Documente Academic
Documente Profesional
Documente Cultură
KEU 110010
KEU110012
KEU110013
Lim Su Yi
KEU110015
Vivian Koh Ci Ai
KEU110047
KEU110049
Table of Contents
1.0
Introduction ................................................................................................................... 2
2.0
2.1
2.1.1
2.1.2
Results ............................................................................................................... 9
2.1.3
Discussion ........................................................................................................ 11
2.1.4
Conclusion ....................................................................................................... 11
2.2
2.2.1
2.2.3
Result ............................................................................................................... 17
2.2.4
Discussion ........................................................................................................ 17
2.2.4
Conclusion ....................................................................................................... 17
3.0
Overall Discussion........................................................................................................ 18
4.0
Conclusion ................................................................................................................... 20
5.0
6.0
Reference .................................................................................................................... 21
1.0
Introduction
2.0
2.1
Distance Approach
In the old time, only one method will be used for iris recognition. For example, the
frist algorithm for iris location was first proposed by Daugman in 1993. Then
researcher had found out that Hamming distance can be used for matching purpose
and an alternative segmentation method (Wildes, 1997) in which edge detection
operator and Hough transform were used. The upper and lower eyelids will be
explicitly modelled with a parabolic arcs. The method in recognizing iris is keep on
improving and researching by researchers. However, a combination of two
classification methods (Rai, 2014) were proposed. The zigzag collarette area of iris is
chosen as feature extraction due to its complex pattern (Roy & Bhattacharya, 2006).
Then parabola detection technique will be used to detect eyelid and eyelash will be
removed by using median filter. HAAR wavelet and 1D Log Gabor filter will be used
for feature extraction. Then support vector machine will be used as main classifier
followed by Hamming distance. By combining two methods of support vector
machine and Hamming distance approach, the accuracy on CASIA and Check iris
database can be increased if compared to single method.
Classification
Result
(1)
y = yc + r sin
(2)
360. The radius can be set as a constant so that the parametric representation of the
circle can be simplified. Edge detection technique can be applied before adopting
circular Hough Transform to find the circles in the given image.
Hough transform with all edge points (xi , yi) where i = 1, 2,... n can be written as
following:
Note: The highest coordinates (xc, yc, r) is selected as the coordinate of centre and
radius of the circle.
Step 2: Selection of zigzag collarette area
The selected zigzag collarette area provides important iris features as most complex
patterns of iris are captured in this area. Zigzag collarette area is unlikely affected by
eyelids and eyelashes due to its close position with the pupil.
Step 3: Eyelid detection
Upper and lower eyelids detection requires two search regions that are confined
within the zigzag collarette area of the iris and pupil. The width of search region can
be determined by following equation:
6
(5)
The edge image of an eye image can be determined using horizontal edge map due to
the presence of eye lids in upper and lower horizontal region. Eyelids detection is
done by adopting parabolic Hough transformation at each edge point within the search
Step 4: Normalization
Right after a successful segmentation, transforming the iris region will be done so that
it has fixed dimensions. Normalization process is done by using rubber sheet model
devised by (Daugman, 1993). After that, eyelash removal method will be used to
remove eyelashes and restore the underlying iris pattern as much as possible by
recreating Zigzag collarette area pixel that occluded by eyelashes. This can be done
by using the information from their non-occluded neighbors. We need to decide if
each of the pixel present in normalized image is occluded by eyelash. The equation
below can be used to determine the occlusion by eyelash.
If I(x,y) < T, pixel is occluded by eyelash.
(6)
Si =
False, otherwise
8
2.1.2 Results
Table 1
Table 2
10
2.1.3 Discussion
There are several reasons of applying two different techniques for feature extraction
and classification. The reasons are as following:
1. Information about the noise mask could not be obtained by Haar decomposed
feature vector. This information plays a significant role in Hamming distance
classification. Thus, Haar is not suitable for Hamming distance classifier but
SVM classification.
2. The feature vector obtained by using 1D Gabor wavelet is suitable for
Hamming distance but not for SVM.
3. According to extensive testing of different networks, combination of SVM and
Hamming distance approach will give a better recognition accuracy than using
a single method. As SVM alone has relatively high false rejection rate,
Hamming distance is added to further overcome this problem, resulting in
higher recognition rate.
Based on the Table which demonstrate the accuracy comparison between the
proposed method and previous reported approaches, combined SVM and Hamming
distance method achieved 99.91% of accuracy, which ranks the highest among all the
approaches.
2.1.4 Conclusion
An efficient approach for iris feature extraction and recognition can be done by using
the method as discussed above. Zigzag collarette area of iris is chosen for the feature
extraction due to its ability to capture the most significant area from the complex
pattern of iris and hence a higher recognition rate can be achieved. HAAR wavelet
and 1D Log Gabor filter are used to extract feature which will then being used in iris
identification using the combined support vector machine and Hamming distance
approach. Parabola detection and trimmed median filter are also been used to detect
eyelid and eyelash. Comparisons among previous reported approaches and the one
currently proposed were made. It indicates that the recognition accuracy is higher
when using combination of SVM and Hamming distance if compared to either SVM
or Hamming distance. Accuracy in terms of FAR and FRR for this method is
11
exceedingly high for the CASIA database. In short, this proposed approach not only
efficient in case of identification but also verification.
2.2
(NBP) is being proposed (Hamouchene & Aouat, 2014). This method is inspired by
Local Binary Pattern (LBP) method as NBP method are able to capture local
information and in the same time iris texture can be describe better.
Image Acquisition
To obtain image from a person
To using sensor
Preprocessing
To remove useless information from iris image
To extract the region of interest (iris)
To include segmentation (isolate iris ring) and normalization (provide
inavariant iris area, form ROI into rectangular region)*
Feature Extraction
There are two approaches:
Local Binary Pattern (LBP)
Neighborhood Binary Pattern (NBP)
Matching step
distasnce measure between generated iris code and stored iris code are being
calculated**
12
Note
*Due to the two ring of the iris are not co-centric, Integro-differential
operator by Daugman is being used to detect the inner and outer boundaries
(Daugman, 2004).
**Daugman using Hamming distance and threshold around 0.34
13
Every neighboring pixel is compared with the central pixel. If the neighbor pixel is
greater than the central pixel, then it is recorded as 1. If the neighbor pixel is less than the
central pixel, then it is recorded as 0.
The weight pixels are matched with the threshold. If the threshold 1 is matched with the
weight pixel, then the pixel will be the weight pixel. If the threshold 0 is matched with the
weight, then 0 is displayed. The new set of data formed is the LBP data.
The pixel values in the LBP is summed and the central pixel is replaced with this
weighted sum value.
than
the
next
neighbor
otherwise
value
of
is
given.
be given and 0 otherwise. Therefore, we will obtain a binary matrix of the variation
means and we can use it as the template of the iris texture.
As show in the above equation, M1 and M2 are the variation binary codes for the
iris images. S value for the ith block is equal to 1 if the value of M1 for ith block is
equal to the value of M2 for ith block. Nb is representing the total number of blocks
and this value is based on the degree of decomposition of the iris image. If value of
Dis is above certain threshold, the two iris images (1 and 2) are referred as same
person.
Public iris database, CASIA is being used to evaluate the performance of the
system. As suggest by (Hamouchene & Aouat, 2014), three images from each person
are taken as reference and 80 images will be used as test images where each image are
referred as query. For each of the image, LBP histogram and mean variation of the
NBP image are being extracted. Between the querys feature and extracted features,
hamming distance is being calculated. By sorting the hamming distance from most
similar to less similar, the top three is being considered and the query iris is classified
by referring the majority (highest similarity).
16
2.2.3 Result
Figure 15: Recognition rate for LBP and NBP method for each person
2.2.4 Discussion
From the above graph (figure 15), we can see that NBP method is way better than
the LBP method as the LBPs global rate is only 58.75% where NBPs rate is 76.25%.
This is because of NBP method is comparing the neighborhood pixel values with its
adjacent pixel values instead of being threshold by the central pixel values. In other
words, we can say there is relationship among the neighborhood pixel values for NBP
method. This result had shown the robustness and efficiency of the NBP method as
compared with LBP method.
2.2.4 Conclusion
We can conclude that NBP method is having good performance as compared with
LBP. This is because of there is relative connection between the neighborhood pixels
as each one of them is being thresholded by the adjacent neighbor and encoded. Not
17
only that, NBP image is being decomposed into number of blocks where their
variation of mean values are being extracted and encoded. This will result the binary
matrix being used as the feature descriptor for the iris image.
3.0
Overall Discussion
There are many approaches for the iris recognition. However, a classical iris
recognition system involved the series common steps which are the image acquisition,
iris preprocessing, feature extraction and matching step. For the preprocessing stage,
it involves the segmentation, filter and normalization.
From the above two methods for the iris recognition, they go through the common
series steps but each method is using different approach for the feature extraction and
matching step. The table below showed the summary for comparison between the
methods.
Table 3: The comparison between method 1 and method 2 for the iris recognition
Method 2
Distance Approach
Approach
Segmentation
Segmentation
Zig-Zag Collarette
Iris region
Feature Extraction
a.
18
Matching
Matching
a. Hamming Distance
Advantage
-
Rotational invariance
Disadvantage
-
Disadvantage
-
From the two methods above, they applied the pattern recognition approaches in
different way to identify the iris pattern. Feature extraction is considered as the most
importance stage for the pattern recognition because it extracts the relatively most
significant information from a given data, in this case is the iris pattern. Method 1
extracts the feature information through the wavelet. Wavelet is a useful tool with rich
mathematical contents and great applications. One of its applications is the image
analysis for pattern recognition. It is a function that is isolated with respect to time
and frequency, while the Fourier transform only respect to the frequency information.
In method 1, one of the wavelet transform used is the efficiently computable Haar
wavelet transform. The iris image is mapped from the space of pixel to that of Haar
wavelet feature that contains rich description of the pattern. On the other hand,
method 2 applies the Neighborhood-Binary Pattern (NBP) in the feature extraction.
19
Generally, NBP is based on the Local Binary Pattern (LBP) which is an effective
texture operator which labels the pixel of an image by thresholding the neighborhood
of each pixel and considers the result as binary number. Thus, two different
approaches in the feature extract may result in different results for the recognition
rate, each approach with its own strength and weakness as shown in the table 3.
After the feature extraction, classifier is applied to classify or identify the iris pattern
for the identification. The method 1 uses extra one classifier known as Support Vector
Machine which is a supervised learning models work with an associated algorithm in
the pattern recognition. With the additional classifier, the error occurred is reduced
compare to method 2.
The results from both methods showed that method 1 is more stable and robust which
was around 99.91% of the average recognition rate, whereas LBP was 58.75% and
NBPs rate was 76.25%. However, it does not mean the method is better because
those methods are carried out in the different designed experiments.
4.0
Conclusion
In conclusion, the approach toward for iris recognition is different for both methods
and each approach with its own advantage and disadvantage respectively. However,
researches can be carried out for novel iris recognition with the combination of both
methods. Wavelet transform is used to decompose a given image to different kinds of
images and then processed with local difference between image pixel and its
neighborhood to build the LBP or NBP. The combination may increase the robustness
for the iris recognition.
5.0
Future Outlook
The future works involve are resolving challenges faced in developing an ultimately
high accuracy and reliable iris recognition system. Besides than solving the issues
caused by special iris condition as mentioned in introduction, noise that caused by
environments factors such as unfavourable lighting, large stand-off distances and
moving objects (Ross, 2010). These conditions can caused nonideal irides images to
20
6.0
Reference
Ali, H., & Salami, M. J. (2008). Iris recognition system by using support vector
machines. In International Conference on computer and communication
engineering (pp. 516-521).
Burge, M. J., & Bowyer, K. W. (2013). Handbook of iris recognition. Springer.
CASIA iris image database (v1.0), The National Laboratory of Pattern Recognition
(NLPR), Institute of Automation, Chinese Academy of Sciences (CAS), 2006
Chen, C., & Chu, C, (2005). High performance iris recognition based on 1-D circular
feature extraction and PSO-PNN classifier. Expert Systems with Applications,
36(7), 10351-10356.
Daugman, J. G. (1993). High confidence visual recognition of persons by a test of
statistical independence. Pattern Analysis and Machine Intelligence, IEEE
Transactions on, 15(11), 1148-1161.
21
Daugman, J. (2004). How iris recognition works. Circuits and Systems for Video
Technology, IEEE Transactions on, 14(1), 21-30.
Grabowski, K., & Napieralski, A. (2011). Hardware architecture optimized for iris
recognition. Circuits and Systems for Video Technology, IEEE Transactions
on, 21(9), 1293-1303.
Hamouchene, I., & Aouat, S. (2014). A New Texture Analysis Approach for Iris
Recognition. AASRI Procedia, 9, 2-7.
Masek,
L.
(2003). Recognition
of
human
iris
patterns
for
biometric
Ross, A. (2010). Iris recognition: The path forward. Computer, 43(2), 30-35.
Westmoreland, B. F., Lemp, M. A., & Snell, R. S. (1998). Clinical Anatomy of the
Eye. Oxford: Blackwell Science Inc.
22