Occluded Iris Recognition using SURF Features

Anca Ignat

and Ioan Păvăloi

Faculty of Computer Science, University “Alexandru Ioan Cuza” of Iași, Romania

Institute of Computer Science, Romanian Academy Iaşi Branch, Iaşi, Romania

Keywords: Iris Recognition, SURF, Occluded Iris Images, Manhattan Distance.

Abstract: In this paper we study the problem of the recognition process for iris images with missing information. Our

approach uses keypoints related features for solving this problem. We present our recognition results obtained

using SURF (Speeded-Up Robust Features) features extracted from occluded iris images. We tested the

influence on the recognition rate of two threshold parameters, one linked with the SURF extraction process

and the other with the keypoint matching scheme. The proposed method was tested on UPOL iris database

using eleven levels of occlusion. The experiments show that the method we describe in this paper produces

better results than Daugman procedure on all considered datasets and the results we previously obtained using

SIFT features. Comparisons were also performed with iris recognition results that use colour for iris

characterization, computed on the same databases of irises with different levels of missing information.

1 INTRODUCTION

There are many applications that use iris biometric

features for automatic security and access control.

This type of authentication that uses iris information

is a non-invasive technology which provides a highly

reliable solution. Generally, irises have a unique

structure for each human being as those provided by

fingerprints or the network of retinal blood vessels.

For each person, the iris has a unique texture pattern

that allows the process of person recognition. It’s true

that one major limitation is related to the image

acquisition conditions, when images of the iris with

occluded parts, or bad illumination can interfere with

the recognition process. These types of problems

require special approaches in order to have acceptable

iris recognition results. One way to treat this problem

is by using keypoint detectors that help extract the

essential iris information.

Since the development of the famous Iris Code

(Daugman, 1993, 2015), a lot of research was

conducted on the problem of iris recognition. An

excellent review of the methods and research

directions in this field can be found in (Bowyer &

Burge, 2016). De Marsico, Petrosino & Ricciardi,.

(2016) and Harakannanavar & Puranikmath (2017)

provide also excellent reviews on the iris recognition

problem. In the pandemic context, when faces are

covered by mask, the iris recognition problem

becomes of increased interest. Nguyen et al. (2017)

present the research on long range iris recognition.

Rattani & Derakhshani (2017) present a survey of

methods that are analysing not only the iris but the

entire region of the eye. A very interesting approach

using deep learning tools are considered in Nguyen

al. (2017).

Considering the iris recognition problem with

SURF descriptors, in (Ali et al., 2016) the effect of

different enhancement methods as CLAHE, HE

(Histogram Equalization), AHE (Adaptive Histogram

Equalization) and classical matching procedure are

tested. The method is tested on CASIA dataset and

the results are compared with those obtained using

SIFT, HOG, MSER and DAISY descriptors. The best

results are in range 99.5% to 100% and are obtained

with CLAHE. Mehrotra, Sa & Majhi (2013) propose

a new iris segmentation procedure. The SURF

features are extracted after segmentation. Image

matching is performed using the Euclidean distance

and the nearest neighbour ratio procedure.

Experiments are conducted on different iris datasets:

BATH, with the best results of 98.24%, UBIRIS, with

96.58%, and CASIA with 97.32%. The authors state

that their method is robust to scale changes and

rotations, occlusion and illumination changes. In the

experiments performed on CASIA dataset by Bakshi

et al. (2012), both SIFT and SURF features are

extracted, the matching between two images being

performed in three stages. After, the matching scores

are combined.

508

Ignat, A. and P

aloi, I.

Occluded Iris Recognition using SURF Features.

DOI: 10.5220/0010255405080515

In Proceedings of the 16th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2021) - Volume 5: VISAPP, pages

508-515

ISBN: 978-989-758-488-6

Mehrotra, Majhi & Gupta (2009) apply SURF

keypoint detection on the annular iris image. The

usual normalization step is skipped to preserve as

much iris information as possible. The matching

procedure uses the Euclidean distance between the

local features, two keypoints are paired if the distance

between them is less than a fixed threshold. Three

datasets are used in experiments, CASIA, BATH and

IITD. In other experiments, Ismail, Ali & Farag

(2015) before computing the SURF descriptors,

CLAHE (Contrast Limited Adaptive Histogram

Localization) enhancement technique is applied. The

experiments are made on CASIA dataset. The

matching procedure uses a fusion process of the

scores obtained at different levels, the results range

from 99.5% to 100%.

Păvăloi & Ignat (2018) introduce a new method

for handling occluded iris images, using colour

features. In Ignat & Vasiliu (2019) the problem of

missing information is approached with an inpainting

procedure. In Păvăloi & Ignat (2019b) SIFT

descriptors are employed for solving the problem of

iris recognition for irises with missing information.

We compute SURF descriptors in the present

work and use them for iris recognition. For testing our

method we used the original and the standardized

segmented UPOL iris databases. For simulating the

missing information situations, starting from UPOL

dataset, we generated eleven datasets with different

levels of occlusion. SURF feature extraction

algorithm and the matching procedure depend on

some threshold parameters. We tested the impact of

these thresholds on the recognition results. We

compare the results of our experiments with those

obtained using Daugman procedure, implemented by

Masek and with Păvăloi & Ignat (2109a, 2019b)

results. We obtain better recognition rates on nine out

of eleven datasets.

In Section 2 the datasets used in the experiments

are presented. Section 3 outlines the SURF features

extraction process and the recognition method. In

section 4 are presented the results and the conclusions

as well as future directions of research are stated in

Section 5.

2 DATABASES

For our computations we used the well-known UPOL

iris dataset (Dobeš et. Al, 2006, 2004) .This dataset

consists of iris images in .PNG format, all the images

having the same dimensions 576 × 768 × 3 pixels (see

the first image from Fig. 1). The collection contains

iris images for 64 persons, six images for each

individual, three for the right eye and three for the left

eye. The background of all these images is black. We

first performed some experiments on the datasets that

contain irises with full information. We have three

versions for UPOL dataset, the original unsegmented,

a manually segmented version (Păvăloi, Ciobanu &

Luca, 2013) and a standardized segmented collection

(Ignat, Luca & Ciobanu, 2016). In Fig. 1, a sample

from each of these datasets are shown.

(a) (b) (c)

Figure 1: Examples of images in UPOL database: (a) -

original, (b)- manually segmented, (c) – automatically

segmented and standardized.

We performed some computations in order to

decide which of the three versions of these datasets to

further use in our experiments. The original,

unsegmented dataset produced results that were

inferior to those obtained on the other two datasets.

The difference between the recognition rates obtained

on the manually segmented dataset and the

standardized one are almost the same (less than 1%).

Figure 2: Samples from occluded UPOL datasets with

missing information from 5%, 10% to 90%.

We decided to employ in our experiments the

standardized UPOL dataset. The images in this

dataset have 404 × 404 × 3 pixels, the region with iris

information having the same size (the pupil zone has

the same size). The eyelid-eyelash occlusion was

simulated by cutting some regions from the lower and

Occluded Iris Recognition using SURF Features

509

upper part of the iris (see Fig 2). We eliminated more

information from the upper part of the iris image than

from the lower part. We eliminated iris information in

such a way as to obtain images that have about 5%,

10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%

and 95% occlusion of the annular region of the iris.

3 SURF FEATURE EXTRACTION

AND RECOGNITION METHOD

SURF, i.e. Speeded-Up Robust Features introduced

in (Bay, Tuytelaars & Van Gool, 2006) is one of the

most employed keypoints detector. As its name

suggests, it is a faster and better version of SIFT

descriptor. Around these keypoints, local features are

extracted, having the same size, either 64 or 128

components. We chose for our experiments the

variant with 128 elements. The number of keypoints

that the SURF method computes, depends on the

content of each image, two different images have

different number of keypoints associated. For each

one of these keypoints a feature vector with 128

components is calculated.

For the same image, one can compute different

number of keypoints, depending on the choice of

parameters involved in the SURF process, such as the

threshold for the Hessian matrix or the number of the

Gaussian pyramid octave or the number of octave

layers within each octave.

We describe in the following the matching

procedure between the keypoints of two images. The

algorithm is an adaptation of the technique developed

in (Păvăloi & Ignat, 2019). Assume we want to match

two images, I and J. We first apply the SURF

procedure to both of them. The SURF algorithm

computes around the detected keypoints, m feature

vectors,

,, ,

tt t for image I and n feature vectors

,,,

dd d for image J.

Each image has a different number of feature

vector associated (n≠m). We match keypoints from

image I with keypoints from image J in the following

way. We first compute all the distances from each

feature vector associated with image I to all the

feature vectors associated to image J.

The keypoint represented by feature vector t

matches the keypoint represented by feature vector d

if the following relation is fulfilled:

dist(t

, d

) ≤ T dist(t

, d

), j≠k (1)

where T>0 is a threshold parameter that allows to

control the matching process. This type of matching

is called nearest neighbor ratio matching procedure.

In equation (1) we tested three distances: Euclidean,

Manhattan and Canberra.

After applying the SURF algorithm and performing

the matching procedure, we define the distance

between two images as the average of the distances

between the coordinates of the matching keypoints.

d(I,J)=average{|| p

– q

||} (2)

where p

denote the coordinates of a keypoint from

image I which is paired with a keypoint from image J

with coordinates q

. We used the Manhattan formula

for computing the distances between the coordinates

of the keypoints.

Denote by I the test image, and by S={J

, p=1,s} the

training set. Assigning a label to image I, using the

above described matching procedure is done by

computing the following steps:

Step 1: Apply the SURF algorithm to I and all the

images in S.

Step 2: Compute the number of matching points

between the test image I and all the images from S

using formula (1).

Step 3: Denote by m

the number of matching points

between the test image I and image J

. Let q be the

maximum number of matching points, i.e.

q=max{m

; p=1,…,s} (3)

Select from the training set a subset of images that

have at least "q-1" matching keypoints with the test

image I.

Step 4: We compute the distances between test image

I and the images selected after Step 3. Choose the

image from the subset at minimum coordinates

distance. The selected image will provide the label for

the test iris image.

We tested the above mentioned three distances and

the best results were obtained with Manhattan

distance, so this is the distance that was used in our

computations.

4 RESULTS

The number of extracted keypoints and feature

vectors computed with SURF depends on the Hessian

matrix threshold and the numbers of the pyramid

octave. For this work we have tested how these

parameters influence the recognition results. We

experimentally found that the most important

parameter, the parameter which made the difference,

is the threshold for the Hessian blob detector in SURF

features extraction method. We denote this threshold

parameter by H. Our computations show that this

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

510

parameter is of most significance for the method we

propose in this paper. Different values for this

parameter produce sets of keypoints of different sizes.

Smaller values for this threshold provide more

detailed information about the analysed image, thus

improving the iris recognition results. On the other

hand, it is of interest to have as few SURF descriptors

as possible, because this reduces the computation

time in the matching process. In our computations we

adopted a Leave-One-Out type of recognition

method.

For each dataset, we have employed personalized

values for the parameter H. We first computed some

statistical values for each set of feature vectors, and

different values of the threshold H parameter. These

values are the total number of keypoints for the entire

dataset, the minimum and maximum number of

keypoints for the analysed images and the average

number of keypoints. For this purpose, we used the

segmented standardized UPOL dataset. The results

are in Table 1.

Table 1: SURF statistics for the standardized segmented

dataset.

Statistics/

H ↓

Total

Min Max Average

100 195575 138 839 509.31

150 140669 81 705 366.33

200 106907 49 580 278.4

250 84644 26 484 220.43

300 68704 10 423 178.92

400 48492 4 321 126.28

500 36368 2 253 94.71

Obviously, small values for the H parameter

produces large numbers of features and as its value

increases the number of feature vectors decreases. We

performed computations on both the original

unsegmented UPOL dataset and for the standardized

segmented dataset. For the first dataset the number of

feature vectors was bigger than for the second one.

For example, for H = 500, the original dataset has an

average of 94840 features (and a recognition rate of

86.46%), and the standardized segmented dataset

only 36368 features (recognition rate 94.71%). The

reason why we use in our further computations the

standardized segmented version of UPOL dataset is

that we obtained better results on this dataset than on

the other datasets.

It is obvious that the number of well recognized

images increases when the number of extracted

keypoints increases. One has to carefully choose the

H parameter in order to balance the recognition

results and the computing time.

Considering the following values for the threshold

parameter involved in the matching process, T∈{0.6,

0.7, 0.8}, and for the Hessian related threshold

H∈{100, 200, 250, 300, 400, 500} we obtain the

recognition results depicted in Table 2.

Table 2: Number of well recognized images on

standardized UPOL dataset.

H 100 200 250 300 400 500 Avg

0.6 382 378 372 369 358 342 366.83

0.7 384 378 373 371 364 350 370.00

0.8 381 375 369 363 347 335 361.67

Although the recognition results are very good,

the processing time for H =100 is very big, because

the number of detected keypoints is large, and each

feature vector associated with these keypoints has to

be compared with all the others. Anyway, for T= 0.7

and H=100 the maximum number of recognized

image (384) is achieved.

We also computed the number of well recognized

images on the original UPOL dataset. The two

threshold parameters that we analyse in this work

were T∈{0.5, 0.6, 0.7}, and H∈{200, 300, 400, 600,

1000}, and the results are in Table 3. Note that the

recognition results are lower than those computed for

the segmented, standardized dataset. This emphasizes

the fact that a good segmentation procedure yields

good recognition results.

Table 3: Number of correctly recognized images on original

UPOL.

T/H 200 300 400 600 1000 Avg

0.5

376 365 351 321 260

334.6

0.6 376 364 346 318 253 331.4

0.7 371 345 330 296 222 312.8

Considering the parameter H=200, and computing

the above mentioned statistics for the eleven datasets

with different levels of missing iris information we

get the results from Table 4.

From the statistics in this table we deduce that the

number of feature vectors decreases as the level of iris

occlusion increases. Starting with 60% missing

information the average number of keypoints is less

than 200 and for 90% and for 95% missing

information the average is less than 100. Sure, one

can force SURF to compute more features by

choosing smaller values for the H threshold

parameter, but this comes with the inconvenience of

a very long computing time.

Occluded Iris Recognition using SURF Features

511

Table 4: SURF statistics for occluded datasets H= 200, T=

0.7.

Statistics/

Occlusion

Total

Min Max Average

05 11095 54 5 288.95

10 11090 57 571 288.81

20 10579 54 541 275.51

30 98312 45 508 256

40 91958 245 458 239.4

50 84561 33 433 220.2

60 73765 29 365 192.1

70 55860 28 284 145.4

80 43606 15 218 113.56

90 3220 15 149 83.87

95 2413 15 117 62.85

In Table 5 we present the number of correctly

classified images for H=200, T=0.7 for irises with

occlusions. We performed computations on the entire

dataset and separately on the left eye and on the right

eye.

Note that the results are very good (98.95% for

70% missing information). For images with 10% or

5% iris information the SURF descriptor is unable to

compute sufficient keypoints to have good

recognition results. Note that, on average, the

recognition results are similar for the complete

dataset, right eye and left eye (95.45% for both eye

dataset, 95.97% for the left eye, and 96.07% for the

right eye). An interesting situation occurs for the left

eye dataset with 70% missing information, in this

case the recognition rate is 100%.

Table 5: Recognition results for dataset with occlusion, H=

200, T= 0.7.

Recogn. /

Occlusion

Left+Right

Left

Right

05 382 192 190

10 381 191 191

20 382 190 192

30 381 192 190

40 380 192 189

50 383 192 191

60 379 190 191

70 380 192 190

80 378 190 188

90 339 168 175

95 267 138 142

In the sequel, we focused our attention on the two

datasets with the smallest amount of iris information.

For H∈{25, 50, 100, 150 our method applied on the

datasets with 90% and 95% of missing information

will produce the results Table 6.

Table 6: SURF statistics for occluded datasets with 90%

and 95% occlusion, T=0.7.

Total no Min Max Av

90%

H=50

66871 89 230 174.14

90%

H=100

48468 36 194 126.22

90%

H=150

39035 20 170 101.65

95%

H=25

57093 96 183 148.68

95%

H=50

50753 59 176 132.17

95%

H=100

36891 25 155 96.07

95%

H=150

29459 17 136 76.74

The recognition results in terms of correctly

classified images, for the datasets and parameters used

in Table 6 are in Table 7. For the complete dataset with

90% missing information one gets a very good

recognition rate of 95.31%, for T=0.7, and H=50. For

the complete dataset with only 5% of iris information,

the best recognition result (89.58%) is obtained for

T=0.7 and H=25. For the right eye dataset with 95%

missing information and H=25 we get an excellent

95.31% recognition result. We remarked that, in this

situation, the right eye dataset has a better average

recognition rate (90.25%) than the complete dataset

(86.97%) and the left eye (87.27%).

Table 7: Recognition results for occluded datasets with 90%

and 95% missing information, T=0.7.

Left+Right

Left eye Right eye

90%

H=50

366 180 187

90%

H=100

356 176 181

90%

H=150

348 177 178

95%

H=25

344 170 183

95%

H=50

335 169 173

95%

H=100

303 154 157

95%

H=150

286 147 154

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

512

In Table 8 we compare the results obtained with

this method, those obtained in Păvăloi & Ignat

(2019a,2019b) in two papers from 2019, one using

only SIFT features, and the other using colour and

SIFT features. We also compare our results with

Daugman’s Iris Code (Daugman, 1993, 2015),

algorithm that was implemented by Masek (Masek,

2003).

Table 8: Comparison results with other methods for the

segmented standardized UPOL dataset, T=0.7

Methods/

Miss.

Info.

(PI,a) (PI,b) Daugmann-

Masek

Our

method

0 384 374 382 384

05 382 372 381 382

10 380 374 380 381

20 377 376 378 382

30 382 378 377 381

40 380 377 377 380

50 381 374 375 383

60 377 374 370 379

70 375 373 361 380

80 369 375 341 378

90 305 369 316 366

95 174 367 275 344

In the last column of Table 8 the results were obtained

with different values for the blob detection related

threshold, H (the values that provided the best

results). We get better recognition results in nine out

of the twelve analysed experiments. For the cases of

90% and 95% missing information the SURF

procedure extracts very few keypoints and thus the

recognition rate is lower. The method that provides

better results in these situations, although uses a

keypoint detector, but the results are improved in

these case, by the colour features (Păvăloi & Ignat,

2016).

The images from the standardized UPOL

collection are very regular, the annular iris region and

the pupil have the same size in all the images. We

tested our method on images with an irregular

structure of the iris, and non-uniform background, by

performing some computations on the images from

the original UPOL collection. We considered three

cases, namely images with 30%, 60% and 90%

missing information (see Fig. 3).

30% 60% 90%

Figure 3: Examples of iris images with missing information

from the original UPOL database: 30%, 60%, 90%.

We used, as before, H= 200 and T=0.7. The

results of our computations are in Table 9. In this

case, due to the fact that the image contains non-iris

information, the recognition rate decreses more

rapidly as the occlusion increases.

Table 9: Recognition results for the original UPOL dataset,

30%, 60%, 90% missing information.

Missing

info.

Left+Right

Left eye Right eye

0 371 186 186

30 354 180 175

60 342 168 176

90 245 126 119

We analyzed the importance of a standard iris

segmentation by applying our method on the

manually segmented UPOL. The iris images in this

dataset have variable iris and pupil areas (see Fig. 4).

The background is black. We used the same values of

the thresold paramaters, H= 200 and T=0.7. The

results are in Table 10. Note that in this case the

results are lower than those obtained for the other two

datasets. One reason for these results is the fact that

SURF extracts orientation and shape information

around the keypoints. Another reason for these

differences is the distance (2) we use in the

classification process. For very regular images, such

as the images from the standardized UPOL collection,

this distance works. For the other datasets one needs

to find a new distance. A third reason is the fact that

the H parameter need to be carefully chosen for each

dataset in order to obtain good recognition results.

0% 40% 80%

Figure 3: Examples of iris images with missing information

from the manually segmented UPOL database: 0%, 40%,

80%.

Occluded Iris Recognition using SURF Features

513

Table 10: Recognition results for the manually segmented

UPOL dataset, 10% to 90% missing information.

Missing

info.

Left+Right

Left eye Right eye

10 346 178 182

20 342 176 179

30 339 178 176

40 329 175 170

50 322 168 168

60 321 164 168

70 290 160 153

80 259 138 135

90 178 93 106

One way to improve the results for datasets with

irregular shape of the iris images is to add texture

information to the feature vectors extracted around

the keypoints.

5 CONCLUSIONS

This paper presents the computation results on

occluded iris image recognition using SURF features

and an adapted method we previously developed for

SIFT keypoint detection. In experiments, the UPOL

iris dataset was employed. We obtain, in some

situations, better results than those computed with

SIFT based features. We observed that the

recognition accuracy depends on the number SURF

features but after a certain level, the recognition rate

reaches a plateau. For each dataset, the value of the

Hessian threshold parameter used for computing

SURF features must be established after some

experiments. Usually, an average bigger than 200

SURF descriptors for an image seems to give very

good recognition results. Sure, for datasets with 90%

or 95% missing information that target cannot be

reached. Experiments have revealed that a good value

for the matching threshold parameter is 0.7.

In our future work we intend to employ also other

datasets, as UBIRIS for example. In our future

experiments we are interested in combining SURF

method with texture features and the colour

information.

REFERENCES

Daugman, J. G. (1993). High confidence visual recognition

of persons by a test of statistical independence. IEEE

transactions on pattern analysis and machine

intelligence, 15(11), 1148-1161.

Daugman, J. (2015). Information theory and the iriscode.

IEEE transactions on information forensics and

security, 11(2), 400-409.

Bowyer, K. W., & Burge, M. J. (Eds.). (2016). Handbook

of iris recognition. Springer London.

De Marsico, M., Petrosino, A., & Ricciardi, S. (2016). Iris

recognition through machine learning techniques: A

survey. Pattern Recognition Letters, 82, 106-115.

Harakannanavar, S. S., & Puranikmath, V. I. (2017,

December). Comparative survey of iris recognition. In

2017 International Conference on Electrical,

Electronics, Communication, Computer, and

Optimization Techniques (ICEECCOT) (pp. 280-283).

IEEE.

Nguyen, K., Fookes, C., Jillela, R., Sridharan, S., & Ross,

A. (2017). Long range iris recognition: A survey.

Pattern Recognition, 72, 123-143.

Rattani, A., & Derakhshani, R. (2017). Ocular biometrics in

the visible spectrum: A survey. Image and Vision

Computing, 59, 1-16.

Nguyen, K., Fookes, C., Ross, A. & Sridharan, S. (2017)

Iris recognition with off-the-shelf CNN features: A

deep learning perspective. IEEE Access, 6, 18848-

18855.

Ali, H. S., Ismail, A. I., Farag, F. A., & Abd El-Samie, F.

E. (2016). Speeded up robust features for efficient iris

recognition. Signal, Image and Video Processing,

10(8), 1385-1391.

Mehrotra, H., Sa, P. K., & Majhi, B. (2013). Fast

segmentation and adaptive SURF descriptor for iris

recognition. Mathematical and Computer Modelling,

58(1-2), 132-146.

Bakshi, S., Das, S., Mehrotra, H., & Sa, P. K. (2012,

March). Score level fusion of SIFT and SURF for iris.

In 2012 International Conference on Devices, Circuits

and Systems (ICDCS) (pp. 527-531). IEEE.

Mehrotra, H., Majhi, B., & Gupta, P. (2009, December).

Annular iris recognition using SURF. In International

Conference on Pattern Recognition and Machine

Intelligence (pp. 464-469). Springer, Berlin,

Heidelberg.

Ismail, A. I., Ali, H. S., & Farag, F. A. (2015, February).

Efficient enhancement and matching for iris

recognition using SURF. In 2015 5th national

symposium on information technology: Towards new

smart world (NSITNSW) (pp. 1-5). IEEE.

Păvăloi, I., & Ignat, A. (2018, September). Experiments on

Iris Recognition Using Partially Occluded Images. In

International Workshop Soft Computing Applications

(pp. 153-173). Springer, Cham.

Ignat, A., & Vasiliu, A. (2018). A study of some fast

inpainting methods with application to iris

reconstruction. Procedia Computer Science, 126, 616-

625.

Păvăloi, I., & Ignat, A. (2019). Iris Image Classification

Using SIFT Features. Procedia Computer Science, 159,

241-250.

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

514

Dobeš, M., Martinek, J., Skoupil, D., Dobešová, Z., &

Pospíšil, J. (2006). Human eye localization using the

modified Hough transform. Optik, 117(10), 468-473.

Dobeš, M., Machala, L., Tichavský, P., & Pospíšil, J.

(2004). Human eye iris recognition using the mutual

information. Optik, 115(9), 399-404.

Păvăloi, I., Ciobanu, A., & Luca, M. (2013, November). Iris

classification using WinICC and LAB color features. In

2013 E-Health and Bioengineering Conference (EHB)

(pp. 1-4). IEEE.

Ignat, A., Luca, M., & Ciobanu, A. (2016). New Method

of Iris Recognition Using Dual Tree Complex Wavelet

Transform. In Soft Computing Applications (pp. 851-

862). Springer, Cham.

Bay, H., Tuytelaars, T., & Van Gool, L. (2006, May). Surf:

Speeded up robust features. In European conference on

computer vision (pp. 404-417). Springer, Berlin,

Heidelberg.

Păvăloi, I., & Ignat, A. (2019, November). Iris Occluded

Image Classification Using Color and SIFT Features. In

2019 E-Health and Bioengineering Conference (EHB)

(pp. 1-4). IEEE.

Masek, L. (2003). Recognition of human iris patterns for

biometric identification (Doctoral dissertation,

Master’s thesis, University of Western Australia).

Masek, L. (2003). Matlab source code for a biometric

identification system based on iris patterns.

http://people. csse. uwa. edu.

au/pk/studentprojects/libor/.

Păvăloi, I., & Ignat, A. (2016, August). Iris recognition

using color and texture features. In International

workshop soft computing applications (pp. 483-497).

Springer, Cham.

Occluded Iris Recognition using SURF Features

515