Selec ion o ou books indexed in he Book Ci a ion Index
in Web o Science™ Co e Collec ion (BKCI)
In e es ed in publishing wi h us?
Con ac book.depa men @in echopen.com
Numbe s displayed abo e a e based on la es da a collec ed.
Fo mo e in o ma ion isi www.in echopen.com
Open access books a ailable
Coun ies deli e ed o Con ibu o s om op 500 uni e si ies
In e na ional au ho s and edi o s
Ou au ho s a e among he
mos ci ed scien is s
Downloads
We a e In echOpen, he wo ld’s leading publishe o Open Access booksBuil by scien is s, o scien is s
12.2%
180,000
195M
TOP 1%
154
6,700
12
Scene Recogni ion h ough Visual A en ion
and Image Fea u es: A Compa ison be ween
SIFT and SURF App oaches
Fe nando López-Ga cía1, Xosé Ramón Fdez-Vidal2,
Xosé Manuel Pa do2 and Raquel Dosil2
1Uni e sidad Poli écnica de Valencia
2Uni e sidade de San iago de Compos ela
Spain
1. In oduc ion
In his wo k we s udy how we can use a no el model o spa ial saliency ( isual a en ion)
combined wi h image ea u es o signi ican ly accele a e a scene ecogni ion applica ion
and, a he same ime, p ese e ecogni ion pe o mance. To do so, we use a mobile obo -
like applica ion whe e scene ecogni ion is ca ied ou h ough he use o image ea u es o
cha ac e ize he di e en scena ios, and he Nea es Neighbo ule o ca y ou he
classi ica ion. SIFT and SURF a e wo ecen and compe i i e al e na i es o image local
ea u ing ha we compa e h ough ex ensi e expe imen al wo k. Resul s om he
expe imen s show ha SIFT ea u es pe o m signi ican ly be e han SURF ea u es
achie ing impo an educ ions in he size o he da abase o p o o ypes wi hou signi ican
losses in ecogni ion pe o mance, and hus, accele a ing scene ecogni ion. Also, om he
expe imen s i is concluded ha SURF ea u es a e less dis inc i e when using e y la ge
da abases o in e es poin s, as i occu s in he p esen case.
Visual a en ion is he p ocess by which he Human Visual Sys em (HVS) is able o selec
om a gi en scene egions o in e es ha con ain salien in o ma ion, and hus, educe he
amoun o in o ma ion o be p ocessed (T eisman, 1980; Koch, 1985). In he las decade,
se e al compu a ional models biologically mo i a ed ha e been eleased o implemen
isual a en ion in image and ideo p ocessing (I i, 2000; Ga cía-Díaz, 2008). Visual
a en ion has also been used o imp o e objec ecogni ion and scene analysis (Bonaiu o,
2005; Wal he , 2005). In his chap e , we s udy he u ili y o using a no el model o spa ial
saliency o imp o e a scene ecogni ion applica ion by educing he amoun o p o o ypes
needed o ca y ou he classi ica ion ask. The applica ion is based on mobile obo -like
ideo sequences aken in indoo acili ies o med by se e al ooms and halls. The aim is o
ecognize he di e en scena ios in o de o p o ide he mobile obo sys em wi h gene al
loca ion da a.
The isual a en ion app oach is a no el model o bo om-up saliency ha uses local phase
in o ma ion o he inpu da a whe e he s a is ic in o ma ion o second o de is dele ed o
achie e a Re inop ical map o saliency. The p oposed app oach join s compu a ional
mechanisms o he wo hypo heses la gely accep ed in ea ly ision: i s , he e icien coding
www.in echopen.com
Objec Recogni ion
186
(Ba low, 1961; A nea e, 1954), which pos ula es ha he mission o he i s s ages o he
isual p ocessing chain is o educe he edundancy o p edic abili y in he incoming da a;
and second, in he isual co ex ele an a ibu es o he image a e ea ly de ec ed using
local phase o ene gy analysis, such as edges o objec s. A hose poin s whe e hese ea u es
a e loca ed he e is an alignmen o he local phase o he Fou ie ha monics (Phase
Cong uency). The model o local ene gy o de ec ea u es (Mo one & Bu , 1988; Mo one
& Owens, 1987; Ko esi, 1999) is based on his idea and demons a ed i s sui abili y o
pe cep ual appea ance and image segmen a ion. Ne e heless, i is no able o p io i ize he
ea u es wi h ega ds o he isual saliency. This ac is illus a ed in Figu e 1, whe e he
inpu image is o med by ba s ha inc emen i s o ien a ion in s eps o 10º om le o igh
and op o bo om, excep o he cen al ba ha b eaks his pe iodici y c ea ing a pop-ou
e ec o he HVS.
a) O iginal Image. b) Salience om PC. c) Salience om ou model.
Fig. 1. Saliency maps o he o iginal image; (a) om Phase Cong uency (b) and he p oposed
model (c).
In Fig. 1b we see he map o saliency achie ed using Ko esi’s model (Ko esi, 1999) based on
Phase Cong uency (PC). I p o ides app oxima ely equal weigh o all ea u es clea ing
away he pop-ou e ec . We hink he eason o ha is he high edundancy in images
which implies co ela ions and hus Gaussianism in ch oma ic and spa ial componen s. I is
known ha o handle in o ma ion abou phase s uc u e is equi alen o use non-Gaussian
in o ma ion in da a dis ibu ion (Hy ä inen e al., 2009). Thus, o ocus in he in o ma ion
ha does no depend on co a iances (local phase) i is necessa y o educe edundancy, ha
is, o deco ela e he da a. One way is h ough da a whi ening. Redundancy in RGB colo
componen s is dele ed h ough PCA and spa ial edundancy is a oided using an s a egy o
il e -based whi ening in equency domain. In Fig. 1c i is shown ha his hypo hesis wo ks
making possible o p io i ize he salience o isual ea u es om local phase.
Scene ecogni ion is pe o med using SIFT (Lowe, 2004) and SURF (Bay, 2008) o image
ea u ing ( wo di e en app oaches ha we compa e) and he Nea es Neighbo ule o
classi ica ion. SIFT ea u es a e dis inc i e image ea u es ha a e in a ian o image scale
and o a ion, and pa ially in a ian o change in illumina ion and 3D iewpoin . They a e
as o compu e and obus o dis up ions due o occlusion, clu e o noise. SIFT ea u es
ha e p o en o be use ul in many objec ecogni ion applica ions and cu en ly hey a e
conside ed he s a e-o - he a o gene al pu pose and eal-wo ld objec lea ning and
ecogni ion, oge he wi h SURF ea u es. SURF is a obus image desc ip o used in
www.in echopen.com
Scene Recogni ion h ough Visual A en ion
and Image Fea u es: A Compa ison be ween SIFT and SURF App oaches
187
compu e ision asks like objec ecogni ion o 3D econs uc ion. The s anda d e sion o
SURF is se e al imes as e han SIFT and i is claimed by i s au ho s o be mo e obus
agains di e en image ans o ma ions han SIFT. Howe e , he esul s o ou expe imen al
wo k showed ha SIFT ea u es pe o m signi ican ly be e han SURF ea u es. In
combina ion wi h saliency maps, SIFT ea u es lead o d as ic educ ions in he numbe o
in e es poin s in oduced in he da abase o p o o ypes (used in 1-NN classi ica ion), also
achie ing e y good pe o mance in scene ecogni ion. Thus, since he compu ing cos s o
classi ica ion a e signi ican ly educed he scene ecogni ion is accele a ed.
The chap e is de eloped as ollows. Nex Sec ion p esen s he model o spa ial saliency. An
o e iew o he image ea u ing me hods is p o ided in Sec ion 3. Sec ion 4 deals wi h he
scene ecogni ion applica ion. Expe imen al wo k and esul s a e p esen ed in Sec ion 5.
Finally, Sec ion 6 is de o ed o conclusions.
2. Model o spa ial saliency
Figu e 2 shows a gene al low diag am o he saliency model. Following we desc ibe each
s age o he model.
2.1 Ea ly s age
The goal o his ini ial s age is o dele e he s a is ical in o ma ion o second o de in colo
componen s (RGB) and spa ial componen s (be ween pixels o each colo componen ),
h ough di e en whi ening p ocesses.
The aim o he ini ial s ep in his s age is o p o ide he model wi h a colo space ha con ains
a mechanism, biologically inspi ed, called sho - e m adap a ion (Simoncelli & Olshausen, 2001;
Ba low & Foldiak, 1989), which main goal is o achie e a inal synch oniza ion in he adap i e
p ocess ha p omo es he mos use ul aspec s o la e p ocessing. Fo ha , he colo RGB
image is decomposed in o h ee channels maximally deco ela ed using P incipal Componen
Analysis (PCA). Ne e heless, we a e no in e es ed in educing he colo space dimension,
hus, we use a ans o med space o he o iginal dimension (3 colo componen s). The i s
componen co esponds o opponen s channel B/W and he emaining wo co espond o
opponen s simila o R/G and Y/B. Howe e , he space, unlike he opponen s space CIE-Lab,
is adap ed o he speci ic s a is ic o he incoming image.
In a second s ep, he goal is o elimina e he spa ial edundancy among he pixels in each
colo channel. In his case, we use a il e -based s a egic in equency domain called
Spec al Whi ening (SW). I is consequence o he Wiene -Khinchin heo em: “ o a
s ochas ic p ocess, he a e age powe spec um is he Fou ie T ans o m o he au oco ela ion
unc ion”. Thus, a whi ened image should ha e a la powe spec um. This can be easily
achie ed using an adap i e il e in he equency domain ha no malizes he spec um o
he ans o med Fou ie co esponding o he incoming image I(x,y) in he ollowing way:
()
()
()
(
)
()
,
,
,,,
xy
xy
xy
Ixy
nIxy
ωω
ωω ωω
⎡⎤
ℑ⎣⎦
==
⎡⎤
ℑ⎣⎦
(1.1)
whe e
[
]
ℑ⋅ is he ans o med Fou ie and 22
sx
y
ω
ωω
=+ is he spa ial equency.
Physically, SW is a edis ibu ion o he spec um ene gy ha will achie e an enhancemen
www.in echopen.com
Objec Recogni ion
188
Fig. 2. Gene al diag am showing how he da a lows h ough he model.
www.in echopen.com

Scene Recogni ion h ough Visual A en ion
and Image Fea u es: A Compa ison be ween SIFT and SURF App oaches
189
o he less edundan pa e ns. This whi ening me hod was p e iously used in he saliency
model p oposed by Guo e al. (Guo e al., 2008) which is based on global phase analysis.
2.2 Analysis s age
In his s age, i is analyzed he maximum alignmen o he local phase o each pixel in each
whi ened colo channel weigh ed by he s eng h o he isual ea u es in he analyzed scale.
Classical me hodology o es ima e he ampli ude and phase in a 1D signal is he Analy ic
Signal. Howe e , he 2D e sion was achie ed pa ially using a quad a u e phase bank il e
(Gabo like il e s), un il he in oduc ion o he Monogenic Signal by Felsbe g & Somme
(Felsbe g & Somme , 2001). Ou model uses his las me hodology ha achie es a new 2D
analy ic signal om he Riesz’s ans o m, which is he 2D iso opic ex ension o Hilbe ’s
ans o m. I s ep esen a ion in Fou ie ’s domain is a se o wo simple il e s in phase-
quad a u e ha a e no selec i e nei he in scale no o ien a ion:
()()
()
12 22 22
,, , , y
x
xy xy
xy xy
HH i i
ω
ω
ωω ωω ωω ωω
⎛⎞
⎜⎟
=⎜⎟
⎜⎟
++
⎝⎠
(1.2)
The Monogenic Signal is a ec o unc ion o h ee componen s o med by he o iginal
signal and wo componen s achie ed by con ol ing i wi h he il e s o Riez’s ans o m,
ha is:
(
)
(
)
(
)
12
, (,),(,) , ,(,) ,
M
xy xy xy h xy xy h xy
⎡
⎤
=∗∗
⎣
⎦

(1.3)
whe e h1(x,y) and h2(x,y) a e he ep esen a ions in he spa ial domain o H1(ωx,ωy) and
H2(ωx,ωy) espec i ely. Because il e s H1 and H2 a e o ien ed in equency domain bu a e
no selec i e in scale, commonly a Gaussian like band-pass il e is used o build scaled
e sions o Riez’s il e s. In ou case, we used he ollowing log-Gauss il e :
()
2
2
log
2log
,
o
o
k
sxy
Ge
ωω
ω
ωω
⎛⎞
⎛⎞
⎜⎟
⎜⎟
⎝⎠
⎜⎟
−⎜⎟
⎛⎞
⎛⎞
⎜⎟
⎜⎟
⎜⎟
⎜⎟
⎝⎠
⎝⎠
⎝⎠
= (1.4)
whe e ω=(ωx,ωy) is he spa ial equency, ωo =( ωox, ωoy) is he cen al equency o he il e
and k is he pa ame e ha go e ns he bandwid h o he il e . I gs(x,y) is he spa ial
ep esen a ion o p e ious il e , he monogenic space o scales is buil as ollows:
(
)
(
)
(
)
(
)
(
)
(
)
,12
1, 2,
, (,) , ,(,) , , ,(,) , ,
(,), (,), (,)
Ms s s s
sss
xy xy g xy xy g xy h xy xy g xy h xy
xyh xyh xy
⎡⎤
=∗ ∗∗ ∗∗ =
⎣⎦
⎡⎤
=⎣⎦

(1.5)
The chosen bank o il e s is o med by h ee scales (s=3) which cen al wa eleng hs we e
dis ibu ed in 1 oc a e om he minimum wa eleng h (assigned o λ1=8 pixels), ha is λi={8,
16, 32} pixels. The k pa ame e was ixed o achie e a bandwid h o 2 oc a es in each il e in
o de o ob ain a good spec al co e age in he bank o il e s. A simple implemen a ion o
he monogenic signal, in he equency domain, can be ound in (Ko esi, 2000).
www.in echopen.com
Objec Recogni ion
190
Once i is achie ed he monogenic decomposi ion, he impo ance o each isual ea u e is
measu ed by maximizing in each pixel o he image and o all he scales, he le el o local
phase alignmen o he Fou ie Ha monics, weigh ed by he s eng h o he isual s uc u e
in each scale (measu ed as local ene gy
(
)
,,
Mi
xy). We call his measu e Weigh ed
Maximum Phase Alignmen (WMAP), and is he ollowing:
(
)
(
)
{
}
() ()
()
()
dn 1 ,
dn 1 ,
1, 2,
11 1
dn 1 1, 2,
1
,wmax ,cos
,
wmax , ,
(,), (,), (,)
=w max (,), (,), (,)
(,)
s
iMi i
local
s
iMi
local
ss s
iii
ii i
s
ii i i
s
i
i
WMPA x y x y
Exy
xy
Exy
xy h xy h xy
xyh xyh xy
xy
θ
=
=
== =
=
=
=⋅ ⋅ =
⎧⎫
⎛⎞
⎪⎪
⎜⎟
=⋅ =
⎨⎬
⎜⎟
⎪⎪
⎝⎠
⎩⎭
⎛⎞
⎜⎟
⎝⎠
⋅⋅
⎛
⎜
⎝
∑∑ ∑
∑



i
1,
222
11
(,) (,)
i
ss
ii
hxy hxy
==
⎧ ⎫
⎛⎞
⎪ ⎪
⎜⎟
⎪ ⎪
⎜⎟
⎪ ⎪
⎜⎟
⎨ ⎬
⎜⎟
⎛⎞
⎪ ⎪
⎜⎟
⎞⎛ ⎞⎛ ⎞
⎜⎟
⎪ ⎪
++
⎜⎟
⎟⎜ ⎟⎜ ⎟
⎜⎟
⎪ ⎪
⎜⎟
⎜⎟
⎠⎝ ⎠⎝ ⎠
⎝⎠
⎝⎠
⎩ ⎭
∑∑
(1.6)
whe e
()
,,
Mi
xy

is he monogenic signal o he i- h scale and θi is he angle be ween ec o s
()
,,
Mi
xy

and local
E

. This angle measu es he de ia ion o he local phase in he monogenic
signal a he i- h scale espec o he local ene gy ec o in pixel (x,y).
We a e only in e es ed on hose pixels whe e local phase is cong uen o he mos o he
used scales. Thus, ou measu e mus inco po a e a ac o ha penalizes oo na ow
equency dis ibu ions. Fac o w dn is achie ed as i was p oposed by Ko esi (Ko esi, 1999)
o his measu e o local Phase Cong uency (PC).
2.3 Ou pu s age
The inal s age o he model has he aim o achie ing a Re inop ic measu e o he salience o
each pixel in he image. Fo ha , we in eg a e in each pixel he WMPA(x,y) measu es o
each colo channel:
()
3
1
,(,)
c
Salienc
y
x
y
WMAP x
y
=
=∑ (1.7)
Finally, a smoo hing is in oduced by a Gaussian il e and also a no maliza ion in o de o
make easy o in e p e he saliency map as a p obabili y unc ion o ecei e a en ion.
2.4 Compu a ional complexi y
The compu a ional e iciency o he model is low due o he load in oduced by he PCA
analysis, which g ows lineally wi h he numbe o pixels in he image (N) and cubically wi h
he numbe o componen s (colo channels), O(M3+N M2). The numbe o componen s is low
and cons an , M=3, hus, he asymp o ic complexi y depends on N. The compu a ional
complexi y o he model depends on he FFT (Fas Fou ie T ans o m) complexi y pe o med
in il e ing p ocessing. This complexi y is O(N log(N)). On he o he hand, he compu a ional
iming o he model is low, by example, o an image o 512x384 pixels using an In el Co e2
Quad p ocesso a 2.4 GHz and 4Gb o RAM memo y, he algo i hm akes 0.91 seconds. We
ha e o ake in o accoun ha he algo i hm is scien i ic so wa e p og ammed in MATLAB.
www.in echopen.com
Scene Recogni ion h ough Visual A en ion
and Image Fea u es: A Compa ison be ween SIFT and SURF App oaches
191
3. Image ea u es
SIFT and SURF belong o a se o me hods aimed o de ec and desc ibe local ea u es in
images. Among hese me hods we can ound (Mikolajczyk, 2005): shape con ex , s ee able
il e s, PCA-SIFT, di e en ial in a ian s, spin images, complex il e s, momen in a ian s
and g adien loca ion and o ien a ion his og ams (GLOH). Ne e heless, SIFT and SURF
ha e cap u ed ecen a en ion o esea che s wo king on applica ions like objec
ecogni ion, obo mapping and na iga ion, image s i ching, 3D modeling, ideo acking,
e c, being i s compa ison a cu en issue in li e a u e (Baue , 2007).
Wi h ega ds o SIFT ea u es, we used he Lowe´s algo i hm (Lowe, 2004) which wo ks as
ollows. To iden i y he in e es poin s (keypoin s), scale space ex ema a e ound in a
di e ence-o -Gaussian (DoG) unc ion con ol ed wi h he image. The ex emas a e ound
by compa ing each poin wi h i s neighbo s in he cu en image and adjacen scales. Poin s
a e selec ed as candida e keypoin loca ions i hey a e he maximum o minimum alue in
hei neighbo hood. Then image g adien s and o ien a ions, a each pixel o he Gaussian
con ol ed image a each scale, a e compu ed. Fo each key loca ion an o ien a ion,
de e mined by he peak o a his og am o p e iously compu ed neighbo hood o ien a ions,
is assigned. Once he o ien a ion, scale, and loca ion o he keypoin s ha e been compu ed,
in a iance o hese alues is achie ed by compu ing he keypoin local ea u e desc ip o s
ela i e o hem. Local ea u e desc ip o s a e 128-dimensional ec o s ob ained om he
p e-compu ed image o ien a ions and g adien s a ound he keypoin s.
SURF ea u es (Bay, 2008) a e based on sums o 2D Haa wa ele esponses and make a e y
e icien use o in eg al images o speed-up he p ocess. As basic image desc ip o s hey use
a Haa wa ele app oxima ion o he de e minan o Hessian blob de ec o . The e a e wo
e sions: he s anda d e sion which uses a desc ip o ec o o 64 componen s (SURF-64),
and he ex ended e sion which uses 128 componen s (SURF-128). SURF a e obus image
ea u es pa ly inspi ed by SIFT, being he s anda d e sion o SURF se e al imes as e
han SIFT. SURF ea u es p o ide signi ican ly less keypoin s han SIFT, app oxima ely he
hal o hem (see Figu e 3).
Fig. 3. SIFT (le ) and SURF ( igh ) keypoin s compu ed o he same ame.
www.in echopen.com
Objec Recogni ion
192
4. Scene ecogni ion applica ion
Scene ecogni ion is ela ed wi h he ecogni ion o gene al scena ios a he han local
objec s. This app oach is use ul in many applica ions such as mobile obo na iga ion, image
e ie al, ex ac ion o con ex ual in o ma ion o objec ecogni ion, and e en o p o ide
access o ou is in o ma ion using came a phones. In ou case, we a e in e es ed in
ecognize a se o di e en scena ios which a e pa o uni e si y acili ies o med by ou
class ooms and h ee halls. The inal aim is o p o ide gene al loca ion da a use ul o he
na iga ion o a mobile obo sys em. Scene ecogni ion is commonly pe o med using
gene ic image ea u es ha y o collec enough in o ma ion o be able o dis inguish among
he di e en scena ios. Fo his pu pose we used SIFT and SURF al e na i es.
To compu e he SIFT ea u es we used he o iginal code by Lowe (h p://people.cs.ubc.ca/
lowe/keypoin s/). We also used he o iginal code o SURF ea u es by Bay e al
(h p://www. ision.ee.e hz.ch/~su /). To ca y ou he classi ica ion ask we used he 1-
NN ule, which is a simple classi ica ion app oach bu as o compu e and obus . Fo he 1-
NN app oach, we need o build p e iously a da abase o p o o ypes ha will collec he
ecogni ion knowledge o he classi ie . These p o o ypes a e a se o labelled SIFT/SURF
keypoin s ob ained om he aining ames. The class o he keypoin s compu ed o a
speci ic aining ame will be ha p e iously assigned o his ame in an o -line
supe ised labeling p ocess. The da abase is hen inco po a ed in o he 1-NN classi ie ,
which uses he Euclidean dis ance o selec he closes p o o ype o he es SIFT/SURF
keypoin being classi ied. The class o e e y es keypoin will be assigned o he class o he
closes p o o ype in he da abase, and inally, he class o he en i e es ame will be ha o
he majo i y o i s keypoin s.
5. Expe imen s and esul s
The expe imen al wo k consis ed in a se o expe imen s ca ied ou using ou ideo
sequences aken in a obo -na iga ion manne . These ideo sequences we e g abbed in an
uni e si y a ea co e ing se e al ooms and halls. Sequences we e aken a 5 ps collec ing a
o al numbe o 2,174 ames (7:15 minu es) o he i s sequence, 1,986 ames o he
second (6:37 minu es), 1,816 ames o he hi d (6:03 minu es) and 1,753 ames o he
ou h (5:50 minu es). Fi s and hi d sequences we e aken in a speci ic o de o halls and
ooms: hall-1, oom-1, hall-1, oom-2, hall-1, oom-3, hall-1, hall-2, hall-3, oom-4, hall-3,
hall-2, hall-1. The second and ou h sequences we e g abbed ollowing he opposi e o de
o collec all possible iewpoin s o he obo na iga ion h ough he acili ies. In all he
expe imen s, we used he i s and second sequences o aining and he hi d and ou h
o es ing.
In he i s expe imen we compu ed he SIFT keypoin s o all he ames o he aining
ideo sequences. Then, we labelled hese keypoin s wi h he co esponding ame class:
oom-1, oom-2, oom-3, oom-4, hall-1, hall-2 o hall-3. The whole se o labelled keypoin s
o med i sel he da abase o p o o ypes o be used by he 1-NN classi ie . Fo each ame o
he es ing sequences hei co esponding SIFT keypoin s we e compu ed and classi ied. The
inal class o he ame was se o he majo i y class among i s keypoin s. Ve y good
pe o mance was achie ed, 95.25% o co ec classi ica ion o ames. Howe e , an
impo an d awback was he compu a ional cos o classi ica ion, which was high despi e
he ac ha 1-NN is known as a low cos classi ie . This was due o he e y la ge size o he
www.in echopen.com
Objec Recogni ion
Edi ed by D . Tam Phuong Cao
ISBN 978-953-307-222-7
Ha d co e , 350 pages
Publishe InTech
Published online 01, Ap il, 2011
Published in p in edi ion Ap il, 2011
InTech Eu ope
Uni e si y Campus STeP Ri
Sla ka K au zeka 83/A
51000 Rijeka, C oa ia
Phone: +385 (51) 770 447
Fax: +385 (51) 686 166
www.in echopen.com
InTech China
Uni 405, O ice Block, Ho el Equa o ial Shanghai
No.65, Yan An Road (Wes ), Shanghai, 200040, China
Phone: +86-21-62489820
Fax: +86-21-62489821
Vision-based objec ecogni ion asks a e e y amilia in ou e e yday ac i i ies, such as d i ing ou ca in he
co ec lane. We do hese asks e o lessly in eal- ime. In he las decades, wi h he ad ancemen o
compu e echnology, esea che s and applica ion de elope s a e ying o mimic he human’s capabili y o
isually ecognising. Such capabili y will allow machine o ee human om bo ing o dange ous jobs.
How o e e ence
In o de o co ec ly e e ence his schola ly wo k, eel ee o copy and pas e he ollowing:
Fe nando Lopez-Ga cia, Xose Ramon Fdez-Vidal, Xose Manuel Pa do and Raquel Dosil (2011). Scene
Recogni ion h ough Visual A en ion and Image Fea u es: A Compa ison be ween SIFT and SURF
App oaches, Objec Recogni ion, D . Tam Phuong Cao (Ed.), ISBN: 978-953-307-222-7, InTech, A ailable
om: h p://www.in echopen.com/books/objec - ecogni ion/scene- ecogni ion- h ough- isual-a en ion-and-
image- ea u es-a-compa ison-be ween-si -and-su -app

© 2011 The Au ho (s). Licensee In echOpen. This chap e is dis ibu ed
unde he e ms o he C ea i e Commons A ibu ion-NonComme cial-
Sha eAlike-3.0 License, which pe mi s use, dis ibu ion and ep oduc ion o
non-comme cial pu poses, p o ided he o iginal is p ope ly ci ed and
de i a i e wo ks building on his con en a e dis ibu ed unde he same
license.