A SIMPLE SCHEME FOR CONTOUR DETECTION

Gopal Datt Joshi

Center for Visual Information Technology, IIIT Hyderabad

Gachibowli, Hyderabad, India-500032

Jayanthi Sivaswamy

Center for Visual Information Technology, IIIT Hyderabad

Gachibowli, Hyderabad, India-500032

Keywords:

Contour detection, Surround suppression, Primary visual cortex, Human visual system.

Abstract:

We present a computationally simple and general purpose scheme for the detection of all salient object con-

tours in real images. The scheme is inspired by the mechanism of surround inﬂuence that is exhibited in

80% of neurons in the primary visual cortex of primates. It is based on the observation that the local context

of a contour signiﬁcantly affects the global saliency of the contour. The proposed scheme consists of two

steps: ﬁrst ﬁnd the edge response at all points in an image using gradient computation and in the second step

modulate the edge response at a point by the response in its surround. In this paper, we present the results of

implementing this scheme using a Sobel edge operator followed by a mask operation for the surround inﬂu-

ence. The proposed scheme has been tested successfully on a large set of images. The performance of the

proposed detector compares favourably both computationally and qualitatively, in comparison with another

contour detector which is also based on surround inﬂuence. Hence, the proposed scheme can serve as a low

cost preprocessing step for high level tasks such shape based recognition and image retrieval.

1 INTRODUCTION

Contour detection in real images is a fundamental

problem in many computer vision tasks. Contours are

distinguished from edges as follows. Edges are vari-

ations in intensity level in a gray level image whereas

contours are salient coarse edges that belong to ob-

jects and region boundaries in the image. By salient

is meant that the contour map drawn by human ob-

servers include these edges as they are considered to

be salient. However, the contours produced by differ-

ent humans for a given image are not identical when

the images are of complex, natural scenes. In such

images, multiple cues are available for the human vi-

sual system (HVS) - low level cues such as coherence

of brightness, texture or continuity of edges, interme-

diate level cues such as symmetry and convexity, as

well as high level cues based on recognition of fa-

miliar objects. Even if two observers have exactly

the same set cues, they may choose contours at vary-

ing levels of granularity. Thus saliency of an edge

is a subjective matter and varies accordingly. Nev-

ertheless, the fact remains that a contour map drawn

by human observers is sparser than an edge map de-

rived by processing the digital image. This can be

seen from Fig 1(a) which shows a test image and the

corresponding ground truth data (Fig. 1(b))indicating

the contours considered relevant by a human observer

(dat, 2003). If we compare this with the edge maps

in Fig. 1(c), (d) extracted by a Canny detector we can

observe that the contour map is sparse. This is de-

spite selecting a low scale (to capture gross informa-

tion) and using two different thresholds for the edge

detection. In general, a contour map is an efﬁcient

representation of an image since it retains only salient

information and hence is more valuable for high level

computer vision tasks. The design of a detector that

can extract all contours from a wide range of images

is therefore of interest.

The key to extracting contours appears, from the

ground truth, to be the ability to assess what is

relevant and what is not in a local neighbourhood.

For instance, the grassy texture has been rejected in

the ground truth while the edges deﬁning the ele-

phants’ feet have been retained. An assessment-based

strategy has been attempted to contour detection us-

ing local information around an edge such as im-

age statistics, topology, texture, colors, edge con-

tinuity, density, etc. Speciﬁcally, these approaches

have used statistical analysis of gradient ﬁeld (Meer

236

Datt Joshi G. and Sivaswamy J. (2006).

A SIMPLE SCHEME FOR CONTOUR DETECTION.

In Proceedings of the First International Conference on Computer Vision Theory and Applications, pages 236-242

DOI: 10.5220/0001374702360242

 SciTePress

(a) (b)

Figure 1: Demonstration of texture as a problem in contour

detection process. (a) Image of elephants (b) Ground truth

image. Canny edge map using low scale and low threshold,

and Georgescu, 2001), anisotropic diffusion (Perona

and Malik, 1990; Black et al., 1998), complementary

analysis of boundaries and regions (Ma and Manju-

nath, 2000) and edge density information (Dubuc and

Zucker, 2001). These approaches, by design, are very

extensive in computation.

The HVS is capable of extracting all important

contour information in its early stages of processing.

Some attempts have also been made to model con-

tour detection in the HVS. One such model assumes

that saliency of contours arise from long-range in-

teraction between orientation-selective cortical cells

(Yen and Finkel, 1998). This model accounts for a

number of experimental ﬁndings from psychophysics

but is computationally intensive and its performance

is unsatisfactory on real images. Another model also

emphasises the role of local information and focuses

on cortical cells which are tuned to bar type features

(Grigorescu et al., 2003). This scheme computes ori-

ented Gabor energy at a single scale and over twelve

different orientations followed by a non-classical re-

ceptive ﬁeld (non-CRF)

inhibition. Results of testing

The classical receptive ﬁeld (CRF) is, by deﬁnition,

the area within which one can activate an individual neu-

ron. The region beyond this area which can modulate the

response of the concerned neuron is called a non-classical

receptive ﬁeld.

of this scheme on images of animals in their natural

habitat, are reasonably good. However, the scheme

is computationally expensive and produces a contour

map which is quite sparser than an edge map though

not as sparse as the ground truth (contour map). In

this paper, we seek to ﬁnd a solution to contour detec-

tion which is computationally low in cost as well as

effective on a wide range of images including natural

images. The paper is organized as follows: section 2

presents the development and details of the scheme;

section 3 proposes an implementation of the scheme;

section 4 summarises the results and section 5 draws

some conclusions based on the performance of the

proposed scheme.

2 PROPOSED SCHEME

The human visual system, in its early stages of

processing, differentiates between isolated bound-

aries such as object contours and region boundaries,

on the one hand, and edge in group, such as those

in texture, on the other hand. This is accomplished

in a series of processing stages. At the retinal level,

the ganglion cells process the visual input from rods

and cones to produce an image similar to that of

an edge detector used in computer vision (Marr and

Hildreth, 1980). The ganglion cells signal the spa-

tial differences in the light intensity falling upon the

retina. Their receptive ﬁeld is organized into a center-

surround fashion, in which the excitatory and in-

hibitory subﬁelds are integrated into circularly sym-

metric regions. The classical work in (Marr and

Hildreth, 1980) modelled this receptive ﬁeld with a

Laplacian of Gaussian function. At the output of this

stage (retina), the visual system provides an efﬁcient

representation of the image by removing redundant

information such as uniform light intensity on adja-

cent retinal locations. The axons of the ganglion cells

project to an area in the brain called the Lateral Genic-

ulate Nucleus (LGN). This area has no known ﬁlter

function but serves mainly to project binocular visual

input to various sites, especially to the visual cortex.

In the visual cortex, Hubel and Wiesel (Hubel and

Wiesel, 1962) found simple and complex cells in cat

primary visual cortex (area V1) that are selective

to intensity changes in speciﬁc orientation. These

orientation-selective cells are organized in columns,

in which all cells in a column have the same pre-

ferred orientation, and adjacent columns have incre-

mental change in orientations (Hubel and Wiesel,

1962). The orientation selectivity of cells is accom-

plished through a spatial summation of the inputs

from LGN. This leads to interesting features. For

example, complex cells differ from simple cells by

showing less speciﬁcity regarding the position of the

A SIMPLE SCHEME FOR CONTOUR DETECTION

237

stimulus. Computationally, this functionality can be

modeled using input from the simple cells.

Figure 2: Contour detection scheme in (Grigorescu et al.,

2003).

Other classes of cells besides the simple and com-

plex cells also exist in V1. These are the end stopped

cells (Dobbins et al., 1987), bar (contour) cells (Bau-

mann et al., 1997), and grating cells (von der Heydt

et al., 1991). Neurophysiological measurements on

cells have showed that the response of an orientation-

selective neuron to an optimal stimulus in its recep-

tive ﬁeld is reduced if the stimulus extends to the

surround. This effect is termed as surround inhibi-

tion and it is exhibited in a majority (80%) of the

orientation-selective cells in the visual cortex of pri-

mates(Knierim and van Essen, 1992). In general,

an orientation-selective cell with surround- inhibition

will respond most strongly to a single bar, line, or

edge in its receptive ﬁeld and will show reduced re-

sponse when more bars are added to the surroundings

like sine wave gratings. These type of cells was called

the bar cell, referring to the preference of the cell for

bars versus gratings. These cells were the source of

inspiration for the contour detector reported in (Grig-

orescu et al., 2003). The detection scheme is based

on ﬁtering the input image with a Gabor ﬁlter bank;

summing the Gabor energy output at 12 different ori-

entations and then applying surround inhibition. This

scheme is shown in Fig. 2. The Gabor ﬁltering stage

is essentially a local energy computation stage (Mor-

rone and Burr, 1988). Edge and line features are

signalled by local maximas in the local energy map.

Hence, the input to the surround inhibition stage in

Fig. 2 is edge information.

Edge information can also be derived using a gra-

dient computation. The difference between edge de-

tection using gradients as opposed to local energy is

that the latter is capabile of detecting step/impulse

discontinuities as well as ramps and other luminance

proﬁles in the image. Most of the edges in natural

images have step proﬁles which can be effectively

picked by the gradient computation. A drawback of

Gabor based commutation is that it leads to a poor

localization since the window over which it is com-

puted needs to be wide enough to attain a good ﬁt to

the Gabor proﬁle. Furthermore, local energy compu-

tation is far too expensive (12×2=24 ﬁlters) compared

to gradient computation which needs to be done only

in two orthogonal directions (2 ﬁlters) to determine

edges at various orientation. Hence, we argue that

a simpler scheme for contour detection would be to

derive the edge information from a gradient compu-

tation followed by an assessment based on the local

context.

Figure 3: Neighboring cells proﬁle of a cortical cell in area

V1.

Next, we turn our attention to the second part of

the contour detection scheme, namely assessment of

the edge information. This assessment needs to be

done based on the local context and a surround in-

hibition mechanism has been used for this purpose

in (Grigorescu et al., 2003). A recent neurophysi-

ological study (Cavanaugh et al., 2002) of cells in

V1 has however shown that the surround region actu-

ally can have an excitatory inﬂuence in addition to an

inhibitory inﬂuence. Speciﬁcally, the ﬁndings about

neuronal behaviour can be summarised as follows: (i)

Every cell responds to a stimulus if it falls on its cen-

tral region (CRF). This is shown by the circle P in

Figure 3; (ii) Besides the CRF, a neuron has a sur-

round region which is made up of two parts, namely

an inner annular shaped region (shown in dark gray

in Figure 3) which is excitatory and an outer annu-

lar shaped region (shown in light grey in Figure 3)

which is inhibitory; (iii) The surround region which

can inﬂuence the response of a neuron is limited in

extent. The behavior of these neurons give us a clue

about the role of local context in the visual perception

of a stimulus which is obtained by combining exci-

tatory and inhibitory inﬂuences. We draw inspiration

from this study and propose a surround mechanism

with excitatory and inhibitory components. An con-

tour detection scheme which computes the gradient

information ﬁrst and then employs this second step

will not respond (sufﬁciently) to edges which belong

to texture regions. Such a scheme is easy to imple-

ment as well.

VISAPP 2006 - IMAGE ANALYSIS

238

3 IMPLEMENTATION OF

SCHEME

The proposed scheme for contour detection can be

implemented in a simple manner as follows. First,

compute a gradient map. Next, incorporate the sur-

round inﬂuence on the gradient map using a mask

operation. Finally, binarise the output of the second

stage using a standard procedure.

3.1 Gradient Estimation in the

Discrete Domain

As mentioned earlier , it sufﬁce to perform the gra-

dient computation in two orthogonal directions using

Sobel or Prewitt masks.

3.2 Surround Interaction

The surround inﬂuence can be implemented by ei-

ther taking into account the direction of the gradient

or ignoring the same. The former will lead to less

amount of surround supression in natural images be-

cause they generally contain texture edges at arbitrary

orientations. In other words, derive contour map will

be noisy. An surround inﬂuence operation which ig-

nore the edge orientation can on the other hand, im-

prove the level of supression. Hence, the edge ass-

esment based on the surround inﬂuence can be imple-

mented as a convolution operation with an appropriate

isotropic mask. We will now explain how this mask

can be designed. As explained in the previous section,

this surround consists of two annular regions, the in-

ner being excitatory and the outer being inhibitory in

nature. Hence, in the mask, we will deﬁne two an-

nular regions surrounding the central pixel. The inner

one is assigned positive weights (excitatory) while the

outer one is assigned negative weights (inhibitory).

The magnitude of the weights is chosen based on the

following neurophysiological ﬁndings: the inﬂuence

of a point in the surround on the cell response is de-

pendent on its distance from the centre and this is

roughly Gaussian in proﬁle. Furthermore, the excita-

tory inﬂuence is weaker than the inhibitory inﬂuence

(Series et al., 2003). The size of the required mask

was determined by varying it from 7 × 7 to 15 × 15.

It was found that a mask of size 7 × 7 is optimal to

achieve the best results. Hence, we propose a 7 × 7

mask as shown in Figure. 4. The weights in the outer

inhibitory region of the optimal mask was found by

sampling a Gaussian proﬁle of standard deviation 1.4.

The sum of the weights inside this mask is set at 0.52

to help enhance a contour pixel.

Figure 4: A 7 × 7 surround mask.

3.3 Binarisation

A binary contour map can be constructed by using

a standard procedure such as nonmaxima supression

followed by hysteresis thresholding (Canny, 1986).

Let the gradient magnitude M (x, y) and orientation

map Θ(x, y) specify the local strength and local edge

direction respectively. Nonmaxima suppression seeks

to thin regions where M(x, y) is non-zero, to gener-

ate candidate contours as follows: two virtual neigh-

bors are deﬁned at the intersections of the gradient

direction with a 3 × 3 sampling grid and the gradi-

ent magnitude for these neighbors is interpolated from

adjacent pixels. The central pixel is retained for fur-

ther processing only if its gradient magnitude is the

largest of the three values. The ﬁnal contour map is

computed from the candidates by hysteresis thresh-

olding. This process involves two threshold values t

and t

, t

. All the pixels with M (x, y) ≥ t

are

retained for the ﬁnal contour map, while all the pix-

els with M(x, y) ≤ t

are discarded. The pixels with

<M(x, y) <t

are retained only if they already

have at least one neighbor in the ﬁnal contour map.

4 EXPERIMENTAL RESULTS

4.1 Ground Truth Image Data

Most of the methods for the evaluation of edge and

contour detectors use natural images with associated

desired output that is subjectively speciﬁed by the

observer [(Martin et al., 2004), (Grigorescu et al.,

2003)]. Some recent studies (Shin et al., 2001) show

that the performance of such an operator must be

A SIMPLE SCHEME FOR CONTOUR DETECTION

239

considered task dependent. For object recognition,

for example, some operators may perform better than

others despite similar performance on synthetic im-

ages. The proposed surround interaction mechanisms

is aimed at a better detection of objects contours in

natural scenes.

We tested the performance of the proposed scheme

on 40 natural images from a database designed to

evaluate the performance of contour detection (dat,

2003). For each test image, an associated desired out-

put binary contour map that was drawn by human is

given. It should be noted that the ground truth data in-

cludes more than one type of pixels: (i) pixels which

are parts of a contour of an object (ii) pixels which

are part of a boundary between two (textured) regions.

Our proposed scheme on the other hand, is designed

to extract only the ﬁrst type of contour pixels.

4.2 Performance Measure

In order to have a quantitative comparison between

the contour detector proposed in (Grigorescu et al.,

2003) we use the performance measure introduced in

the same. Let f

and f

are number of false positive

and false negative pixels detected in the ﬁnal contour

map, respectively. The performance measure is de-

ﬁned as follows:

P =

+ f

(1)

where, t

is the number of correctly detected con-

tour pixels (True positive). The performance measure

P is a scalar taking values in the interval [0, 1].If

the desired output pixels are correctly detected and no

background pixels are falsely detected as contour pix-

els, then P =1. For all other cases, P takes values

smaller than one, being closer to zero as more con-

tour pixels are falsely detected and/or missed by the

operator.

For computing the performance measure, we must

determine which true positives are correctly detected,

and which detection is false. The binary contour map

in the ground truth data can be used for this pur-

pose. Let us consider how to compute P of a out-

put contour image given a ground truth contour map.

One could simply correspond coincident contour pix-

els and declare all unmatched pixels either as false

positives or misses. However, this approach would

not tolerate any localization error and result in a poor

performance measure. For robustness, it is desirable

that the correspondence of output contour pixels to

ground truth tolerate localization errors since ground

truth data is not accurately localized. The approach

proposed in (Grigorescu et al., 2003) considers a con-

tour pixel as correctly detected if a corresponding

ground truth contour pixel is present in a 5×5 (empir-

ically ﬁnd) square neighborhood (window) centered

Table 1: Performance of proposed scheme and reported

by Cosmin (Grigorescu et al., 2003) on 3 natural im-

ages(elephant, goat and hyena).

Image Method Performance

Goat (Grigorescu et al., 2003) 0.34

Proposed Scheme 0.51

Elephant (Grigorescu et al., 2003) 0.42

Proposed Scheme 0.61

Hyena (Grigorescu et al., 2003) 0.55

Proposed Scheme 0.76

at the respective pixel coordinate. A window based

approach leads to a less robust performance measure,

as different sizes of the window can be shown to af-

fect the performance value signiﬁcantly, which is a

undesirable. A large window will boost the number

of true positive. However, an explicit correspondence

of the detected contour and ground truth contour pix-

els is the only way to robustly count the hits, misses

and false positives that we need to compute P .We

have used the algorithm presented in (Martin et al.,

2004) for the correspondence between output contour

map and a ground truth contour map. The algorithm

converts the corresponding problem into a minimum

cost bipartite assignment problem, where the weight

between a output contour pixel and ground truth con-

tour pixel is proportional to their relative distance in

the image plane. One can then declare all contour

pixels matched beyond some threshold to be non-hits.

The correspondence computation is detailed in (Mar-

tin et al., 2004).

4.3 Results

The proposed contour detection scheme was tested on

40 images from a database reported in (Grigorescu

et al., 2003). Of these, we present results on 3 im-

ages in Fig 5 for illustrative purposes. A qualita-

tive comparison between the results of our contour

detection scheme and the contour reported in (Grig-

orescu et al., 2003) can be made by observing the re-

sults in this ﬁgure. The Canny edge detector outputs

are also included for reference. The ﬁrst and second

columns show the input images and the corresponding

ground truth images, respectively; the third column

shows the best results of the Canny edge detector; the

fourth column shows the results of the contour detec-

tion reported by (Grigorescu et al., 2003); and ﬁnally,

the ﬁfth column shows the results of the proposed

VISAPP 2006 - IMAGE ANALYSIS

240

Figure 5: Results of contour detection on test images. (a) & (b) show input images and associated ground truth images

respectively (c) Canny edge map (d) Best contour map reported in (Grigorescu et al., 2003) (e) Best results obtained by the

proposed contour detection scheme.

scheme. The ﬁrst point to note is that the obtained

contours in the fourth and ﬁfth columns are closer to

the ground truth than the Canny output which is jus-

tiﬁably very ’noisy’. Furthermore, it can be seen that

the results of the proposed scheme are closer to the

ground truth. For instance, the grassy region in the

elephant image (top row) is suppressed well. The best

result of the contour map can be seen from the hyena

images (in the bottom row) where the output is very

close to the ground truth image.

A quantitative comparison of the contour and edge

detectors is shown in Table 1. The ﬁgures in the ta-

ble show that the proposed scheme outperforms the

scheme in (Grigorescu et al., 2003) in all 3 images

shown in Fig 5. This is consistent with our visual

analysis of the results.

Fig. 6 shows statistical Box- and - Whisker plots

for ten of the images used in our experiments. These

plots are helpful in interpreting the distribution of per-

formance value over ten different images. The aver-

age obtained value is above 0.5 which is encouraging

since the measure is computed using ground truth im-

ages which included contour pixels belonging to ob-

ject as well as texture region boundaries, whereas our

scheme is designed for extracting only the former.

5 CONCLUSION

Though centre-surround receptive ﬁelds have been

explored as a possible solution for good edge detec-

tion, a centre-surround mechanism is also applica-

ble at a wider level (summation across neighbour-

ing cells) to achieve contour detection. The proposed

scheme for contour detection is motivated from such

center-surround inﬂuence in the cortical cells of pri-

mates. It contributes to better contour detection not by

enhancing responses to contours, but by selectively

suppressing edges based on the surround. Speciﬁ-

cally, an edge (signalled by a strong gradient) qual-

iﬁes to be a contour only if it is salient in a local con-

text where saliency implies that either the surround

has no edges or the surround has weaker edges. Thus,

the proposed scheme is not dependent on any prior

knowledge which makes it a general preprocessing

step for high order tasks. A further attractive feature

of the scheme is that it is also computationally low in

cost compared with many of the earlier approaches to

general purpose contour detection.

In practice, contour detection is an intermediate

level operation in computer vision with its output of-

ten used as input for further stages performing higher

level processing. It is hence of interest to know the

A SIMPLE SCHEME FOR CONTOUR DETECTION

241

Figure 6: The distribution of performance value over ten

different images.

appropriateness of its use given a speciﬁc high level

task. As can be seen from the results, the proposed

contour scheme largely suppresses the local back-

ground information and hence it is not appropriate

to deploy it in tasks where the background informa-

tion is essential, e.g. texture classiﬁcation or region

based segmentation. In other high-level tasks such as

shape-based recognition and image retrieval, the pro-

posed scheme can play a very useful role in their per-

formance improvement.

REFERENCES

(2003). http://www.cs.rug.nl/∼imaging/databases/contour

database/contour database.html.

Baumann, R., van der Zwan, R., and Peterhans, E. (1997).

Figure-ground segregation at contours: a neural mech-

anism in the visual cortex of the alert monkey. In Eu-

ropean Journal of Neuroscience.

Black, M., Sapiro, G., Marimont, D., and Heeger, D.

(1998). Robust anisotropic diffusion. In IEEE Trans-

action on Image Processing.

Canny, J. (1986). A computational approach to edge detec-

tion. In IEEE Transactions on Pattern Analysis and

Machine Intelligence.

Cavanaugh, J., Bair, W., and Movshon, J. (2002). Nature

and interaction of signals from the receptive ﬁeld cen-

ter and surround in macaque v1 neurons. In Journal

of Neurophysiology.

Dobbins, A., Zucker, S. W., and Cynader, M. S. (1987).

Endstopped neurons in the visual cortex as a substrate

for calculating curvature. In Nature.

Dubuc, B. and Zucker, S. (2001). Complexity, confusion

and perceptual grouping. part ii: mapping complexity.

In International Journal on Computer Vision.

Grigorescu, C., Petkov, N., and Westenberg, M. (2003).

Contour detection based on nonclassical receptive

ﬁeld inhibition. In IEEE Transactions on Image

Processing.

Hubel, D. H. and Wiesel, T. N. (1962). Receptive ﬁelds,

binocular interaction and functional architecture in the

cats visual cortex. In Journal of Psychology.

Knierim, J. and van Essen, D. (1992). Neuronal re-

sponses to static texture patterns in area v1 of the alert

macaque monkey. In Journal of Neurophysiology.

Ma, W.-Y. and Manjunath, B. (2000). Edgeﬂow: A tech-

nique for boundary detection and image segmentation.

In IEEE Transactions on Image Processing.

Marr, D. and Hildreth, E. (1980). Theory of edge detection.

In Proceedings of the Royal Society.

Martin, D. R., Fowlkes, C. C., and Malik, J. (2004). Learn-

ing to detect natural image boundaries using local

brightness, color, and texture cues. In IEEE Trans-

actions on Pattern Analysis and Machine Intelligence.

Meer, P. and Georgescu, B. (2001). Edge detection with em-

bedded conﬁdence. In IEEE Transactions on Pattern

Analysis and Machine Intelligence.

Morrone, M. C. and Burr, D. C. (1988). Feature detection in

human vision: A phase-dependent energy model. In

Proceedings of the Royal Society, London Series B.

Perona, P. and Malik, J. (1990). Scale-space and edge detec-

tion using anisotropic diffusion. In IEEE Transactions

on Pattern Analysis and Machine Intelligence.

Series, P., Lorenceau, J., and Fregnac, Y. (2003). The silent

surround of v1 receptive ﬁelds: theory and experi-

ments. In Journal of Physiology Paris.

Shin, M. C., Glodgof, D. B., and Bowyer, K. (2001). Com-

parision of edge detectors using an object recognition

task. In Computer Vision and Image Understanding.

von der Heydt, R., Peterhans, E., and Drsteler, M. R. (1991).

Grating cells in monkey visual cortex: coding texture?

In Channels in the Visual Nervous System: Neuro-

physiology, Psychophysics and Models (Blum B, ed).

Yen, S. and Finkel, L. (1998). Extraction of perceptually

salient contours by striate cortical networks. In Vision

Research.

VISAPP 2006 - IMAGE ANALYSIS

242