Hellinger Kernel-based Distance and Local Image Region Descriptors for

Sky Region Detection from Fisheye Images

Y. El Merabet

, Y. Ruichek

, S. Ghaffarian

, Z. Samir

, T. Boujiha

, R. Touahni

, R. Messoussi

and A. Sbihi

Laboratoire LASTID, D

epartement de Physique, Facult

e des Sciences, Universit

e Ibn Tofail,

B.P 133, 14000, K

enitra, Morocco

Laboratoire IRTES-SeT, Universit

e de Technologie de Belfort-Montb

eliard,

13 rue Ernest Thierry-Mieg, 90010, Belfort, France

Faculty of Geo-Information Science and Earth Observation (ITC), University of Twente,

Enschede, 7500 AE, The Netherlands

Laboratoire LABTIC, ENSA, Universit

e Abdelmalek Essadi, Route Ziaten, km 10, BP 1818 Tanger, Morocco

Keywords:

GNSS, Region Classiﬁcation, Image Segmentation, Fisheye, Color Invariance, Hellinger Kernel, Local Image

Region Descriptors.

Abstract:

Characterizing GNSS signals reception environment using ﬁsheye camera oriented to the sky is one of the rel-

evant approaches which have been proposed to compensate the lack of performance of GNSS occurring when

operating in constrained environments (dense urbain areas). This solution consists, after classiﬁcation of ac-

quired images into two regions (sky and not-sky), in identifying satellites as line-of-sight (LOS) satellites or

non-line-of-sight (NLOS) satellites by repositioning the satellites in the classiﬁed images. This paper proposes

a region-based image classiﬁcation method through local image region descriptors and Hellinger kernel-based

distance. The objective is to try to improve results obtained previously by a state of the art method. The

proposed approach starts by simplifying the acquired image with a suitable couple of colorimetric invariant

and exponential transform. After that, a segmentation step is performed in order to extract from the simpliﬁed

image regions of interest using Statistical Region Merging method. The next step consists of characterizing

the obtained regions with local RGB color and a number of local color texture descriptors using image quan-

tization. Finally, the characterized regions are classiﬁed into sky and non sky regions by using supervised

M SR C (Maximal Similarity Based Region Classiﬁcation) method through Hellinger kernel-based distance.

Extensive experiments have been performed to prove the effectiveness of the proposed approach.

1 INTRODUCTION

GNSS systems (Global Navigation Satellites Sys-

tems), such as COMPASS, GLONASS, GPS and Eu-

ropean GALILEO have the potential to advance the

development of intelligent transport systems (ITSs)

and associated services. They contribute widely to

localization and navigation systems and provide use-

ful information that can for example be exploited in

a meaningful way in the ﬁeld of transport market

such as ﬂeet management, monitoring of containers,

etc. One of the main drawbacks of GNSSs systems

in constrained environments such as urban zones is

that signals may arrive at the receiver antenna only in

non-line-of-sight (NLOS) conditions. Indeed, even if

most of them give satisfying accuracy in terms of po-

sition of localization, they cannot avoid propagation

problems caused by multi-path phenomena (cf. Fig-

ure 1) of GNSS signals. This can be explained by the

fact that in dense environments, like city centers, sig-

nals can be shadowed (signal received after reﬂections

without any direct ray), blocked (no signal received)

and directly received. Consequently, the evaluation

of estimated position reliability remains challenging

in presence of these constraints. For applications like

containers monitoring, ﬂot management, etc., not re-

quiring high availability, integrity and accuracy of the

positioning system, this drawback is only a marginal

problem. But this constitutes a real challenge for spe-

ciﬁc applications dealing with liability issues (toll, in-

surance, etc.) as well as safety-related applications

(automatic guidance or control), requiring more strin-

gent performances. Several solutions have been pro-

posed in the literature in order to enhance localization

El Merabet Y., Ruichek Y., Ghaffarian S., Samir Z., Boujiha T., Touahni R., Messoussi R. and Sbihi A.

Hellinger Kernel-based Distance and Local Image Region Descriptors for Sky Region Detection from Fisheye Images.

DOI: 10.5220/0006092404190427

In Proceedings of the 12th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2017), pages 419-427

ISBN: 978-989-758-225-7

419

accuracy. One can cite those increasing the system

complexity like multi-sensor-based approaches that

consist to add other sensors (odometer, Inertial Mea-

surement Unit, etc.) to compensate the lack of per-

formance of GNSS (Wang and Gao, 2007; Lentmaier

et al., 2008). Using complementarity between local-

ization systems and computer vision to characterize

the environment of reception of satellites is another

workaround that is recently proposed (Marais et al.,

2013). The idea consists in analyzing the structure

of the environment traveled by a vehicle using a sin-

gle camera delivering visible range to overcome prob-

lems such as time computation, lack of precision of

3D models, etc. The approach relies, at each acqui-

sition of a ﬁsheye image using a wide-angle camera

(ﬁsheye camera with a large ﬁeld of view of 180

◦

)

mounted near to a GPS receiver on the roof of the ve-

hicle and oriented to the sky, on sequentially applying

of two major steps: 1/ image processing and 2/ repo-

sitioning. The objective of the ﬁrst step is to automat-

ically detect sky region from the acquired image. For

that, a geodesic reconstruction based technique with

an optimal contrast parameter is used in order to sim-

plify the ﬁsheye image. Then, the pixels of the ob-

tained simpliﬁed image are classiﬁed into two classes

(sky and not-sky) using image clustering. A set of

unsupervised (Km local, Fuzzy Cmeans, Fisher and

Statistical region Merging) and supervised (Bayes, K-

Nearest Neighbor and Support Vector Machine) clus-

tering algorithms have been tested and compared. In

the second step, the satellites are repositioned in clas-

siﬁed image to identify GNSS signals with direct path

(resp. blocked/reﬂected signals) i.e. located in sky re-

gion of the image (resp. located in not-sky region).

More details of this repositioning step can be found

in (Marais et al., 2013). Obviously, the reliability of

the proposed procedure depends greatly on the results

of image processing step, i.e. the classiﬁcation re-

sults. In this paper, our challenge consists in mak-

ing this step more effective in terms of image classi-

ﬁcation results. Figure 2 illustrates the ﬂowshart of

the image processing-based method for localization.

The method is composed of two major steps : image

processing and localization. Our contribution is con-

cerned with the image processing part. The method

we propose is composed of several substeps: 1/ im-

age simpliﬁcation, 2/ image segmentation, 3/ region

features extraction and 4/ region classiﬁcation.

The paper is organized as follows: Section II

presents the ﬁsheye image simpliﬁcation step of the

proposed procedure. Section III introduces brieﬂy

the Statistical Region Merging algorithm used to ob-

tain the preliminary ﬁsheye image segmentation. In

section IV, we introduce the implemented local color

RGB, color local texture and color hybrid histograms

as local image region descriptors. Section V presents

the proposed M S R C algorithm. Experimental re-

sults and a comparison with the method of (Marais

et al., 2013) are shown in Section VI. Conclusions are

derived in Section VII.

Figure 1: Illustration of the multipath phenomenon in urban

areas.

Figure 2: Flowshart of the image processing-based method

for localization.

2 FISHEYE IMAGE

SIMPLIFICATION

Image simpliﬁcation is a very important basic ingre-

dient of a lot of practical image-based applications.

This useful basic pre-processing step permits to re-

duce content information of an image by suppressing

undesired details such as noise. In this work, in order

to simplify the acquired ﬁsheye image, we opted to

use color invariance and exponential transform. The

sequential use of these two tools permits us to consid-

VISAPP 2017 - International Conference on Computer Vision Theory and Applications

420

erably reduce the over-segmentation problem and to

reach a more precise segmentation that is faithful to

the desired real objects.

2.1 Color Invariance

In mobility with a mobile platform, as is the case of

our application, several factors such as the surface

reﬂectance, illuminant color, lighting geometry, re-

sponse of the sensor, etc. (see Figure 4), may af-

fect the quality of the acquired images. Consequently,

attempting to segment the acquired image without

any pre-processing step, leads to a strongly over-

segmented image caused by insigniﬁcant structures or

noise. To overcome this shortcoming and therefore to

achieve satisfactory segmentation results (extract sky

regions with high accuracy), the trend towards obtain-

ing invariant signatures seems to be the best way for-

ward (El merabet et al., 2011; El merabet et al., 2014;

El merabet et al., 2015). Indeed, we propose to trans-

form the input ﬁsheye image using a colorimetric in-

variant in order to obtain a color-invariant ﬁsheye im-

age whatever the illumination conditions and artifacts

present in the acquired images (noise and unimportant

ﬁne-scale details). In this work, we have used Afﬁne

Normalization (AN) expressed below (Fusiello et al.,

1999). As it will be shown in section 6.2, tests have

validated the interest in using this colorimetric invari-

ant in the simpliﬁcation process.











(p) =

(p)−µ(I

(p))

std(I

(p))

(p) =

(p)−µ(I

(p))

std(I

(p))

(p) =

(p)−µ(I

(p))

std(I

(p))

(1)

where I

(p) is the pixel value at position p in

the color component K={R, G, B}, µ(I

(p)) and

std(I

(p)) are, respectively, the mean value and stan-

dard deviation calculated in a window of interest W

centered on the pixel p. This normalization ensures

the invariance under afﬁne changes of illumination

that is achieved by independently normalizing each

channel to have zero mean and unit variance.

2.2 Exponential Transform (ET)

Besides obtaining invariant signatures, image en-

hancement is another effective technique allowing to

improve the robustness of image simpliﬁcation pro-

cess. The principal objective of this second compo-

nent of image simpliﬁcation module is to modify at-

tributes of an image to make it more suitable for the

considered application. In this paper, in order to efﬁ-

ciently improve segmentation quality results of ﬁsh-

eye images, we have opted to use the exponential

transform (ET). ET permits to approximate the expo-

nential correction factor of grayscale images which

maximizes the contrast of the images in the class of

exponential intensity mapping functions. Mathemati-

cally, ET is given by (cf. Eq. 2):







i j

= exp(χ/ξ) − 1 + I

min

χ = I

i j

− I

min

ξ = (I

max

− I

min

)/(log(I

max

− I

min

+ 1))

(2)

where I

i j

is the intensity of the pixel at position

(i, j), I

max

and I

min

are the highest and lowest inten-

sities of the image I, respectively and ξ is a normal-

ization factor for stretching output values between the

new lowest I

min

and highest I

max

intensities of the re-

sultant image I

3 PRELIMINARY FISHEYE

IMAGE SEGMENTATION

As indicated previously, the second step of our ap-

proach relies on image segmentation in order to seg-

ment the simpliﬁed images into homogeneous regions

with the same properties. Obviously, the quality

of the classiﬁcation results which are the output of

the proposed region based classiﬁcation procedure is

strongly dependent on the segmentation results. In

this paper, in order to obtain a preliminary ﬁsheye

image segmentation, we have used SRM (Statisti-

cal Region Merging) algorithm (Nock and Nielsen,

2004) that seems to be more adapted when consid-

ering the objectives of our application. Indeed, using

this method, we can correctly extracts all signiﬁcant

regions where the boundaries hypothesized coincide

with the signiﬁcant segment boundaries in the simpli-

ﬁed ﬁsheye images. SRM algorithm presents several

advantages : 1/ it dispenses dynamical maintenance of

region adjacency graph (RAG); 2/ it allows deﬁning

a hierarchy of partitions; 3/ it runs in linear-time by

using bucket sorting algorithm while transversing the

RAG and 4/ it not only considers spectral, shape and

scale information, but also has the ability to cope with

signiﬁcant noise corruption and handle occlusions.

4 REGION FEATURES

This stage of our approach consists in characterizing

the segmented regions, obtained by using SRM algo-

rithm, with suitable descriptors to identify the regions

corresponding to the sky in ﬁsheye images. The ex-

traction of these descriptors, used as inputs to the pro-

posed region based image classiﬁcation, permits to

Hellinger Kernel-based Distance and Local Image Region Descriptors for Sky Region Detection from Fisheye Images

421

provide a global representation of a region that is a

higher level than that of the raw image pixels allow-

ing thus, to discriminate robustly between the differ-

ent regions in the treated images. The descriptors we

have choosen for the tests are explained below.

4.1 Local Color RGB Histograms

In spite of the fact that RGB color histogram is very

simple, easiest feature to implement and low level

method, it still being relevant for color based image

classiﬁcation. In the present study, RGB color his-

togram is implemented as follows: each RGB color

channel is ﬁrst uniformly quantized into l levels, af-

ter that, the color histogram of each segmented re-

gion is produced in the feature space of z = l × l × l

bins. Let I be an image containing N pixels quantized

in z=16×16×16=4096 color bins, the RGB color his-

togram of a segmented region R is represented as

RGB

(R ) = [H

,...., H

] (3)

where

(

∑

j=1

i| j

);j ∈ R and 0 6 i 6 z. (4)

is the ith normalized histogram bin and

τ=card(R ) is the number of pixels in the region R .

i| j

is the conditional probability of the selected jth

pixel belonging to the ith color bin. It is expressed as

follows:

i| j



1,if the jth pixel is quantized into the ith color bin

0,otherwise.

(5)

4.2 Color Local Texture Histograms

Texture analysis plays an important role in many dis-

ciplines and related applications: defect detection

and food inspection, surface grading, computer as-

sisted diagnosis, remote sensing, etc. (Dornaika et al.,

2016), due to its potential in extracting prominent fea-

tures with very high discriminating power.

It ressorts from the literature that the most pro-

posed texture descriptors are developed for gray level

images. Exploiting color aspects of textured images,

which is one of the objectives of this work, has un-

fortunately received much less attention (Chao et al.,

2013; Choi et al., 2012). Generally, the choice of suit-

able texture descriptor is closely related to the partic-

ularities of objects to be extracted from the image. In

this work, we propose to investigate the impact of sev-

eral well-known texture descriptors on the outcome

of the proposed region based ﬁsheye image classiﬁca-

tion. A set of 10 texture features calculated for each

region of the segmented ﬁsheye image are used and

tested (cf. Table 1). In order to incorporate color in-

formation, these texture descriptors are extended to

RGB color space producing thus 10 color local tex-

ture histograms (H

LBP

, H

CSLBP

, H

LDP

, H

ILT P

, etc.).

The method consists in calculating the unichrome tex-

ture feature independently over different channels in

RGB color space, concatenate them to get a descrip-

tor color image (for example the LBP color image in

Figure 3) and then the color local texture histogram

is calculated for each region of the segmented image

by following the same steps as for RGB color his-

tograms, as shown in Figure 3.

Original Image

Channel B

LBP color

Image

Quantization

LBP descriptor

over each channel

Channel

extraction

Final color LBP

histograms over each

segmented region

, R

.....R

)

LBP descriptors

concatenation

Channel GChannel R

LBP R LBP G

LBP B

Segmented

image

Figure 3: Calculation of color LBP histogram over each

region of the segmented image.

4.3 Local Hybrid Histograms

In order to build local hybrid histograms, we pro-

pose to concatenate the feature vectors provided by

different descriptors. Given two local color his-

tograms H (R ) = [H

, H

,...., H

] and H

(R ) =

,...., H

] of a region R , the corresponding

local hybrid color histogram is mathematically repre-

sented as:

hyb

(R ) =[H

]

=[H

,.., H

]

(6)

where H

is the ith histogram bins and the dimen-

sion of the obtained H

hyb

(R ) will be (2 × z) sized;

z=16

=4096.

VISAPP 2017 - International Conference on Computer Vision Theory and Applications

422

Table 1: Texture descriptors extracted and used in this work.

Descriptors Acronym Ref.

Local binary patterns LBP (Ojala et al., 1996)

Orthogonal combination of local binary pattern OCLBP (Chao et al., 2013)

Local quinary patterns LQP (Nanni et al., 2010b)

Local ternary patterns LTP (Tan and Triggs, 2007)

Improved local ternary patterns ILTP (Nanni et al., 2010a)

Center-symmetric local binary patterns CS-LBP (Heikkil

a et al., 2006)

3D Local binary patterns 3DLBP (Huang et al., 2006)

Local derivative pattern LDP (Zhang et al., 2010)

Local phase quantization LPQ (Ojansivu et al., 2008)

Sum and difference histograms SDH (Unser, 1986)

In the present work, each one of the proposed

color local texture histograms is concatenated with

the color RGB histogram leading to 10 hybrid color

texture histograms (H

RGB

LBP

, H

RGB

ILT P

, H

RGB

OCLBP

, H

RGB

LPQ

etc.). While this can enrich the discrimination capac-

ity of the resulting descriptor, it has the disadvantage

that the dimensionality of the resulting feature vector

could be very high, increasing thus the computation

time.

5 MAXIMAL SIMILARITY

BASED REGION

CLASSIFICATION

Since the segmented regions M

SRM

, obtained via

SRM algorithm, are now characterized with the de-

scriptors introduced previously, our challenge is to

classify them into sky and non-sky regions. For this

purpose, we need to calculate similarity between the

characterized regions (R ∈ M

SRM

) and those of two

learning databases B

obj

and B

back

that are constructed

respectively with m distinctive textures of sky regions

and n distinctive textures of non-sky regions such

as building, road, tree, etc. We have then to deﬁne

a similarity measure Ψ(R ,Q ) allowing to calculate

similarity between two regions R and Q basing on

their descriptors. Given two image region features

and f

, Ψ( f

, f

) considers these image region

features as points in the vector space and calculate

close degree of two points. There are several well-

known goodness-of-ﬁt statistical metrics in the litera-

ture. One can cite second type distance (Stricker and

Orengo, 1995), log-likelihood ratio statistic (Fuku-

naga, 1990), Minkowski measure, histogram intersec-

tion method (Swain and Ballard, 2002), Hellinger dis-

tance (Kailath, 1967; Ninga et al., 2010), etc. Let H

be the normalized histogram of a region R , the super-

script i represents its i

element. z = l × l × l = 4096

represents the feature space. In this study, we have

used Hellinger kernel (also known as Bhattacharyya

coefﬁcient given by Eq. 7), which represents the co-

sine of angle between the unit vectors representing the

two regions to be compared:

(

,........,

)

and

(

,........,

)

The higher the Hellinger distance Ψ(R ,Q ) be-

tween regions R and Q is, the higher the similarity

between them is. That is to say that the angle be-

tween the two histogram vectors is very small involv-

ing that their histograms are very similar. Certainly,

two similar histograms do not necessarily involve that

the two corresponding regions are perceptually simi-

lar. Nevertheless, coupled with the proposed M SR C

algorithm summarized in algorithm 1, Hellinger ker-

nel based similarity works well in the proposed ap-

proach.

Ψ(R ,Q ) =

∑

i=1

(7)

In our implementation, the main assumption of the

proposed M SR C algorithm is that it assigns an un-

known region R to a class C

, if the average of the K

ﬁrst high similarity measures calculated between the

region R and the regions of the learning database cor-

responding to the class C

is maximal, i.e,

∗

(R ) = argmax

∈C

∑

i=1

Ψ(R ,Q

), Q

∈ B

(8)

where B

,..., B

are the learning databases corre-

sponding to the classes C

,...,C

, R is a query,

and Ψ is the similarity measure.

6 RESULTS AND DISCUSSION

To evaluate and shown the effectiveness of the pro-

posed M SR C method, the RGB color and the pro-

Hellinger Kernel-based Distance and Local Image Region Descriptors for Sky Region Detection from Fisheye Images

423

Algorithm 1: Maximal similarity based region classiﬁcation.

Require: I ← The set M

SRM

of segmented regions (after simpliﬁcation of the input image using AN/ET conjointly).

obj

← learning database of sky regions.

back

← learning database of non sky regions (building, road, tree, etc.).

1: Calculate the local image region descriptor for all regions of M

SRM

and for those composing B

obj

and B

back

2: for each candidate region R ∈ M

SRM

3: Calculate V

ob j

= {Ψ(R , Q

);(Q

)

i=1..m

∈ B

obj

}, the similarity vector between R and B

obj

. Ψ(R , Q

) is the similarity

between R and the region Q

∈ B

obj

4: Calculate V

back

= {Ψ(R , Q

);(Q

)

j=1..n

∈ B

back

}, the similarity vector between R and B

back

. Ψ(R , Q

) is the

similarity between R and the region Q

∈ B

back

5: Get the order of V

ob j

and V

back

by decreasing sorting;

6: Calculate µ

ob j

∑

i=1

Ψ(R ,Q

)

, K ≤ m, the mean of the K ﬁrst elements of V

ob j

7: Calculate µ

back

∑

j=1

Ψ(R ,Q

)

, K ≤ n, the mean of the K ﬁrst elements of V

back

8: if (µ

ob j

≥ µ

back

) then

9: The region R has the maximal similarity with B

obj

, it is then classiﬁed as a part of sky regions.

10: else

11: The region R has the maximal similarity with B

back

, it is then classiﬁed as a part of background.

12: end if

13: end for

14: return The ﬁnal segmentation map.

posed groups of color local texture and local hybrid

image region descriptors for classiﬁcation of regions

extracted from ﬁsheye images into two class (sky or

non sky regions), several tests were carried out. As

pointed out in the introduction, the main objective of

this work consists in improving the results presented

in (Marais et al., 2013). For that, a comparison step

was also performed to show the improvement that our

method provides.

6.1 Dataset

The image database acquired in the framework of

the CAPLOC project (Marais et al., 2013), was

used to validate the proposed sky detection approach

from ﬁsheye images. It has heterogeneous data and

varying complex scenarios (overexposure, brightness

changes, vegetation abundance, urban canyon, etc.).

The image database contains 150 ﬁsheye images ex-

hibiting the mentioned various complex situations.

Figure 4 Illustrates Six images of the database.

Figure 4: Six images of the database (acquired in the frame-

work of the CAPLOC project).

6.2 Impact of the Couple AN/ET

As indicated in Section 2, the acquired ﬁsheye image

is ﬁrstly simpliﬁed using the couple of AN/ET where

the objective is to limit illumination changes and thus

reduce over-segmentation problem.

Table 2 highlights the ability of the couple AN/ET

to reduce the number of segmented regions. It appears

from this Table that when using Afﬁne Normalization

and the Exponential Transform conjointly, the num-

ber of regions is considerably reduced. Indeed, the

use of this couple leads to a number of segmented re-

gions greatly reduced (in average 10.43 regions per

image) in comparison with that obtained without any

simpliﬁcation step. The reduction rate is 73.46%,

i.e. the couple of AN/ET permits to eliminate (38.08-

10.43)*150 images=4147.5 undesirable regions. In

addition, the couple AN/ET provides good classiﬁca-

tion results where the produced classiﬁed images are

most close to the corresponding ground truth (judged

by the evaluation results given in Table 2). Indeed, the

couple AN/ET allows obtaining 99.17% as a value for

accuracy measure Vs 98.98% Vs 99.02% Vs 99.00%

using AN without ET, RGB/ET and RGB without ET

respectively.

6.3 M SR C and Descriptors

Performance

In this section, we study the ability of the proposed

M S R C method and the proposed groups of local im-

age region descriptors to classify all regions of the

segmented image into sky and non sky class.

VISAPP 2017 - International Conference on Computer Vision Theory and Applications

424

Table 2: Average number of segmented regions per image

and performance of the M S R C method using OCLBP as

local image region descriptor.

RGB RGB/ET AN AN/ET

Number

of regions

38.08 12.20 11.23 10.43

Accuracy

(%)

99.00 99.02 98.98 99.17

In Table 3, we report the obtained experimental re-

sults according to the local color RGB, color local tex-

ture and color hybrid histograms used. The analysis

of the accuracy measure leads us to highlight the in-

ﬂuence of the use of hybrid descriptors. Indeed, when

used in a hybrid form with RGB color histogram, all

tested color local texture histograms give good re-

sults as they permit to increase the classiﬁcation rate.

However, it is worthy to notice that the RGB ∪ LBP

and RGB ∪ ILT P descriptors are the most promising

because they give the maximum classiﬁcation rates

compared to the other tested descriptors.

6.4 Comparative Evaluation and

Discussion

As pointed out in Section 1, the main aim of this

study is to improve the performance of the method in-

troduced within the CAPLOC project (Marais et al.,

2013). In this framework, the proposed approach

mainly consists in classifying the acquired ﬁsheye im-

ages (after a geodesic reconstruction based simpliﬁ-

cation) into sky and non-sky class. In this work, the

authors have compared the performance of different

well known clustering algorithms including unsuper-

vised (Fisher, KMlocal, Fuzzy-Cmeans, SRM) and

supervised (Bayes, KNN, SVM) classiﬁers. More de-

tails related to this work can be found in (Marais et al.,

2013). This section is intended to compare our pro-

posed approach with the framework of (Marais et al.,

2013) used with all these clustering algorithms.

Figure 5 highlights the recognition accuracy ob-

tained for each tested method. It can be seen that our

approach shows higher performance than the method

in (Marais et al., 2013) used with the tested popu-

lar classiﬁers. Indeed, the proposed method permits

to reach 99.24% as a value for accuracy measure vs

97,71% with Fisher vs 97.67% with KNN, etc. This

shows that our method allows to increase the accuracy

with 1,53%.

We support these results by the illustration of Fig-

ure 6 that shows some examples of visual comparison

of the region classiﬁcation results. It appears from

this visual analysis that the proposed method demon-

strates excellent precision in terms of sky boundary

extraction. In fact, for all of the images in the ﬁrst

row of Figure 6, the produced sky extraction results

agree most closely with the corresponding ground

truth. Note that the majority of the sky regions are

detected with good boundary delineation whereas the

method introduced in (Marais et al., 2013) leads to

many false positives on buildings and vegetation ar-

eas and false negatives within sky area accompanied

by a loss of several parts of sky.

Figure 5: Classiﬁcation results according to the method of

(Marais et al., 2013) used with the tested popular classiﬁers

and according to our proposed method.

7 CONCLUSION

The paper introduced a complete processing chain for

horizon line detection from ﬁsheye images, in the

framework of enhancing GNSS based localization.

The proposed method relies on a region based clas-

siﬁcation method using Hellinger kernel-based dis-

tance and a number of color, texture and hybrid im-

age region descriptors. As a ﬁrst step, we proposed

to simplify the input images using Afﬁne Normaliza-

tion coupled with Exponential Transform in order to

limit illumination changes (shadows, brightness, etc.)

affecting the acquired ﬁsheye images. Then, we have

introduced several new local color descriptors for im-

age region description based on texture features in

order to characterize the regions obtained with the

SRM algorithm. RGB color, a number of 10 color lo-

cal texture and 10 local hybrid histograms have been

introduced. As a ﬁnal step, a Maximal Similarity

Based Region Classiﬁcation using Hellinger kernel-

based distance has been proposed in order to identify

candidate regions corresponding to targeted objects

(sky regions in our case). The proposed approach was

tested on image database containing 150 ﬁsheye im-

ages allowing to obtain satisfying results with higher

performance than that obtained with the method of

(Marais et al., 2013) tested with several unsupervised

(KMlocal, Fisher, fuzzy C-means and SRM) and su-

pervised (Bayes, KNN and SVM) classiﬁers.

Hellinger Kernel-based Distance and Local Image Region Descriptors for Sky Region Detection from Fisheye Images

425

Table 3: Classiﬁcation results according to the local color RGB histogram, color local texture and color hybrid histograms

used.

Descriptors Accuracy (%) Hybrid Descriptors Accuracy (%)

RGB 99.13 - -

LBP 98.43 RGB ∪ LBP 99.24

CSLBP 96.92 RGB ∪ CSLBP 97.97

LPQ 93.42 RGB ∪ LPQ 99.18

3DLBP 93.33 RGB ∪ 3DLBP 97.97

LQP 96.34 RGB ∪ LQP 99.18

SDH 99.07 RGB ∪ SDH 99.13

LDP 96.11 RGB ∪ LDP 98.23

OCLBP 99.17 RGB ∪ OCLBP 99.22

LTP 97.29 RGB ∪ LTP 97.27

ILTP 97.81 RGB ∪ ILTP 99.24

Figure 6: Visual comparison of region classiﬁcation results. acquired images (ﬁrst row); classiﬁed image into two classes (sky

and non-sky) obtained by the best classiﬁer defained in (Marais et al., 2013) (second row) and classiﬁcation result obtained

by the proposed M SR C approach (third row).

In future works, we envisage to extend the pro-

posed approach to other data sets related to applica-

tions dealing with automatic objects recognition.

REFERENCES

Chao, Z., Charles, E., and Liming, C. (2013). Image re-

gion description using orthogonal combination of lo-

cal binary patterns enhanced with color information.

Pattern Recognition, 46:1949–1963.

Choi, J., Ro, Y., and Plataniotis, K. (2012). Color local tex-

ture features for color face recognition. IEEE Trans.

Image Proces, 21(3):1366–1380.

Dornaika, F., Moujahid, A., El merabet, Y., and Ruichek,

Y. (2016). Building detection from orthophotos using

a machine learning approach: An empirical study on

image segmentation and descriptors. Expert Syst Appl,

58:130–142.

El merabet, Y., Meurie, C., Ruichek, Y., Sbihi, A., and

Touahni, R. (2011). Orthophotoplan segmentation and

colorimetric invariants for roof detection. In Interna-

tional Conference on Image Analysis and Processing,

Springer Berlin Heidelberg, pages 394–403.

El merabet, Y., Meurie, C., Ruichek, Y., Sbihi, A., and

Touahni, R. (2014). Segmentation d’images a

eriennes

par coop

eration lpe-r

egions et lpe-contours. applica-

tion

a la caract

erisation de toitures. Revue Francaise

de Photogrammetrie et de Teledetection, (206):9–44.

El merabet, Y., Meurie, C., Ruichek, Y., Sbihi, A., and

Touahni, R. (2015). Building roof segmentation from

aerial images using a line and region-based watershed

segmentation technique. Sensors, 15:3172–3203.

Fukunaga, K. (1990). Introduction to statistical pattern

recognition. Second ed., Academic Press.

Fusiello, A., Trucco, E., Tommasini, T., and Roberto, V.

(1999). Improving feature tracking with robust statis-

tics. Pattern Analysis and Applications, 2:312–320.

Heikkil

a, M., Pietik

ainen, M., and Schmid, C. (2006). De-

scription of interest regions with center-symmetric lo-

cal binary patterns. In Madurai, I. L., editor, 5th In-

dian Conference on Computer Vision, Graphics and

Image Processing, volume 4338, pages 58–69.

VISAPP 2017 - International Conference on Computer Vision Theory and Applications

426

Huang, Y., Wang, Y., and Tan, T. (2006). Combining statis-

tics of geometrical and correlative features for 3d face

recognition. In 17th British Machine Vision Confer-

ence, pages 879–888.

Kailath, T. (1967). The divergence and bhattacharyya dis-

tance measures in signal selection. IEEE Transactions

on Communications Technology, 15:52–60.

Lentmaier, M., Krach, B., and Robertson, P. (2008).

Bayesian time delay estimation of gnss signals in dy-

namic multipath environments. International Journal

of Navigation and Observation, vol. 2008.

Marais, J., Meurie, C., Attia, D., Ruichek, Y., and Flanc-

quart, A. (2013). Toward accurate localization in

guided transport: Combining gnss data and imaging

information. Transportation Research Part C Emerg-

ing Technologies, 43:88–197.

Nanni, L., Brahnam, S., and Lumini, A. (2010a). A local

approach based on a local binary patterns variant tex-

ture descriptor for classifying pain states. Expert Syst.

Appl, 37(12):7888–7894.

Nanni, L., Lumini, A., and Brahnam, S. (2010b). Local bi-

nary patterns variants as texture descriptors for medi-

cal image analysis. Artif. Intell. Med, 49(2):117–125.

Ninga, J., Zhanga, L., and Wub, D. (2010). Interactive im-

age segmentation by maximal similarity based region

merging. Pattern Recognition, 43(2):445–456.

Nock, R. and Nielsen, F. (2004). Statistical region merg-

ing. IEEE Transactions on Pattern Analysis Machine

Intelligence, 26:1452–1458.

Ojala, T., Pietik

ainen, M., and Harwood, D. (1996). A

comparative study of texture measures with classiﬁca-

tion based on feature distributions. Pattern Recognit,

29(1):51–59.

Ojansivu, V., , and Heikkila, J. (2008). Blur insensitive tex-

ture classiﬁcation using local phase quantization. In

ICISP, pages 236–243.

Stricker, M. and Orengo, M. (1995). Similarity of color

images. In SPIE Storage and Retrieval for Image and

Video Databases, volume 3, pages 381–392.

Swain, M. and Ballard, D. (2002). Color indexing. Interna-

tional Journal of Computer Vision, 7:11–32.

Tan, X. and Triggs, B. (2007). Enhanced local texture fea-

ture sets for face recognition under difﬁcult lighting

conditions. In LNCS, editor, Analysis and Modelling

of Faces and Gestures, volume 4778, pages 168–182.

Unser, M. (1986). Sum and difference histograms for tex-

ture classiﬁcation. IEEE Trans. Pattern Anal. Mach.

Intell . PAMI-8(1), pages 118–125.

Wang, J. and Gao, Y. (2007). High-sensitivity gps data

classiﬁcation based on signal degradation conditions.

IEEE Trans. Veh. Technol, 56(2):566–574.

Zhang, B., Gao, Y., Zhao, S., and Liu, J. (2010). Lo-

cal derivative pattern versus local binary pattern: face

recognition with high-order local pattern descriptor.

IEEE Trans. Image Proces, 19(2):533–544.

Hellinger Kernel-based Distance and Local Image Region Descriptors for Sky Region Detection from Fisheye Images

427