A region-basedactive contour model of Chan et al. [J. Vis. Commun. Image Repre-sent., vol. 11, pp. 130–141, 2000] is then applied to the vector-valuedimage extracted from the original data. Our implementation in-cludes a new automatic feature selection step used to readjust theweights attached to each feature in the curve evolution equationthat drives the segmentation. Results are shown on simulated andreal data. The influence of the algorithm parameters and contourinitialization are also analyzed.

Index Terms—Image processing, seabed classification, segmen-tation, sidescan sonar.

I. INTRODUCTION

THE role of underwater seabed analysis is becoming in-creasingly important for many applications including ma-

rine science (habitat mapping, environmental monitoring), off-shore industry, and military applications. There are two typesof existing approaches to the problem. The first is based on theanalysis of acoustic and geoacoustic data of the sea bottom (see[2], [3], and references therein). The second approach is imagebased and it makes use of, primarily, sidescan and bathymetricsonar data [4]–[7].

Sidescan sonar systems are used to generate 2-D, high-resolution images that represent large areas of the seabed. Theyare characterized by an acoustic signal which is emitted fromthe sides of the sonar system in a direction perpendicular to thedirection of travel. The beam produced from each pulse is narrowin the horizontal direction and wide in the vertical direction, itspurpose being to insonify a narrow, long strip of the seabed oneither side of the platform and perpendicular to the navigationtrack. The intensity of the backscattered signal along a strip isdisplayed as a function of time corresponding to a row of data inthe sidescan image, each time instant referring to a point in theseabed. The sidescan image is then produced by putting togethera sequence of these signals along the navigation track to createa 2-D representation of large areas of seabed.

High-resolution sidescan images are characterized by visu-ally distinct areas corresponding to objects on the seabed, vi-sualized as high-intensity areas caused by the reflection of the

Manuscript received November 8, 2004; accepted May 18, 2006.Associate Editor: B. R. Calder.The authors are with the School of Engineering and Physical Sciences, Heriot

Watt University, Edinburgh EH14 4AS, Scotland, U.K. (e-mail: [email protected]; [email protected]).

Digital Object Identifier 10.1109/JOE.2007.893683

acoustic wave; shadows, visualized as low-intensity, texturedareas caused by the lack of acoustic reverberation from areasneighboring the objects; and background, visualized as distinctareas with strong texture characteristics. The latter are the resultof the backscatter caused by the seabed, with different types ofseabed producing a distinct backscatter response.

Sidescan sonar image analysis is used in a number of appli-cations, from object localization and identification to seabedclassification and 3-D reconstruction. The analysis falls into twocategories: classification using texture features and supervisedlearning [4], [5] and unsupervised segmentation/classificationbased on Bayesian clustering methods using gray levels or tex-ture features [6], [7]. The first of these approaches has produceduseful results for seabed classification but existing methods arevery sensitive both to training and to the viewpoint in the sonardata [8]. Existing unsupervised segmentation algorithms forsidescan data are largely concerned with identifying objects onthe seabed by segmenting the image in shadow, nonshadow areas([6] and references therein), or into three regions correspondingto shadow, echo, and sea-bottom reverberation [7]. The approachused in [6] and [7] is based on a novel hierarchical Markov treemodel with prior knowledge being incorporated in the model.

In this paper, we concentrate on a different type of segmenta-tion task which has received very little treatment in the literature,namely that of segmenting a sidescan sonar image in differenttypes of seabed in an unsupervised manner. In particular, suchan algorithm can be used to segment the areas of sonar imagescorresponding to sea-bottom reverberation. This task is viewedhere as an unsupervised texture segmentation problem. In ad-dition, the extra assumption that the underlying image containstwo distinct regions is made.

There is a large literature on unsupervised texture segmenta-tion and a number of different approaches to the problem. In thispaper, a variational approach using region-based active contoursbased on level sets is adopted.

Active contour methods have been extensively used over thepast decade for boundary detection and image segmentation[9]–[18]. Originating from the classical parametric active con-tours (or snakes) [9], geometric active contours are representedimplicitly as level sets of functions of two variables. Like thesnakes, they are usually derived by minimizing a suitable data-dependent energy functional, the central idea being that a min-imum is attained when the contour optimally segments the un-derlying image. Unlike the snakes, however, they are able tomake use of the level set techniques introduced by Osher andSethian [19] to handle topological changes and convergenceproblems. Energy functionals may depend on the data in a va-riety of ways but can be broadly divided into boundary and

0364-9059/$25.00 © 2007 IEEE

Authorized licensed use limited to: Heriot-Watt University. Downloaded on November 12, 2009 at 13:22 from IEEE Xplore. Restrictions apply.

LIANANTONAKIS AND PETILLOT: SIDESCAN SONAR SEGMENTATION USING TEXTURE DESCRIPTORS AND ACTIVE CONTOURS 745

region-based models depending on the type of information usedto evolve the curve toward the boundary of the distinct regionsin the image. Region-based models aim to evolve a curve by dy-namically calculating some homogeneity measure over the en-tire region to be segmented. They are thus more suitable for im-ages where the different regions are not defined through stronggradients as, for example, in blurry, noisy, or textured images.

One way of using this type of method for texture segmen-tation is to evolve a curve simultaneously over a set of fea-ture images extracted from the original image. Several worksbased on this approach exist in the literature, mostly using Gaborfilters and wavelets as the texture feature extraction technique[20]–[25]. Of these works, only [24] and [25] present unsuper-vised segmentation methods and are closely related to the tech-niques used here. In particular, the active contour model andfeature selection method are very similar to those in [25]. Theprimary aim of the work in this paper, however, is to apply thesetechniques referred to sonar data, and in particular, to sidescansonar images, while taking advantage of the narrower domainof application to improve and refine existing methods.

The active contour model used in this paper was first intro-duced by Chan and Vese in [14]. It is obtained as a special caseof the Mumford–Shah functional [26] and uses the mean in-tensity as a region homogeneity measure. One of the attractivefeatures of this model is that it is free of any a priori assump-tions or statistical modeling of the image. The latter is desirableas textures present in sidescan images cannot be modeled byone type of distribution. Moreover, this model is robust to noiseand holds good regularization properties, similar to those of aMarkov random field, as a result of the speed dependence onthe global region statistics and the curvature of the contour. In[1], the model has been extended to include vector-valued im-ages and it is this form of the model that will be used here fortexture segmentation of sidescan images. More specifically, theChan–Vese vector-valued model is combined with the Haralicktexture features calculated from the gray level cooccurrence ma-trices. The Chan–Vese model is used in a novel way makinguse of its ability to perform automatic feature selection. Thisis implemented by automatically readjusting the weights of thefeature images to ensure that the contour evolution is driven bythose features that are the most discriminant.

The outline of this paper is as follows. In Section II, we givesome necessary background on level sets and briefly describe ageneral framework for region-based geometric active contours.The Chan–Vese active contour models (scalar- and vector-valued) are then described. In Section III, the feature set usedto perform the segmentation is briefly described. Finally, inSection IV, the implementation of the segmentation algorithmis described and several numerical results on simulated and realsonar images are presented.

II. CHAN–VESE ACTIVE CONTOUR MODEL FOR BINARY

IMAGE SEGMENTATION

A. Background

Let be an image defined on a domain . It willbe assumed throughout this paper that consists of two ho-mogeneous regions and with denoting their common

boundary. Let be a closed, simple curve withcurve parameter . The idea behind active contour segmenta-tion methods is to deform the initial curve intime using a suitable, data-dependent partial differential equa-tion (PDE) of the form

(1)

so that the family of curvessatisfying (1) converges to the boundary as . Here,and are the curve and time parameters, respectively, standsfor the inward normal vector of , and is the speed withwhich evolves in the direction of the vector . The speedmay depend on many factors (e.g., local or global properties ofthe front and image data) but it is assumed that it is independentof the curve parametrization. It has to be chosen in such a waythat the evolving curve is attracted by the boundary ofthe two regions and becomes stationary at .

The strength of this approach lies in its ability to make use ofthe level set methods introduced by Osher and Sethian in [19]and further developed by several authors over the past decade[27]–[29]. The level set formulation of (1) is given by

(2)

with initial condition where the initialsurface is chosen so that its zero-level set is given bythe initial curve in (1), that is

(3)

The family of curves , , satisfying (1) will then begiven by the zero-level sets of the surfaces , , thatsatisfy (2). In this way, any topological changes in the evolvingcurve , as splitting or merging, can be handled naturallyand powerful numerical schemes able to approximate the correctviscosity (weak) solution can be employed [27], [29].

The speed function is usually derived through minimizinga suitable image-dependent energy although it is also possibleto synthesize directly from the image data [10], [30]. Thereare the following two types of approaches when choosing aspeed : boundary- and region-based. The former rely on theboundary being described as the points in the image where

is maximized and, therefore, tend to depend only on localinformation [10], [12], [30]. On the other hand, region-basedmethods aim to segment the two regions by considering variousmeasures of homogeneity of each region. In this way, globalimage information can be incorporated in the speed function

[11], [14], [16]–[18]. A review of region-based methods canbe found in [18].

A general framework that can unify many of the region-basedapproaches in the two region cases was presented in [18]. Thisframework considers energies of the general form

(4)



where and denote the outside and inside of the curve .The kernels and play the role of “region descriptors”and are modeled as a combination of features globally attached tothe evolving regions and . The main result in [18] is thecomputation of the speed that will make a contour obeying (1)evolve toward a minimum of the energy in (4).

B. Chan–Vese Model for Scalar Images

The approach used in this paper was first introduced in [14]and generalized for vector-valued images in [1]. It can be seenas a restricted form of the Mumford–Shah functional for seg-mentation first proposed in [26]. Alternatively, it can be viewedas a special case of the framework in [18].

The general form of the Mumford–Shah functional is givenby

Length

(5)

where is continuous and piecewise smooth andand are positive parameters. The segmentation problem thatthe minimization of (5) is designed to solve can be stated asfollows: Given , find a decomposition of and an op-timal approximation of such that varies smoothly withineach and rapidly or discontinuously across the boundariesof . A minimizer of will be an “optimal”piecewise-smooth approximation of the initial, possibly noisy,image while will approximate the edges of .

In the case where consists of two distinct regions and whenrestricting the approximations of to piecewise constant func-tions (functions taking only two values and in the regioninside and the region outside , respectively), the Mum-ford–Shah functional becomes

Length

(6)

In this case, the constants and are given by

average inside

average outside

This leads to the minimization problem

Length

(7)

for and . Now, energy (7) has the form of(4) considered in [18]. Following [18], the evolution equationderived from the minimization of (7) is given by

(8)

where is the local curvature of the curve .Equation (8) can be solved by using a level set method as

explained previously. Note that in this case the speed functionis given by

(9)

and can, therefore, be extended naturally outside the curve. As demonstrated in [14], this model has several advan-

tages: ability to detect boundaries with very smooth or blurredboundaries (boundaries without gradient), automatic changeof topology and automatic detection of interior contours, scaleadaptivity (through the parameter ), and robustness to noise.

The main limitation of the model comes from the fact thatit can only discriminate regions which have different meanintensities. In general, it is unable to segment images withstrong textures. One way to overcome this is to extract features

from the initial image and apply the previousalgorithm directly to one of them. An even better segmenta-tion result can be achieved by combining the information in

. This is explained in the following.

C. Extending the Model to Vector-Valued Images

In this paper, the approach used to extend the model of theprevious section to vector-valued images is presented as follows[1].

Let be as before and let be feature mapson the domain that have been extracted from . For example,

may be the output of a filter bank applied on . The ideais to evolve a curve in as before but making use of the infor-mation contained in all of the images . One wayto achieve this is to evolve under (1) where the speed function

is given as a weighted average of terms over all images

(10)where and are the mean values of images in-side and outside . This approach was first implemented in [1]where it is also shown that the speed will evolve under (1)toward the minimum of the energy

(11)

As in the 1-D case, this approach also fits into the frameworkin [18].

The coefficients and can be used as weights at-tached to each image depending on the amount of informationthat it contains. In our implementation, the weights are ini-tially set equal to 1 and readjusted automatically as the curveevolves depending on the magnitude of the quantities

. In this way, the active contour also performs a featureselection.



III. FEATURE EXTRACTION

Sidescan sonar images are often characterized by texturedregions corresponding to different types of seabed. Manyapproaches have been proposed to model texture (filter banks,morphology, run length encoding, Markov random fields,fractals, and cooccurrence) [4]–[7], [21], [31]. The appropriatechoice of feature sets is a complex and largely unresolvedproblem. Cooccurrence matrices belong to the statistical repre-sentation of textures (fractals and Markov random fields) whichview texture as a stationary stochastic process. A stochasticmodel is well adapted to sonar imagery due to the amount ofnoise present in the images. Cooccurrence matrices have beenchosen here as they are simple, well established, and fast to cal-culate. There is also evidence that they are well adapted to sonarimagery [31]–[33]. The main emphasis of this paper is not theselection of the best texture feature set but the demonstrationthat texture features can be integrated into a curve evolutionframework to efficiently segment sidescan sonar images. Otherfeature sets could be readily integrated in the existing schemeif they were better adapted to the specific nature of the imagesto be analyzed, i.e., sidescan or other.

The feature set used here is the well-known Haralick set [34].Only a subset of the 14 Haralick features (energy, contrast,correlation, entropy, homogeneity, cluster shade, and clusterprominence) representing the most commonly chosen ones isused here. They are referenced throughout the text as to .For computational reasons, these features are not extracted foreach pixel of the original image but on a subsampled lattice de-termined by horizontal and vertical subsampling steps .For each point of the lattice, the features are extracted usingan estimation window of size . The distance betweenthe couples of pixels for which the cooccurrence is estimated isdenoted and the orientation of the couples is denoted . Moredetails on the extraction of Haralick features can be found in [34].A fast implementation of an estimate of the Haralick set usingsum and differences of histograms introduced by Unser [35] isused here. Our implementation is based on nonoptimized C coderunning on a Pentium 1.8-GHz, 512-MB random access memory(RAM) computer with Linux operating system. In Fig. 1, anexample sonar image and the extracted features to with

and is shown. The extraction timefor this image (400 360 pixels) was 45 s. It is worth noting,however, that a local histogram equalization step used to makethe textures more stationary, performing as a preprocessingstage, takes up about half of this time ( 21 s).

IV. IMPLEMENTATION AND EXPERIMENTAL RESULTS

A. Implementation of the Segmentation Algorithm

Let be a two-region grayscale image and letdenote a set of feature images extracted from

and defined on a common domain . The PDE (12) is im-plemented numerically by making use of a level set method asexplained in Section II

(12)

Fig. 1. (a) Example of sonar image and extracted Haralick features (b) energy,(c) contrast, (d) correlation, (e) entropy, (f) homogeneity, (g) cluster shade, and(h) cluster prominence. This figure clearly shows that some features are bettersuited than others for the segmentation of this example image. Therefore, auto-matic feature selection will be important.

with initial condition and where are themean values of images inside and outside . More precisely,the PDE

(13)

with initial condition is discretized.Here, denotes the signed distance function fromthe initial curve and is defined by

if lies insideif lies outside

(14)

where denotes the Euclidean distance between apoint of coordinates and a curve .

We have used a first-order monotone scheme to approximatethe term and a first-order central difference approxi-mation to the curvature term (see [27] and referencestherein for details). For computational efficiency, the values of



are updated only in a narrowband around the zero-level setof . To ensure that the evolving curve remains well within thenarrowband domain, it is necessary to reinitialize when thezero-level set of gets close to the boundary of the band. Thisis done by resetting to be equal to the signed distance from itszero-level set using (14). We have found that, in some cases,it is necessary to reinitialize more often as the gradient oftends to get very large affecting the convergence of the numer-ical solution.

The stopping criterion for the algorithm is formulated interms of the stationarity of the zero-level set of the solution :If the zero-level set of has remained unchanged over a givennumber of iterations, then the algorithm is declared to haveconverged. In our implementation, the algorithms is said tohave converged after 50 stationary iterations of the zero-levelset of . Convergence of the scheme can also be monitoredby calculating the energy (11) and determining whether it isdecreasing over time but it has not been used here.

It is important to note that convergence to the global min-imum of the energy functional (11) associated with the speed

cannot, however, be guaranteed. First, the proposed methodis a “gradient–descent” technique and it will only reach theclosest local minimum. Second, such an energy has in generalmore than one local minimum. This makes the algorithmsusceptible to initialization. The following two initializationschemes have been used here: initialization with a sequenceof uniformly distributed circles over the entire image and ini-tialization with a single central circle (see Fig. 10). Generally,uniformly distributed circles give better segmentation resultsand convergence is faster than when initializing with a singlecurve. However, for certain types of images where the objector region in the central part of the image is small relative tothe background region, the final segmentation is better wheninitializing with a single curve that intersects with the centralregion. Generally, it was found that the algorithm was notoverly sensitive to initialization.

The choice of features and the setting of the coefficientsis another issue that needs to be addressed. The

parameters and determine the degree of contribu-tion of image in the final result. In [1], the parametersare used to filter high-frequency noise from different channels.In the case where the s are the output of feature extraction(e.g., Haralick features), the level of noise is the same in allchannels and the primary role of the parameters is one of featureselection. Here, as in [25], the automatic feature selection isbased on maximizing the difference between the means of agiven feature inside and outside the curve. In [25], the numberof features to be selected is fixed a priori and the coefficients

and are set manually. In our implementation, thecoefficients and are set automatically and thenumber of selected features can be indirectly determined via athreshold (in our implementation, is set at 0.5).The parameters and are initially set to be equalto 1 for the first 300 iterations. They are subsequently resetregularly following:

(15)

where

Setting the coefficients in this manner ensures that the featureswith maximum discriminatory capacity drive the curve evolu-tion. In addition, the number of features contributing in the al-gorithm can be controlled using the criterion . As exam-ples in Section IV-B illustrate, this type of feature selection cangreatly improve the segmentation result.

The parameter controls the smoothness in the contour andthus the sharpness of the boundaries of the segmented regions.The setting of is somewhat empirical as it weights the con-tribution of the smoothness term against the contribution of thedata-driven term, the latter being difficult to bound and depen-dent on the image size. However, the results do not seem to bevery sensitive to the value of which only needs to be adjustedwithin the correct operational range for a given type of images.

B. Experimental Results

In this section, we present numerical results of the proposedmethod on synthetic and real sonar images. These were obtainedusing a Pentium 1.8-GHz processor with 512 MB of RAM. Thecode is written in nonoptimized C++ compiled with g++.

One of the problems when segmenting sonar images is thenoise or other artifacts resulting from imperfect data acquisi-tion methods. It is thus desirable to first validate the proposedmethod on synthetic data. We then present some examples ofimages which segment well and for which the segmentation isrobust under different initialization. Finally, examples of imageswhich are difficult to segment with the proposed method are alsogiven.

The choice of the window size used to extract the texture fea-tures is ultimately driven by the scale of the textures to be seg-mented. Large window sizes will provide good texture featuresbut blur the edges between texture regions while smaller win-dows will preserve the edges at the expense of feature stability.There is also a computational cost to large windows. A compro-mise must be found between those parameters. A multiresolu-tion implementation is always possible to obtain a fixed analysiswindow size irrespective of the scale of the texture but this hasa high implementation cost and the scale at which the analysishas to be performed needs to be determined. For those reasons,we have used a fixed window size representing a good compro-mise for the textures under analysis in the underwater images ofour database.

In the experiments described here, the algorithm parameterswere chosen as follows. In the feature extraction algorithm, thewindow sizes and were set to . Finally,the step sizes and were set to , the graylevels were quantized to 32, and the set of displacement vectors

for the calculation of the cooccurrence matrices was deter-mined by , . In the curveevolution, scheme parameters and were chosen as in(15) at iterations but we have also set

for those indices where, effectively removing “bad” features from the curve

evolution process.



Fig. 2. (a) Simulated image and (b) segmentation result. The simulated imagewas realized using a sonar simulator developed in [40]. It is based on ray tracingand provides realistic ground-truthed sonar images enabling the validation of themethod proposed.

Fig. 3. Curve evolution on simulated image at iterations (a) 0, (b) 20, (c) 40,and (d) 100. Image (a) corresponds to the initial curve. The others correspondto the zero-level set of � at different stages of the process. The image here iscomposed of the following two distinct textures: sand ripples and flat seabed.The final curve very precisely follows the boundary between the two regionswithout being perturbed by the vertical artifact which locally interferes with thefeature extraction. This example demonstrates the global region-based nature ofthe technique.

For sonar images, the value , whereimages were of size 256 256 pixels, was found to be a goodcompromise between the level of accuracy in boundary detec-tion and the level of noise or scale of detail within each separateregion that needed to be left undetected. The setting of followssuggestions from [14].

Figs. 2–4 show the segmentation results for three simulatedsidescan sonar images. The images in Figs. 2 and 3 contain twotypes of seabed. The image in Fig. 4 consists of three types ofseabed with the middle and top regions having a higher degreeof similarity than any of the other two pairs. All three segmen-tation results are robust to different initializations. Initializingwith an array of circles uniformly distributed over the entireimage speeds up convergence by a factor of two compared toinitializing with a curve whose interior consists of a single con-nected component (e.g., a single circle or rectangle). In Fig. 4,the two visually more similar regions (top and middle) havebeen given the same label. This is an example where the tech-nique is used on a three-class segmentation problem (it is de-signed for two-class segmentation problems only). In this case,a segmentation in three regions could be achieved by reapplyingthe method on both regions separately. In general, however, thisis not a robust approach for multiregion segmentation, at least

Fig. 4. (a) Curve evolution on simulated image containing three types of seabedat iterations (b) 0, (c) 200, and (d) 1000. Image (a) corresponds to the initialcurve. The other images correspond to the zero-level set of � at different stagesof the process. The image here is composed of three distinct textures. The algo-rithm is currently limited to two classes and amalgamates two classes together.Note that they are perceptively the closest as well.

Fig. 5. Contour evolution at (a) 0, (b) 10, (c) 20, (d) 40, (e) 60, and (f) 200 iter-ations. Image (a) corresponds to the initial curve. The other images correspondto the zero-level set of � at different stages of the process. The image here iscomposed of the following two distinct textures: sand ripples and flat seabed.The final curve very precisely follows the boundary between the two regions.

when the curve evolution is used to minimize the particular en-ergy functional used here. Initial experiments on real sonar dataconfirm this.

Figs. 5, 7, and 8 contain segmentation results on real rawsonar data. As can be seen, the data is noisy and contains variousartifacts. These results are robust to initialization and to featureselection: Running the algorithm with for all givesthe same segmentation as setting the coefficients, as explainedpreviously. The speed of convergence, however, is greatly in-creased when the feature selection is activated as only a fewof the feature images contribute to the calculation of the curveflow. Fig. 6 displays the six most discriminant features extracted



Fig. 6. Six most discriminant features selected for segmentation of the image inFig. 5. (a) Contrast at 0 . (b) Contrast at 135 . (c) Contrast at 90 . (d) Contrastat 45 . (e) Homogeneity at 0 . (f) Homogeneity at 90 . This clearly shows thatthe feature selection is effective at identifying the most discriminant features inthe feature set.

Fig. 7. Contour evolution at (a) 0, (b) 80, (c) 140, (d) 240, and (e) 460 iterations.Image (a) corresponds to the initial curve. The other images correspond to thezero-level set of � at different stages of the process. The initialization in thiscase is a single large circle. Convergence is still demonstrated showing that theproposed algorithm is not very sensitive to initialization in this case.

from the image in Fig. 5 and selected automatically. These cor-respond to the features (contrast) in (1) in four different di-rections ( and ) and (homogeneity)in the directions and . Execution time per iter-ation for an image of 400 300 was 0.2 s.

Fig. 9 shows examples of images for which some feature se-lection is necessary for correct segmentation. The images in themiddle column were obtained by setting the coefficients as

Fig. 8. Contour evolution at (a) 0, (b) 20, (c) 50, (d) 70, (e) 90, and (f) 120 iter-ations. Image (a) corresponds to the initial curve. The other images correspondto the zero-level set of � at different stages of the process. The final curve veryprecisely follows the boundary between the two regions.

Fig. 9. (a), (d), and (g) original images and corresponding segmentation results(b), (e), and (h) with and (c), (f), and (i) without feature selection. The resultsshow the importance of feature selection for convergence. Nondiscriminant fea-tures will create extra local minima in the energy functional.

described in the beginning of this section. The images in theright column were obtained by setting all of the equal to 1.

Fig. 10 contains examples of images that are harder to seg-ment using the proposed algorithm. The image in the top rowcan be successfully segmented when initializing with a smallcircle (about a third of the image radius) placed in the middleof the image. This result is shown in the middle column. Theresult on the right is the one obtained when the initial curve hasa large intersection with the surrounding region. The image inthe middle row exhibits the opposite behavior: initializing withtoo small a circle does not allow for region growing to capturethe top half of the rippled seabed (middle). Initializing with a



Fig. 10. (a), (d), and (g) original images and segmentation results when initial-izing (b), (e), and (h) with one circle and (c), (f), and (i) with an array of circles.This figure demonstrates the importance of initialization in specific cases. In thefirst image, initialization with a big central circle outperforms the initializationwith an array of small circles while the situation is inverted for the last two im-ages. There is no rule as yet to decide on a better initialization scheme. Priorsegmentation using more rudimentary techniques (k-means) could be used toinitialize the curves and improve convergence.

3 4 array of circles, however, gives a good segmentation re-sult (right). The difficulties in the last image come from the ver-tical line across the middle and the similarity of the two typesof seabed present. As in the top image, initializing with a circlehaving a small intersection with the coarser texture gives a goodresult. In all, for such “harder images,” the difficulty seems tocome from the lack of a good initialization rather than the choiceof features. Problems of this type can be addressed by usinga multiresolution approach. For example, evolving a curve onfeatures extracted at a lower resolution can be used for initial-ization. Alternatively, a different pixel-based segmentation al-gorithm can be used after the feature extraction, as an initialstep—e.g., -means or fuzzy -means—followed by regulariza-tion with the proposed algorithm.

V. CONCLUSION AND FUTURE WORK

In this paper, an unsupervised binary segmentation algorithmis proposed and applied, in particular, to sidescan sonar images.It combines the Haralick features for texture and an active con-tour model for vector-valued images proposed in [1].

The results obtained strongly suggest that the Haralick set offeatures is suitable for texture discrimination in sidescan sonardata. In particular, using the automatic feature selection step in

the algorithm, the features “contrast” and “homogeneity” turnout to be consistently the most significant.

The algorithm is also shown to have good regularization prop-erties and to be robust to noise and other artifacts resulting fromthe data acquisition process.

As a result of the nonconvexity of the energy functional, oneof the limitations of the algorithm is its sensitivity to initial-ization. One way to address this problem is by presegmentingthe image at a coarser scale using a pixel-based segmentationmethod. A multiresolution approach would also be beneficial asthe value of parameters for the cooccurrence matrices dependson the scale of the textures in the sonar image. This may beachieved by enlarging the feature set using different parameterscorresponding to different scales and using the feature selectionstep as a means of identifying the most suitable values.

The results obtained so far and presented in this paper suggestthat curve evolution is a useful tool in sonar analysis.

The proposed method can be readily extended to other con-texts. Different types of features may be used. In particular, it ispossible to “fuse” different feature sets to allow information notcaptured by the Haralick features to be considered. By takinga slightly different viewpoint, the method may also be used forsensor fusion (for example, bathymetry or video). Finally, it ispossible to extend the method to segment sonar images toclasses, , by considering coupled evolutions of severalcontours. The main difficulty in achieving this is the representa-tion of different regions through the evolving curves. A numberof different approaches exist, such as [36]–[39]. All of thesemethods can, in principle, be used in the context of this paperwith suitable adjustments. This is a direction of research that wewill pursue in the future.

ACKNOWLEDGMENT

The authors would like to thank the NATO Undersea Re-search Centre for providing the sidescan sonar data used in thispaper. They would also like to thank the editor and reviewers ofthis paper for their helpful comments which have improved theexposition considerably.

REFERENCES

[1] T. Chan, B. Sanberg, and L. Vese, “Active contours without edges forvector-valued images,” J. Vis. Commun. Image Represent., vol. 11, pp.130–141, 2000.

[2] L. Bjorno, J. Papadakis, P. Papadakis, J. Sageloli, J. Sessarego, S. Sun,and M. Taroudakis, “Identification of seabed data from acoustic reflec-tions: Theory and experiment,” Acta Acoustica, vol. 2, pp. 359–374,1994.

[3] D. Sternlicht and C. de Moustier, “Time-dependent seafloor acousticbackscatter (10–100 KHz),” J. Acoust. Soc. Amer., vol. 114, pp.2709–2721, 2003.

[4] B. Zerr, E. Maillard, and D. Gueriot, “Sea-floor classification by neuralhybrid system,” in Proc. IEEE OCEANS, 1994, pp. 239–243.

[5] D. R. Carmichael, L. M. Linnet, S. J. Clarke, and B. R. Calder, “Seabedclassification through multi-fractal analysis,” Proc. Inst. Electr. Eng.Radar Sonar Navigat., vol. 143, no. 3, pp. 140–148, 1996.

[6] M. Mignotte, C. Collet, P. Perez, and P. Bouthemy, “Sonar imagesegmentation using an unsupervised hierarchical MRF model,” IEEETrans. Image Process., vol. 9, no. 7, pp. 1216–1231, Jul. 2000.

[7] S. Reed, Y. Petillot, and J. Bell, “An automatic approach to the detec-tion and extraction of mine features in side-scan sonar,” IEEE J. Ocean.Eng., vol. 28, no. 1, pp. 90–105, Jan. 2003.

[8] J. M. Bell, M. J. Chantler, and T. Wittig, “Side-scan sonar: A direc-tional filter of seabed texture?,” Proc. Inst. Electr. Eng. Radar SonarNavigat., vol. 146, no. 1, pp. 65–72, 1999.



[9] M. Kass, A. Witkin, and D. Terzopoulos, “Snakes: Active contourmodels,” Int. J. Comput. Vis., vol. 1, pp. 321–331, 1988.

[10] R. Malladi, J. Sethian, and B. C. Vemuri, “Shape modeling with frontpropagation: A level set approach,” IEEE Trans. Pattern Anal. Mach.Intell., vol. 17, no. 2, pp. 158–175, Feb. 1995.

[11] S. Zhu, T. Lee, and A. Yuille, “Region competition: Unifying snakes,region growing and Bayes/MDL for multi-band image segmentation,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 18, no. 9, pp. 884–900,Sep. 1996.

[12] V. Caselles, R. Kimmel, and G. Sapiro, “Geodesic active contours,”Int. J. Comput. Vis., vol. 22, no. 1, pp. 61–79, 1997.

[13] A. Yezzi, A. Tsai, and A. Willsky, “A statistical approach to snakes forbimodal and trimodal imagery,” in Proc. IEEE Int. Conf. Comput. Vis.,1999, vol. 2, pp. 898–903.

[14] T. Chan and L. Vese, “Active contours without edges,” IEEE Trans.Image Process., vol. 10, no. 2, pp. 266–277, Feb. 2001.

[15] T. Chan and L. Vese, “A level set algorithm for minimizing the Mum-ford-Shah functional in image processing,” in Comput. Soc. Proc. 1stIEEE Workshop Variat. Level Set Methods Comput. Vis., 2001, pp.161–168.

[16] J. Kim, J. Fisher, A. Yezzi, M. Cetin, and A. Willsky, “Non-parametricmethods for image segmentation using information theory and curveevolution,” in Int. Conf. Image Process., 2002, pp. 797–800.

[17] N. Paragios and R. Deriche, “Geodesic active regions: A new frame-work to deal with frame partition problems in computer vision,” J. Vis.Commun. Image Represent., vol. 13, pp. 249–268, 2002.

[18] S. Jehan-Besson, M. Barlaud, and G. Aubert, “Deformable regionsdriven by an Eulerian accurate minimization method for image andvideo segmentation,” Int. J. Comput. Vis., vol. 53, no. 1, pp. 45–70,2003.

[19] S. Osher and J. Sethian, “Fronts propagating with curvature-dependentspeed: Algorithms based on Hamilton-Jacobi formulation,” J. Comput.Phys., vol. 79, pp. 12–49, 1988.

[20] C. Sagiv, N. A. Sochen, and Y. Y. Zeevi, “Integrated active contoursfor texture segmentation,” IEEE Trans. Image Process., vol. 15, no. 6,pp. 1633–1646, Jun. 2006.

[21] S. Liapis, E. Sifakis, and G. Tziritas, “Color and texture segmenta-tion using wavelet frame analysis, deterministic relaxation, and fastmarching algorithms,” J. Vis. Commun. Image Represent., vol. 15, no.1, pp. 1–26, 2004.

[22] N. Paragios and R. Deriche, “Geodesic active regions and level setmethods for supervised texture segmentation,” Int. J. Comput. Vis., vol.46, no. 3, pp. 223–247, 2002.

[23] J.-F. Aujol, G. Aubert, and L. Blanc-Féraud, “Supervised classificationfor textured images,” INRIA, France, Tech. Rep. 4640, 2002.

[24] M. Rousson, T. Brox, and R. Deriche, “Active unsupervised texturesegmentation on a diffusion based feature space,” in Int. Conf. Comput.Vis. Pattern Recognit., 2003, vol. 2, pp. 699–706.

[25] B. Sandberg, T. Chan, and L. Vese, “A level-set and Gabor-based ac-tive contour algorithm for segmenting textured images,” UCLA, LosAngeles, CA, CAM Rep. 02-39, 2002.

[26] D. Mumford and J. Shah, “Optimal approximations by piecewisesmooth functions and associated variational problems,” Commun.Pure Appl. Math., pp. 577–685, 1989.

[27] J. Sethian, Level Set Methods and Fast Marching Methods. Cam-bridge, U.K.: Cambridge Univ. Press, 1999, ch. 6, 7, 16, 17.

[28] S. Osher and N. Paragios, Geometric Level Set Methods in Imaging,Vision and Graphics. New York: Springer-Verlag, 2003, ch. 1, 7, 11,12.

[29] S. Osher and R. Fedwick, Level Set Methods and Dynamic ImplicitSurfaces. New York: Springer-Verlag, 2003, ch. 3.

[30] G. Sapiro, Geometric Partial Differential Equations and Image Anal-ysis. Cambridge, U.K.: Cambridge Univ. Press, 2001, ch. 2.

[31] I. Karoui, J.-M Boucher, R. Fablet, and J.-M Agustin, “Multi-scale seg-mentation of texture sonar images using occurrence statistics,” in Proc.Int. Conf. Image Process., 2004, pp. 1691–1694.

[32] N. Pican, E. Trucco, M. Ross, D. M. Lane, Y. Petillot, and I. T. Ruiz,“Texture analysis for seabed classification: Co-occurrence matricesversus self-organizing maps,” in Proc. IEEE OCEANS, Nice, France,Sep. 1998, pp. 424–428.

[33] J. M. Preston, A. C. Christney, S. F. Bloomer, and I. L. Beaudet,“Seabed classification of multi-beam sonar images,” in Proc. IEEEOCEANS, 2001, vol. 4, pp. 2616–2623.

[34] R. M. Haralick, K. Shanmugam, and I. Dinstein, “Textural features forimage classification,” IEEE Trans. Syst., Man, Cybern., vol. 3, no. 6,pp. 610–621, 1973.

[35] M. Unser, “Sum and difference histograms for texture classification,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 11, no. 7, pp. 717–727,Jul. 1989.

[36] L. Vese and T. Chan, “A multiple level set framework for image seg-mentation using the Mumford and Shah model,” Int. J. Comput. Vis.,vol. 50, no. 3, pp. 271–293, 2002.

[37] J.-F. Aujol, G. Aubert, and L. Blanc-Féraud, “Wavelet-based level setevolution for classification of textured images,” IEEE Trans. ImageProcess., vol. 12, no. 12, pp. 1634–1641, Dec. 2003.

[38] I. Ayed, A. Mitiche, and Z. Belhadj, “Multiregion level set partitioningof synthetic aperture radar images,” IEEE Trans. Image Process., vol.27, no. 5, pp. 793–800, May 2005.

[39] C. Vazquez, A. Mitiche, and I. B. Ayed, “Image segmentation as reg-ularized clustering: A fully global curve evolution method,” in Proc.IEEE Int. Conf. Image Process., Oct. 2001, pp. 3467–3470.

[40] J. M. Bell and L. M. Linnett, “Simulation and analysis of synthetic side-scan sonar images,” Proc. Inst. Electr. Eng. Radar, Sonar, Navigat., vol.144, no. 4, pp. 219–226, 1997.

Maria Lianantonakis received the Ph.D. degree inmathematics from King’s College, London, U.K., in1991. Her dissertation was in the area of elliptic par-tial differential equations.

Between 1991 and 1995, she held research posi-tions at King’s College and the University of Sussex,Sussex, U.K. Since 2002, she has been holdingresearch positions at Heriot-Watt University, Edin-burgh, Scotland, U.K. Her current research interestsinclude image segmentation, level set methods inimage analysis, and sonar simulation.

Yvan R. Petillot received the French Engineeringdegree in telecommunications from École NationaleSupérieure des Télécommunications de Bretagne(ENSTBr), France, with a specialization in imageand signal processing and the Ph.D. degree inreal-time pattern recognition using optical proces-sors from the Université de Bretagne Occidentale,Brest, France, in 1996.

Currently, he is a Lecturer at Heriot-Watt Univer-sity, Edinburgh, Scotland, U.K., leading several re-search projects in image processing and robotics in-

cluding texture and shape analysis, segmentation and classification of sonar andvideo images, and simultaneous mapping and localization and navigation. Hehas a particular interest in seabed mapping and mine detection and classificationusing statistical techniques.

Dr. Petillot is a Reviewer for various journals of the IEEE and the Instituteof Electrical Engineers (IEE) including the IEEE JOURNAL OF OCEANIC

ENGINEERING, the IEEE TRANSACTIONS ON IMAGE PROCESSING, the IEEEPATTERN ANALYSIS AND MACHINE INTELLIGENCE, the IEEE TRANSACTIONS

ON ROBOTICS, and IEE Sonar and Navigation.


744 ieee journal of oceanic engineering, vol. 32...

Documents