DETECTION OF POINTS OF INTEREST

FOR

GEODESIC CONTOURS

Application on Road Images for Crack Detection

Sylvie Chambon

Institut Franc¸ais des Sciences et Technologies des Transports, de l’Am´enagement et des R´eseaux (IFSTTAR)

Champs-sur-Marne, France

Keywords:

Detection, Segmentation, Thin structures, Road cracks, Point of interest, Geodesic contour, Auto-correlation.

Abstract:

A new algorithm of automatic extraction of thin structures in textured images is introduced, and, more speciﬁ-

cally, is applied to detection of road cracks. The method is based on two steps: the ﬁrst one consists in detecting

points of interest inside the thin structure whereas the second step connects the points with a geodesic contour

process. The main contribution of this work is the study of automatic detection of points of interest inside thin

structures in a high-textured background. The results are compared with a Markovian segmentation.

1 INTRODUCTION

This work concerns the detection of road cracks.

The threshold-based methods are simple but the re-

sults contain a lot of false detections (Koutsopoulos

and Downey, 1993). The methods based on mor-

phology allow to reduce false detections but strongly

depend on the parameter settings (Tanaka and Ue-

matsu, 1998). The learning-based methods alleviate

the problems of the two ﬁrst categories (Oliveira and

Correia, 2008) but they need a learning phase which

is not well appropriate to the application. The ﬁlter-

ing methods are the most recent ones (Oliveira and

Correia, 2009). In (Chambon et al., 2010), we have

introduced a segmentation based on a wavelet de-

composition and a Markovian modelling. To have a

dense detection with less false detections, we intro-

duce a Geodesic Contour-based method with initial

automatic detection of Points Of Interest (GCPOID)

inside the thin structure. We present, ﬁrst the data

set, second, the method GCPOID, and, the results and

conclusions.

2 DATA SET

Evaluation is done on 14 synthetic images and 42 real

ones. The synthetic images were built using, for 8,

a random dot texture and, for 6, real images without

cracks (2 with a static camera and 4 with a camera

embedded on a vehicle and controlled lights). The

(a) (b)

(c)

Figure 1: Tested images – one synthetic image (a), one real

mage with a simulated default (b) and one real image (c).

cracks were randomly added, with a random shape

and gray-levels (Figure 1). For real images, estimate

a reference segmentation is more complicated but the

images are more realistic than the synthetic ones. 4

experts have manually segmented the images and the

results are merged (Chambon et al., 2010). On the 42

real images, 17 were acquired using a static camera

and 25 using a dynamic system.

3 DESCRIPTION OF GCPOID

Existing methods have obtained quite good results,

particularly, with multiscale analysis (Chambonet al.,

2010). However, the results still present a lot of false

detections and incomplete detection of the cracks, see

Figure 2. Another way to deal with this problem is to

consider a sparse but reliable detection and, then, to

propagate this detection to obtain a dense detection.

In consequence, we propose a Geodesic Contour-

based method initialized by Points Of Interest Detec-

tion (GCPOID), an algorithm in 2-steps: (1) Detec-

tion of POIs (inside the cracks), (2) Propagation of

210

Chambon S..

DETECTION OF POINTS OF INTEREST FOR GEODESIC CONTOURS - Application on Road Images for Crack Detection.

DOI: 10.5220/0003333002100213

In Proceedings of the International Conference on Computer Vision Theory and Applications (VISAPP-2011), pages 210-213

ISBN: 978-989-8425-47-8

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

the initial detection and the goal of this paper is to

propose a detector of POIs efﬁcient for extracting thin

structures in textured environment.

IMAGE F379 IMAGE F382

Figure 2: Results obtained with Markovian Modelling

(MM) (Chambon et al., 2010) (in red) – This method is

based on a ﬁrst step with adaptive ﬁltering and a reﬁnement

step with a Markov model-based segmentation.

Detection of POIs for Road Cracks – We suppose

that points inside the cracks are the POIs in the im-

age because the crack is a signiﬁcant element of the

image, and, moreover, it is a high perturbation in the

road texture. Some existing detectors are based on

grey level, ﬁrst derivatives of the image (Harris and

Stephens, 1988) or second derivatives (Lowe, 2004).

In this work, the most adapted detectors seem to be:

Harris operator and grey-level auto-correlation(noted

AC) (Moravec, 1977). As an important point of this

study, we detail and comment the auto-correlation de-

tector whose principle is:

The grey level of a point of interest must be low cor-

related to the grey levels of its neighbors:

R(p

i, j

) = fusion

(k,l)∈V(i, j)

M(f

, f

), (1)

R depends on f

and f

, the two vectors that con-

tain, respectively, the grey levels of the studied pixel

i, j

, and a neighbor pixel p

k,l

and their respective

neighborhoods, M corresponds to a similarity crite-

rion (correlation measure), ”fusion” is an operator

to merge the correlation scores in the neighborhood

V(i, j) of the studied pixel. In consequence, we have

to determined: (V), the size of the neighborhood,

V(i, j), (S), the correlation measure and the size of

the correlation window (f

) and (parameter F), the

fusion operator.

Parameter V – For POIs located inside ﬁne cracks

on road images, it seems natural and appropriate to

take a 3× 3 square area centered on the studied pixel

for the size of V(i, j). A different size can be used in

an other kind of application.

Parameter S – We have studied the inﬂuence

of different correlation measures based on our

work (Chambon and Crouzil, 2003). For reason of

space, the experimentation is not described but it re-

veals the most appropriate measure, the Normalized

Cross Correlation (NCC), a scalar product between

the vectors f

and f

NCC(f

, f

) =

· f

kkf

. (2)

The Zero mean NCC (ZNCC) has been used but this

centered version of NCC does not afford discrimina-

tive values for R(p

i, j

Parameter F – Three different kinds of fusion have

been proposed:

(1) MAX – When the POIs correspond to the high-

est dissimilarity in the neighbors, equation (1) be-

comes:

R(p

i, j

) = max

(k,l)∈V(i, j)

M(f

, f

). (3)

(2) MAX-MIN – We suppose that a crack POI has

the highest local dissimilarity but also the highest

local contrast and equation (1) becomes:

R(p

i, j

) = max

(k,l)∈V(i, j)

M(f

, f

)

− min

′

)∈V(i, j)

M(f

, f

′

(4)

(3) ORIENTATION – We suppose that the contrast

should be oriented and equation (4) becomes:

R(p

i, j

) =argmax

( max

)

M(f

i, j

, f

)

− min

)

M(f

i, j

, f

)),

(5)

where the orientation o corresponds to one possi-

ble orientation over [0;

;

3π

] and o = π + o.

For each detector, the N

ﬁrst responses (global selec-

tion) are chosen for extracting the N

selected POIs.

The inﬂuence of N

is studied in § 4.

Geodesic Contour Propagation – The goal is to

ﬁnd a curve that ﬁts an object or a region by itera-

tively minimizing a cost or energy function, E, of the

curve C, deﬁned by (Kass et al., 1988):

E(C ) =



′

(v)k

+ w

′′

(v)k

+ P (C (v)



dv,

(6)

where c(v) represents a curve drawn on a 2D im-

age, [0, L] its domain of deﬁnition, and L the length

of the curve. The two ﬁrst terms represent the inter-

nal forces (regularity of the curve) based on ﬁrst and

second derivatives with w

, w

the weights of each of

these two parts. The third term is the external force,

P . These choices can be made:

) P based on grey levels: P (C (v)) = I(v);

) P based on the gradients: P (C(v) = g(k∇I(v)k)

where g is a decreasing function.

DETECTION OF POINTS OF INTEREST FOR GEODESIC CONTOURS - Application on Road Images for Crack

Detection

211

When the evolution curve is based on a planar curve

evolution, this is a particular case, the geometric con-

tours (Delagnes and Barba, 1995) that are the most

efﬁcient in the presence of topology changes. To com-

bine the advantages of both methods (active and geo-

metric contours) geodesic contours (Cohen and Kim-

mel, 1997) have been introduced. In fact, it is a par-

ticular case of active contours where equation (6) be-

comes:

E(C ) =



′

(v)k

+ P (C (v)



dv, (7)

In examining the shape of the element to detect, it

seems very natural to use the geodesic contours algo-

rithm and shape (S

) has been chosen for P . Indeed,

the energy function associated to the curve depends

on the grey levels (for the external force) instead of

the gradients (that are not signiﬁcant in this kind of

images that are highly textured). A Fast Marching al-

gorithm is used for the optimisation. Moreover, we

suppose that the cracks are not opened. To rely the

POIs , denoted by p

, i ∈ {1;. . . ;N

}, we link the

points two by two, by pairs of nearest points:

1. G ← {p

}

2. p ← argmin

p∈G

(0,0)

− pk

where p

(0,0)

is the origin of the image

3. While G 6=

0 do

(a) G ← G − {p}

(b) p

← argmin

∈G

kp− p

(d) p ← p

4 EXPERIMENTAL RESULTS

Detection of Points of Interest – The results, cf.

Figure 3, show that for each method, the best results

are obtained when n = 10, and the larger the number

of POIs, the larger the errors. The maximal percent-

age of detections is always found with a window of

size 3 × 3. Moreover, auto-correlation always gives

better results than Harris, and, the best operator fu-

sion is the maximum. The orientation fusion leads

to results similar to those of the maximum. For 38

images (over the 55 tested), we obtained from 80%

to 100% of points that are well detected inside the

cracks. There is no image (over the 55), with a 3× 3

window size, where Harris is the best.

Crack Detection – These two methods are com-

pared: the proposed method GCPOID and the pre-

vious method based on Markovian Modelling and de-

noted by MM (Chambon et al., 2010). These criteria

Number of points of interest that are selected

Harris

Percentage of correct detections

NCC

MaxMin

Orientation

Figure 3: Percentage of POIs detected inside cracks – The

ean (solid line) and the maximum (dotted line) percent-

ages are computed with all the results obtained with the 55

images.

IMAGE F379

Harris Auto-correlation

IMAGE F382

Harris Auto-correlation

Figure 4: Detection of POIs (in red) in road images – The

left results are obtained with the detector of Harris whereas

the right, with the auto-correlation (AC) operator. For AC,

there is no errors, i.e. no POIs detected outside the cracks.

are used in order to compare the results: (1) the per-

centage of correct matches (TP), (2) the percentage

of false negatives (FN) which corresponds to pixels

that are not detected as cracks whereas they are inside

a crack, (3) the similarity coefﬁcient or Dice similar-

ity, DICE, deﬁned by:

2TP

FN+TP+P

, where P represents

the number of pixels that have been detected as crack

pixels. This criterion illustrates the proportion of cor-

rect detections compared to the whole detection and

it is important for our application because we want

to have an accurate and complete detection of the de-

faults. The higher the Dice (with maximum 1), the

better the result.

In the results of GCPOID, for all the images with

a percentage of correct detections of POIs of 100%

(27), there is very few false detections. Unfortunately,

the detection of the crack is incomplete. For the im-

ages with more than 80% of correct detections inside

the cracks, the results are also interesting, i.e better

than our initial method MM, cf. ﬁgure 5.

In Figure 6, a global analysis over all the tested

images is given. It highlights these two facts:

(1) GCPOID outperforms MM for the percentage of

correct matches (TP) ;

VISAPP 2011 - International Conference on Computer Vision Theory and Applications

212

IMAGE F379

IMAGE F382

Figure 5: Detection of road cracks with GCPOID – The

POIs, obtained with AC (NCC), are linked in red with

geodesic contours. Detection is incomplete but with very

few false detections when there is no wrong POI.

Figure 6: Results obtained with GCPOID and MM – These

results are represented by boxes with one box for each cri-

teria (TP, true positives, and FN, false negatives) and each

method. The boxes have lines at the lower quartile, the

minimum, the median, the maximum and the upper quar-

tile. These results show that the median of the percentages

of correct detections is highly improved with GCPOID, be-

cause it becomes 90.54% against 50.68% for MM whereas

the percentage of false negatives for GCPOID is higher (the

median is 55.63%) than for MM (the median is 15.68%).

(2) On the contrary, MM obtains lower percentages

of false negatives (FN) than with GCPOID.

This is why the mean Dice similarity is better with

MM, 0.54, than with GCPOID, 0.46. However, the

result with GCPOID is reliable: the mean of the cor-

rect detections reached 74.43% (MM obtained only

50.89%). In consequence, we can consider that we

have reached our ﬁrst goal: to obtain very few false

negatives. The second aim (complete detection) can

be obtained if we try to propagate the results of the

detection at the extremities of the crack detection, see

image F382 in Figure 5.

5 CONCLUSIONS

We introduced a new method of road crack detec-

tion based on two steps: automatic detection of

POIs followed by the estimation of geodesic con-

tours. The study reveals the most appropriate detec-

tor of POIs (by introducing the principle and the pa-

rameter settings): auto-correlation detector with nor-

malised cross correlation and fusion with the maxi-

mum. Moreover,we demonstrate the interest of using

geodesic contours for the ﬁnal detection by compar-

ing to a Markovian segmentation.

Our future work will focus on how adding more

POIs (to have a complete/dense detection), with

the combination of preﬁltered images with auto-

correlation detection, and how removing wrong POIs,

by adding some constraints (based on the variation

of grey levels in the neighborhood or on the shape

of the curves determined by each pair of POIs). For

the geodesic contours, we will propose a more real-

istic algorithm for connecting the POIs (possibility to

connect one point with numerous others, to remove a

point if the energy is low with all the paths passing

through this point).

REFERENCES

Chambon, S. and Crouzil, A. (2003). Dense matching using

correlation: new measures that are robust near occlu-

sions. In BMVC, volume 1, pages 143–152.

Chambon, S., Gourraud, C., Moliard, J., and Nicolle, P.

(2010). Road crack extraction with adapted ﬁltering

and markov model-based segmentation – introduction

and validation. In VISAPP.

Cohen, L. and Kimmel, R. (1997). Global minimum for ac-

tive contour models: A minimal path approach. IJCV,

24(1):57–78.

Delagnes, P. and Barba, D. (1995). A markov random ﬁeld

for rectilinearstructure extraction inpavement distress

image analysis. In ICIP, volume 1, pages 446–449.

Harris, C. and Stephens, M. (1988). A combined corner and

edge detector. In Alvey Vision Conf., pages 147–151.

Kass, M., Witkin, A., and Terzopoulos, D. (1988). Snakes:

Active contour models. IJCV, 1(4):321–331.

Koutsopoulos, H. and Downey, A. (1993). Primitive-based

classiﬁcation of pavement cracking images. Journal

of Transportation Engineering, 119(3):402–418.

Lowe, D. (2004). Distinctive image features from scale-

invariant keypoints. IJCV, 60(2):91–110.

Moravec, H. (1977). Toward automatic visual obstacle

avoidance. In IJCAI, volume 2, page 584.

Oliveira, H. and Correia, P. (2008). Supervised strategies

for cracks detection in images of road pavement ﬂexi-

ble surfaces. In EUSIPCO.

Oliveira, H. and Correia, P. (2009). Automatic road

crack segmentation using entropy and image dynamic

thresholding. In EUSIPCO.

Tanaka, N. and Uematsu, K. (1998). A crack detection

method in road surface images using morphology. In

Machine Vision Applications, pages 154–157.

DETECTION OF POINTS OF INTEREST FOR GEODESIC CONTOURS - Application on Road Images for Crack

Detection

213