COMPUTATIONALLY EFFICIENT SERIAL COMBINATION OF

ROTATION-INVARIANT AND ROTATION COMPENSATING IRIS

RECOGNITION ALGORITHMS

Mario Konrad, Herbert St

ogner

School of Communication Engineering for IT, Carinthia Tech Institute, Austria

Andreas Uhl, Peter Wild

Department of Computer Sciences, University of Salzburg, Austria

Keywords:

Iris recognition, Multibiometrics, Serial classiﬁer combination, Rotation compensation.

Abstract:

Rotation compensation is one of the computational bottlenecks in large scale iris-based identiﬁcation schemes,

since a signiﬁcant amount of Hamming distance computations is required in a single match due to the neces-

sary shifting of the iris codes to compensate for eye tilt. To cope with this problem, a serial classiﬁer com-

bination approach is proposed for iris-based identiﬁcation, combining rotation-invariant pre-selection with a

traditional rotation compensating iris code-based scheme. The primary aim, a reduction of computational com-

plexity, can easily be met - at comparable recognition accuracy, the computational effort required is reduced

to 20% or even less of the fully ﬂedged iris code based scheme. As a by-product, the recognition accuracy is

shown to be additionally improved in open-set scenarios.

1 INTRODUCTION

Iris recognition technology has been dominated over

years by the commercially successful algorithm of

J. Daugman (Daugman, 2004). This algorithm ba-

sically extracts local iris features from polar iris im-

ages by convolution with 2-dimensional complex Ga-

bor atoms, quantizing the resulting phase information

into 2 bits per coefﬁcient. The basic idea of extract-

ing local intensity variations from iris texture has been

followed employing other types of transforms and

methods as well, e.g. in the spatial domain or in the

wavelet domain. All these approaches share the prop-

erty of being sensitive against eye tilt, i.e. they are

intrinsically not rotation invariant due to the usage of

local spatial information. Therefore, in order to com-

pensate potential rotation, in all these algorithms the

templates in the matching process are shifted against

each other for a certain amount, and taking the mini-

mal template distance among all shifted versions as

the actual distance. Obviously, depending on the

amount of shift that is required for a certain applica-

tion (i.e. the amount of rotation that is to be expected),

these operations may amount to a signiﬁcant number

of matching operations performed, which can become

prohibitive in an identiﬁcation scenario.

Rotation-invariant iris features therefore represent

an attractive alternative. Due to the signiﬁcant com-

putational demand associated with transform domain

processing, spatial domain techniques working di-

rectly on the iris texture are of speciﬁc interest in our

context. Du et al. (Du et al., 2006) employ ﬁrst order

moments of the iris texture line-histograms. While

this technique is successful in providing rotation in-

variance and consequently fast matching procedures

independent of the eye’s position, it fails in terms of

recognition accuracy.

This is where our approach comes in. In this work

we combine a spatially-based rotation invariant iris

recognition approach with a traditional local-feature

based scheme into a serial classiﬁer combination.

The aim is to result in reduced overall computa-

tional demand as compared to classical rotation com-

pensating schemes while at least maintaining their

recognition accuracy. This is achieved by using the

ﬁrst scheme to determine a certain amount of the

highest matching ranks of the entire database (this can

be done quickly due to the high speed of the ﬁrst

Konrad M., Stögner H., Uhl A. and Wild P. (2010).

COMPUTATIONALLY EFFICIENT SERIAL COMBINATION OF ROTATION-INVARIANT AND ROTATION COMPENSATING IRIS RECOGNITION

ALGORITHMS.

In Proceedings of the International Conference on Computer Vision Theory and Applications, pages 85-90

DOI: 10.5220/0002821100850090

 SciTePress

scheme), while the second (and more accurate)

scheme is then only applied to this predetermined

subset to determine the ﬁnal matching result. Serial

classiﬁer combination aiming at a reduction of iden-

tiﬁcation time is a relatively new research topic in

biometrics, see (Uhl and Wild, 2009; Gentile et al.,

2009) for applications in hand biometrics and iris

biometrics. This work is the ﬁrst multi-algorithm

multibiometric (Ross et al., 2006) approach combin-

ing rotation-invariant with rotation compensating iris

recognition algorithms serially to reduce computa-

tional demands. Besides the aimed reduction in com-

putational effort it is of particular interest, if the com-

bination of two very different feature types can also

lead to improved recognition results. Since features

used in classiﬁer combinations need to be as uncor-

related as possible to result in better results as com-

pared to their single classiﬁer counterparts, a recog-

nition improvement might be expected as well in our

case, since the features of the two classiﬁers used are

of signiﬁcantly different nature.

In iris recognition, several single-sensor multibio-

metric approaches have been suggested yet, however,

mostly focussing on an improvement of recognition

accuracy. Sun et al. (Sun et al., 2005) “cascade” two

feature types employing global features only in ad-

dition to a Daugman-like approach if the matching

value of the latter is in a questionable range. Also

Zhang et al. (Zhang et al., 2004) use a similar strat-

egy while interchanging the role of the global and the

local features. Vatsa et al. (Vatsa et al., 2005) com-

pute the Euler number of connected components as

global feature, while again using an iris code as local

texture feature. Park and Lee (Park and Lee, 2006)

decompose the iris data with a directional ﬁlterbank

and extract two different feature types from this do-

main. Combining the results leads to an improvement

of the single techniques applied.

Approaches closer related to our work are quick

screening or pre-classiﬁcation techniques. Several au-

thors have developed techniques to divide iris data

into a certain number of categories in order to achieve

a rough pre-classiﬁcation before applying a more ac-

curate matching technique. Qui et al. (Qui et al.,

2007) use iris textons to generate ﬁve classes, and

Yu et al. (Yu et al., 2006) use fractal dimension to

generate four classes. The approach as described in

this work is different from the techniques proposed so

far in several ways. First, it is different compared to

the single-sensor multibiometric approaches combin-

ing global and local features suggested since we

• mainly focus on a reduction of computational ef-

fort by limiting the required rotation compensa-

tion in matching instead of aiming at better recog-

nition accuracy and

• we apply a speciﬁc type of serial classiﬁer combi-

nation to accomplish this (i.e. contrasting to many

other approaches not both classiﬁers are applied to

the entire dataset).

Second, it is different to the developed screening

approaches since we

• do not partition the templates into a certain num-

ber of classes and do therefore not limit the actual

matching to the common class of database tem-

plate and sample template and

• we actually combine our “screening approach”

with a classical technique in serial manner and

provide actual computational performance and

classiﬁcation accuracy results.

In contrast to recently introduced serial iris fea-

ture combinations (Gentile et al., 2009), we employ a

rotational-invariant pre-selecting feature making bit-

shifts in the ﬁrst classiﬁer unnecessary.

Section 2 describes the iris recognition tech-

niques, which are part of our serial matching approach

and explains the applied serial classiﬁer combination

technique. In Section 3, the experimental setup is dis-

cussed and experimental results are provided. First,

we illustrate the results of the classical approach us-

ing local features and rotation compensation. Sub-

sequently, we discuss and present the experimental

results for the serial classiﬁer combination for the

open-set scenario with respect to time consumption

as well as recognition performance of the different ap-

proaches. Section 4 concludes the paper.

2 SERIAL CLASSIFIER

COMBINATION IN IRIS

RECOGNITION

The applied serial classiﬁer combination mainly aims

at reducing computation demand in identiﬁcation sce-

narios. The basic idea is to employ a faster recog-

nition scheme with rotation invariant (more global)

features for a ﬁrst screening sweep across the entire

database. The screening procedure does not result in

a certain number of classes, which can then be used to

restrict the subsequent search to a single class, but re-

sults in a ranking of the enrolled database templates.

This ranking is subsequently used to determine actu-

ally two classes, where the subsequent search is lim-

ited to one of these classes then – obviously this is

the class with the highest ranks. An important pa-

rameter of this approach is the amount of top ranked

VISAPP 2010 - International Conference on Computer Vision Theory and Applications

templates that is contained in the class subject to fur-

ther search. We denote this parameter p and it is ex-

pressed in percent of the database, e.g. if p = 25, the

25% top ranked templates are subject to further pro-

cessing by the second, computationally more expen-

sive approach (which relies on local texture features

and requires rotation compensation). Obviously, for

increasing p the computational demand is increased.

The main question addressed in the experimental sec-

tion is whether the recognition accuracy can be main-

tained for decreasing p, or otherwise, in how far the

decrease of p is coupled to a decrease of recognition

accuracy.

In the next two subsections we will brieﬂy review

the two iris recognition schemes, which we use to

combine into a serial classiﬁer combination.

2.1 Rotation-invariant Screening

Du et al. (Du et al., 2006) have proposed a rotation

invariant 1D signature approach, which we employ

as the ﬁrst screening stage. As preprocessing, lo-

cal texture patterns (LTPs) are generated, which sub-

stract a localized mean value from the data. By av-

eraging LTP values of an entire row of the LTP polar

iris image (which is generated in a ﬁrst preprocessing

stage after iris detection), one value of the 1D sig-

nature is generated. While the upper and lower three

rows of the polar iris data are discarded, the remaining

rows are used to create the ﬁnal 1D signature. In re-

cent work (Matschitsch et al., 2008) this approach has

been extended by using entire row histograms instead

of ﬁrst order moments only, which results in 2D sig-

natures in some sense. Clearly, both the original 1D

and the extended 2D signatures are rotation invariant.

Two 1D signatures are compared by using the Du

measure (Du et al., 2006). For the extension to 2D

signatures, we apply the “accumulated errors” ap-

proach (Matschitsch et al., 2008), where the Du mea-

sure is computed for each row between the single his-

tograms and the resulting distances are accumulated

afterwards for all rows. We apply the accumulated er-

rors strategy using 256 histogram bins. Moreover, a

weighting factor > 1 is used for the distance of polar

image rows, which are close to the pupil and a fac-

tor = 1 is proposed for rows, which are close to the

sclera of the eye. The weighting factors for the rows

in between are obtained by linear interpolation.

2.2 Local Texture Features – Iris Code

For the original iris code (Daugman, 2004), the po-

lar images are subject to a 2D complex Gabor ﬁlter-

ing process, subsequently the available phase infor-

mation is quantized into four different levels, one for

each of the four possible quadrants in the complex

plane. Hence, for each pixel of the polar image, two

bits are obtained, which are combined and form the

iris template (i.e. iris code), which can be compared

to other iris codes by computing the Hamming dis-

tance. This measure is highly localized and needs

to compensate for possible rotation between irises –

this is done by applying the Hamming distance cal-

culation several times while shifting the polar images

against each other. The lowest matching value then

determines the ﬁnal distance.

In this work, we use an open-source MatLAB im-

plementation, which applies a 1D Gabor-ﬁlter version

of the iris code strategy for iris recognition. Due to its

free availability

and the lack of other publicly avail-

able iris recognition software, it has gained signiﬁcant

popularity in the community.

3 EXPERIMENTS

3.1 Experimental Setup

We use the CASIA V1.0 and the CASIA V3.0 Inter-

val

as well as the MMU V1.0

datasets where for

each database two different subsets are selected. For

the CASIA V1.0, which consists of 756 images ac-

quired from 108 eyes (7 images per eye), the ﬁrst

subset contains 630 iris images and the second subset

comprises 126 images. For the CASIA V3.0, the ﬁrst

subset consists of 1705 iris images acquired from 341

different eyes (5 images for each eye). The second

subset includes 117 images from 53 eyes with vari-

ous numbers of images per eye. For MMU, the ﬁrst

dataset is composed of 400 iris images from 80 eyes,

while the second set contains 50 images (again 5 im-

ages per eye). For all three databases, the ﬁrst dataset

is used to serve as database of enrolled persons and

the second dataset contains images that are unknown

to the recognition system (which are needed in the

open-set scenario). Out of all datasets, we extract po-

lar iris images with 360 × 65 pixels, which results in

1D/2D signatures with length 59 since only 59 out of

65 LTP rows are used.

We consider the open-set scenario (or watchlist

scenario). Here it is not guaranteed that the per-

son that should be identiﬁed is truly member of the

database. Hence, an identiﬁcation attempt results in

http://www.csse.uwa.edu.au/˜pk/

studentprojects/libor/sourcecode.html

http://www.sinobiometrics.com

http://pesona.mmu.edu.my/ccteo/

COMPUTATIONALLY EFFICIENT SERIAL COMBINATION OF ROTATION-INVARIANT AND ROTATION

COMPENSATING IRIS RECOGNITION ALGORITHMS

a correct identiﬁcation whenever an enrolled person

is correctly recognized. If a not enrolled person is

falsely labelled as database member, the attempt will

result in a false accept. The rank-1 recognition rate as

well as the false accept rate is used in order to assess

this type of system.

With respect to serial classiﬁer combination,

we investigate set reductions to top-ranked p =

1, 5, 10, 15 candidates subjected to further processing.

Rotation compensation for the “pure” iris code tech-

nique is conducted with 2, 4, and 8 shifts of the iris

code in each direction (which sums up to 17 Ham-

ming distance calculations in the case of 8 shifts), in

the context of the serial combination rotation compen-

sation is conducted using 8 shifts in the second stage

of the identiﬁcation.

3.2 Results

First, we create reference results by analyzing the per-

formance of the “pure” iris code approach. Table 1

illustrates the time consumption to search the entire

database, which obviously strongly depends on the

amount of shifts (AMD Athlon 2200+ processor, 512

MB RAM, Windows XP and MatLAB R2007b). For

comparison, the required amount of time to search

the database with the Du approach is also displayed.

When conducting only a single Hamming distance

computation (instead of already 5 in the case of two

shifts), the iris code algorithm would be only slightly

slower compared to the Du approach. Only rotation

compensation degrades its performance signiﬁcantly.

Note, that the image segmentation and generation of

the polar iris image was performed prior to the analy-

sis and is excluded from the timing results. The time

effort for these computations is 52 seconds for the

CASIA V1.0 dataset and 59 seconds for the CASIA

V3.0 dataset, for example.

In Figure 1, we show the effect of applying a vari-

able amount of shifts on recognition accuracy in the

iris code approach for CASIA V3.0 as an example.

In accordance to literature, 8 shifts yield only slightly

better results than 4 shifts overall, but this effect is

more pronounced for the important case of lower false

accept rates, which makes the application of 8 shifts

sensible in this case as well. Therefore, further results

refer to 8 shifts in the “pure” iris code approach as

Table 1: Time consumption of iris code-based identiﬁcation

with a various number of shifts and Du’s approach.

Du Iris code 2s Iris code 4s Iris code 8s

Cv1 12.5 s 77.2 s 105.8 s 176.3 s

Cv3 32.1 s 215.7 s 322.7 s 507.1 s

MMU 8.1s 48.3 s 69.3 s 111.5 s

Figure 1: Performance in terms of recognition accuracy of

the iris code approach for CASIA V3.0 with varying amount

of shifts.

Figure 2: Time consumption (in seconds) of the iris code

approach and the serial combination approach with p =

1, 5, 10,15 for 1D signatures.

well as in the second stage of the serial combination.

Figure 2 displays the time needed to perform iden-

tiﬁcation over the entire datasets and compares the

iris code approach using 8 shifts (denoted Masek’s-

daugman in this plot) with the serial approach using

1D signatures with p = 1, 5, 10, 15. The time reduc-

tion achieved even for p = 15 is impressive.

Table 2 summarizes the time consumption of the

different serial combination variants for 1D (S1DpX)

and 2D signatures (S2DpX). As expected, it is con-

ﬁrmed that the serial approach is much faster than the

classical iris code method (IC8s). Although the ap-

plication of 2D signatures raises the computational

effort, the results are also clearly superior. Apply-

ing a weighting factor (w=4) to the signatures does

only slightly increase the computational demand in

a negligible manner (not shown). For example, for

1D signatures with p = 10 (which is better in term of

recognition accuracy in many cases compared to the

iris code case as we shall see), the serial approach is

faster more than a factor of 6, and still for 2D signa-

tures the serial approach is faster by a factor of 5.

Having shown that the aim of signiﬁcant reduction

of the computational effort has been met, we investi-

gate the impact of the serial classiﬁer combination on

VISAPP 2010 - International Conference on Computer Vision Theory and Applications

Table 2: Time consumption (in seconds) of the iris code

approach and the serial combination approach with p =

1, 5, 10, 15 for 1D (above) and 2D signatures (below).

1-D signatures

IC8s S1Dp1 S1Dp5 S1Dp10 S1Dp15

Cv1 181.5 s 14.1 s 21.3 s 30.3 s 39.3 s

Cv3 507.1 s 36.9 s 56.6 s 79.7 s 104.7 s

MMU 111.5 s 9.5 s 14.1 s 19.8 s 25.5 s

2-D signatures

IC8s S2Dp1 S2Dp5 S2Dp10 S2Dp15

Cv1 181.5 s 19.7 s 27.0 s 35.0 s 45.0 s

Cv3 507.1 s 52.5 s 71.7 s 95.7 s 120.0 s

MMU 111.5 s 13.4 s 18.1 s 23.9 s 29.8 s

Table 3: Rank-1 Recognition Rate (RR-1) at certain False

Accept Rate (FAR) values.

Casia V1.0

FAR IC8s S1Dp1% S1Dp5% S1Dp10% S1Dp15%

0% 90.9% 81.6% 88.1% 90.3% 91.9%

0.79% 91.1% 81.6% 88.6% 90.3% 91.9%

11.1% 91.1% 81.9% 89.2% 91.1% 92.5%

CASIA V3.0

FAR IC8s S1Dp1% S1Dp5% S1Dp10% S1Dp15%

0% 68.3% 69.2% 76.1% 65.5% 66.2%

0.85% 86.7% 77.9% 85.8% 87.7% 88.4%

11.1% 87.4% 78.4% 86.3% 88.3% 89.0%

MMU V1.0

FAR IC8s S1Dp1% S1Dp5% S1Dp10% S1Dp15%

0% 84.5% 49.8% 68.8% 73.8% 76.8%

0.85% 85.8% 52.5% 71.3% 75.8% 78.8%

11.1% 86.0% 53.3% 71.8% 76.3% 79.3%

the recognition results in the following. Table 3 shows

the results of the serial approach with 1D signatures

and using p = 1, 5, 10, 15 compared to the iris code

8-shifts variant for the three databases.

We notice, that the CASIA V1.0 dataset yields

overall better results than the CASIA V3.0 and MMU

V1.0 datasets. This fact may be due to the fact that

the pupil areas of the CASIA V1.0 dataset have been

post-processed and set to a uniformly dark gray value,

which eases the segmentation process and therefore

reduces errors coming from incorrect segmentation.

For the CASIA V1.0 dataset, only the p = 15 se-

rial variant (S1Dp15) yields better identiﬁcation rates

than the iris code approach (IC8s). For higher false

accept rates (FAR=11%), the p = 10 variant leads to

similar results as compared to the iris code reference

case.

For the CASIA V3.0 dataset and zero false ac-

cepts, the p = 5 serial variant yields the best identi-

ﬁcation rate (76.1%) and also p = 1 is surprisingly

better than the iris code approach. The decreasing

identiﬁcation accuracy for increasing values of p may

be surprising at ﬁrst, however, these results are due

to the fact that a larger set of preselected templates

can increase the chance for false positives, which have

(a) CASIA V1.0

(b) CASIA V3.0

Figure 3: Performance of the 2D serial approach compared

to the iris code approach for the CASIA datasets.

been excluded for lower values of p. For higher false

accept rates (FAR), the p = 15 percent variant pro-

duces the best results (89% @ 0.85% FAR) while also

p = 10 performs better as compared to the iris code

reference and still p = 5 delivers comparable recog-

nition performance.

For both datasets these recognition results are not

obvious since a coarse pre-selection could be ex-

pected to exclude potential genuine templates too. In

contrary, the pre-selection helps to avoid false positive

matches. This may be explained by the fact that the

Du approach yields good results for partial iris recog-

nition (e.g., heavy eyelid occlusion or segmentation

errors) and noisy images (Du et al., 2006).

In contrast, the results for the MMU dataset are

worse for the serial combinations as compared to the

classical iris code approach for all values of p =

1, 5, 10, 15. We also notice that recognition accuracy

increases monotonically with increasing p so that we

can expect at least equal performance for larger values

of p.

In Figure 3 we further analyze the possible im-

provement of the serial approach for p = 10 by us-

ing 2D signatures and weighting of inner rows with

weight 4. For both datasets shown (CASIA V1.0 and

COMPUTATIONALLY EFFICIENT SERIAL COMBINATION OF ROTATION-INVARIANT AND ROTATION

COMPENSATING IRIS RECOGNITION ALGORITHMS

Figure 4: Performance of the 2D serial approach compared

to the iris code approach for the MMU dataset.

V3.0), the 2D variant (hist-256b-w1) improves per-

formance compared to 1D signatures considerably,

while weighting (hist-256b-w4) only results in minor

improvement. The performance of the serial approach

with 2D signatures is better as the iris code technique

for all cases for both CASIA datasets.

Figure 4 shows that also for the MMU dataset

improvements in terms of recognition performance

as compared to the iris-code approach are possible

if 2D signatures are used and p is set sufﬁciently

high. We observe that for p = 15 and 2D signa-

tures with weighting the iris code approach is outper-

formed across the entire range of considered false ac-

cept rates. Note that this improvement still is achieved

at a reduction of computational effort by a factor of

3.5 (compare Table 2).

4 CONCLUSIONS

As expected, we are able to reduce computational de-

mands with our proposed serial classiﬁer combination

considerably. At a comparable recognition accuracy

we sufﬁce with 20% - 30% or even less computa-

tion time for identiﬁcation (the actual value depends

on the speciﬁc dataset considered). Interestingly, we

are even able to outperform the recognition accuracy

of iris code based recognition, since the serial clas-

siﬁer combination technique turns out to be more ro-

bust against false acceptances. This is especially in-

teresting, since we reveal that the rotation-invariant

ﬁrst stage of the serial combination is able to exclude

candidates, which lead to false accepts in the entirely

rotation-compensating iris-code approach.

ACKNOWLEDGEMENTS

This work has been partially supported by the Aus-

trian Research Promotion Agency (FFG) FIT-IT Trust

in IT-Systems project 819382.

REFERENCES

Daugman, J. (2004). How iris recognition works. IEEE

Transactions on Circiuts and Systems for Video Tech-

nology, 14(1):21–30.

Du, Y., Ives, R., Etter, D., and Welch, T. (2006). Use

of one-dimensional iris signatures to rank iris pattern

similarities. Optical Engineering, 45(3):037201–1 –

037201–10.

Gentile, J. E., Ratha, N., and Connell, J. (2009). An efﬁ-

cient, two-stage iris recognition system. In Biomet-

rics: Theory, Applications and Systems (BTAS’09),

pages 1–5.

Matschitsch, S., St

ogner, H., Tschinder, M., and Uhl, A.

(2008). Rotation-invariant iris recognition: boosting

1D spatial-domain signatures to 2D. In Filipe, J.,

Cetto, J., and Ferrier, J.-L., editors, ICINCO 2008:

Proceedings of the 5th International Conference on

Informatics in Control, Automation and Robotics, vol-

ume SPSMC, pages 232–235. INSTICC Press.

Park, C.-H. and Lee, J.-J. (2006). Extracting and com-

bining multimodal iris features. In Proceedings of

the 1st IAPR International Conference on Biometrics

(ICB’06), number 3832 in Lecture Notes on Computer

Science, pages 389–396.

Qui, X., Sun, Z., and Tan, T. (2007). Coarse iris classiﬁ-

cation by learned visual dictionary. In Proceedings of

the 2nd IAPR/IEEE International Conference on Bio-

metrics (ICB’07), number 4642 in Lecture Notes on

Computer Science, pages 770–779.

Ross, A., Nandakumar, K., and Jain, A. (2006). Handbook

of Multibiometrics. Springer.

Sun, Z., Wang, Y., Tan, T., and Cui, J. (2005). Improv-

ing iris recognition accuracy via cascaded classiﬁers.

IEEE Transactions on Systems, Man and Cybernetics,

35(3):435–441.

Uhl, A. and Wild, P. (2009). Parallel versus serial classiﬁer

combination for multibiometric hand-based identiﬁca-

tion. In Tistarelli, M. and Nixon, M., editors, Pro-

ceedings of the 3rd International Conference on Bio-

metrics 2009 (ICB’09), volume 5558 of LNCS, pages

950–959. Springer Verlag.

Vatsa, M., Singh, R., and Noore, A. (2005). Reducing the

false rejection rate of iris recognition using textural

and topological features. Int. Journal of Signal Pro-

cessing, 2(2):66–72.

Yu, L., Wang, K., and Zhang, D. (2006). A novel

method for coarse iris classiﬁcation. In Proceedings of

the 1st IAPR International Conference on Biometrics

(ICB’06), number 3832 in Lecture Notes on Computer

Science, pages 404–410.

Zhang, P.-F., Li, D.-S., and Wang, Q. (2004). A novel iris

recognition method based on feature fusion. In Pro-

ceedings of the International Conference on Machine

Learning and Cybernetics, pages 3661–3665.

VISAPP 2010 - International Conference on Computer Vision Theory and Applications