2010 International Conference on Technologies and Applications of Artificial Intelligence
A Fusion Technique for Iris Localization and Detection
Chai Tong Yuen1, Saied Ali Hosseini Noudeh2, Mohammad Shazri2 and Mohamed Rizon3
Department of Mechatronics and Biomedical Engineering Universiti Tunku Abdul Rahman (UTAR) Kuala Lumpur, Malaysia
Research and Development Department Extol MSC Sdn Bhd Kuala Lumpur, Malaysia
Department of Electrical Engineering King Saud University Riyadh, Kingdom of Arab Saudi
Email: chaity@utar.edu.my, Ali.Hosseini@extolcorp.com, Shazri.Shahrir@extolcorp.com, mjuhari@ksu.edu.sa
AbstractAn automated algorithm to localize irises for Middle East individuals had been developed in this research. Histogram equalization, Logabout, Difference of Gaussian (DoG), wavelet transformation, Principle Component Analysis (PCA) and Artificial Neural Network are popular techniques used for image processing, feature extraction and classification. A fusion of these techniques had been introduced to compensate the effects of illumination and head orientation for iris detection. The algorithm was tested with Middle East face database through experiments. In this paper, iris candidates are extracted from the valley of the detected face region after being pre-processed. All the detected iris candidates will go through wavelet transform. The wavelet coefficients are then reduced and extracted by PCA. Finally, Softmax Backpropagation Neural Network (SBNN) works as the iris classifier. The impact of the pre-processing techniques on the performance of the proposed algorithm was studied. The proposed algorithm had achieved a success rate of 90.5% with 0% false positives being reported. Keywords-iris detection; principle component analysis; neural network; wavelet transform

INTRODUCTION

Face recognition has been applied in surveillance, access control, security system and identity verification. In past few decades, face recognition has experienced a tremendous growth in researches and attracted substantial attention from various discipline. It is never an easy task for a computer to recognize human faces automatically compared to human. Thus, automatic face and facial features detections are vital for face recognition development. Face detection is crucial in the field of automatic face recognition. The detected faces can be used to allocate facial features and create samples for face recognition. Face detection can be developed based on still-image [1], [2] and video sequences [3]. Iris recognition [8]-[10] has been one of the alternatives for identity verification. Iris could be a
salient and stable feature for face recognition. However, the performance of the system still depends on how consistent the input samples are. Variability from one image to another is caused by location and orientation of face relative to the image border, lighting, background and facial expressions. Therefore, iris location can be a reference point to locate the face accurately [4]. Eye detection can be divided into two categories, active [5] and passive [6], [21]. Active eye detection uses external source for illumination. This will evoke the physical characteristic to utilize the eye localization. Although it seems to be more accurate and robust but a lot of false positives have been captured [5]. Passive method uses the visual spectrum to deduce the location of eyes. Iris detection can be done by using passive method. The most challenging part for iris detection is to eliminate features with low intensity such as eyebrow, hair, beard and moustache. In recent research, the boundary of the pupil is found by using gray level information. A number of optimizations have been made on traditional Hough Transform [7]. Canny edge detection and Circular Hough Transform are still popularly used to refine and locate the iris position [7], [8], [11]. 2D Log-Gabor Filters is being used for feature extraction together with hamming distance for iris matching [8]. Eyelids and eyelashes have been defined as occlusions in [11]. To detect eyelids, the contrast between eyelids and iris is calculated as the deviation from average intensity. Eyelashes detection is actually the high frequency component which falls into the iris area after applying the low pass filtering. Improvement on 2D Log-Gabor has been achieved in [9]. However, the difficulty in finding an accurate threshold and higher false acceptance rate have been noticed. Wavelet decomposition [10] is introduced to divide the image into three wavelet bands before smoothing operation and multiscale edge detection. Gradient value is calculated from each

978-0-7695-4253-9/10 $26.IEEE DOI 10.1109/TAAI.2010.16
level of wavelet transform in iris detection. In [12], Daugman Circular Differential Operator is proposed to search by assuming that the center of the image is pupil. Average Square Shrinking (ASS) is then used to refine the pupil search. There is a possible contribution to failed detection if the assumed center is not the pupil. Daugman integro-differential operator [13] has been utilized for circles searching. It calculates Gaussian size and gradient coefficient by using Daugmans method. Besides, neural network [14] is trained to find the center coordinates of the iris. However, false positives of iris detection such as eyelash and eyelids are still an issue. We observed, the recent researches in [11]-[14] have tested with iris dataset instead of face dataset. It could be invasive and inconvenient since the user needs to put their eyes in front of the camera in terms of practical implementation. The aim of this research is to improve the performance of iris detection and implement on the Middle East face database. In Section II, the outline of the proposed algorithm which includes face localization stage, pre-processing stage, iris candidate detection, wavelet transform, feature extraction stage and classification stage will be discussed. The results will be tabled and discussion in Section III. Finally, conclusion and future works for this research are shared in Section IV. II. PROPOSED METHOD
C. Iris Candidates Detection After the illumination normalization, valley extraction has to be carried out. Each pixel in the face region:
V ( x, y ) = G( x, y ) I ( x, y )
Where G x, y and I x, y denote the value obtained from grayscale closing and intensity value. Region which consists of pixels x, y such that V x, y is greater than or equal to a threshold value are determined to be valleys. In iris candidates detection, the proposed algorithm performs similar method as [2] in iris candidate detection. First, it computes the costs C x, y for all pixels in the valleys and selects m pixels according to non-increasing order that give the local maxima of candidate locations as shown in (2).

C (x, y ) as the iris

C(x, y ) = C1 (x, y ) + C2 (x, y )
C1 (x, y ) is the mean crossing of the row and column pixels and C 2 ( x, y ) is the intensity difference
between the centre part and boundary part of a square region. All the selected candidates will then be passed for Wavelet Transform. D. Wavelet Transform A few processing steps have been implemented to select the best pairing of iris candidates. In this stage, Haar Wavelet has been chosen as the mother wavelet as shown in (3). The mother wavelet is a continuous function in both the time domain and the frequency domain. The main purpose of is to generate the daughter wavelets which are the translated and scaled version of the mother wavelet. The translation can be done via variable while the variable is to scale the size of the window. Wavelet Transform does a better job of describing the local details of an image than standard Fourier Transform [18]. This is because the actual implementation of (3) consists of Discrete Wavelet Transform (DWT) which is done by using Dyadic Sampling [19]. (3) The scanner convolves through the face image via a pixels. At each instance of window with the size of window, a multiple sets of coefficients at different combination of high and low pass filters is created. The set of coefficients chosen for Principle Component Analysis (PCA) is the high (row) and low (column) pass filter sub-

A. Face Localization Stage A total of 60 images from the Middle East face database are captured by using Panasonic SDR-H280. In this experiment OpenCV Haar Cascade based face detector [15] has been used to localize the faces of Middle East individuals. B. Image Pre-processing Stage This section describes the techniques chosen for image processing. The proposed algorithm will first apply grayscale closing on the image followed by illumination normalization. The illumination normalization methods are Difference of Gaussian (DoG) [16], histogram equalization and Logabout [17] which will be applied to the images respectively. These methods aim to compensate the effect of lighting and improve the quality of the images for iris detection. The impact of applying the each suggested method on the performance of iris detection had been compared and tested through the experiments. Second, the images have to undergo white spot deletion [4]. This is to reduce the reflection of lighting on the iris. A 3x3 kernel had been used to remove the light spot on the iris. The center pixel of the kernel will be replaced by the smallest intensity value within the kernel as it moves along the face region.
band. E. Feature Extraction Stage PCA is used as the reduction method for feature extraction. The motivation of using PCA is to manage the output of Wavelet Transform. It eliminates less significant basis which is generated through Eigen-decomposition thus focusing more on the important features. The wavelet coefficient of the HL sub-band for the iris and non-iris images after the wavelet transform will be extracted by using PCA. The resolution of the final images will be 10x10 pixels. (4) shows the projection of a set of Eigen Vectors generated from Wavelet Coefficient (V, Column Vector) onto Wavelet Coefficients (W) to generate a set of projected vectors (P). (4) F. Classification Stage In this stage, Softmax Backpropagation Neural Network (SBNN) has been developed as the classifier for iris. The features extracted by PCA are fed to neural network which then creates a non-linear mapping from input to output. The neural network used in this research is a Softmax Cross Entropy variant. Cross Entropy is chosen because of its definition of error (5) which reflects the nature of problem which is a classification problem. It shows the error definition of a two class classification problem where is the exemplar and is the probability interpretation via softmax transfer function (a sigmoid transfer function is more suited for regression [20]). (5)

The subjects are captured with the eyes open to be eligible for iris detection. Training and testing data have been prepared separately to reflect the performance of the SBNN. The ability of the classifier to recognize irises and reject non-iris objects has been evaluated through experiment. 56 positive samples and 26 negative samples are used for training while 44 samples have been used for validation set. The SBNNs training parameters used in this experiment are shown in Table I. The standard for successful iris detection has been set as less than 5 pixels deviation.
TABLE I. TRAINING PARAMETERS FOR SBNN Parameters Input nodes Hidden nodes Output nodes Learning Rate Sigmoid Gradient Value 2 0.03125 3
TABLE II. PERFORMANCE FOR CONVENTIONAL IRIS DETECTION ALGORITHM Processing Technique Histogram Equalization Logabout DoG Iris Detection Rate 0.524 0.770 0.443 False Positive Rate 0.082 0.098 0.290
TABLE III. PERFORMANCE OF THE PROPOSED IRIS DETECTION ALGORITHM Processing Technique Histogram Equalization Logabout DoG Iris Detection Rate 0.897 0.905 0.857 False Positive Rate 0.143 0.000 0.143

RESULTS AND DISCUSSION

The proposed algorithm had been tested with Middle East face database. There are 60 images captured from 6 different individuals in this database. The images are pixels. captured in grayscale with the size of Most of the captured images are frontal with scarf and plain background as shown in Fig. 1.

EER 0.043 0.033 0.071

From Table II, the conventional iris detection algorithm [4] had shown some changes in performance when different processing techniques were implemented. The highest iris detection rate is 77% with 9.8% false positive detection when Logabout chosen as the processing method followed by 52.4% successful iris detection with 8.2% false positive detection from histogram equalization. DoG has the worst performance on iris detection which is 44.3% with 29% of false positives being detected.
Figure 1. Sample images from Middle East face database.
The success rate for iris detection, false positive rate and Equal Error Rate (EER) for the proposed method are shown in Table III. The proposed algorithm with Logabout has achieved the best iris detection rate, 90.5% and 0% false positive rate. The reported EER is 0.033. Histogram equalization and DoG have shown similar performance in iris detection, which are 89.7% and 85.7% respectively. Both of the methods have registered 14.3% false positive rate. The proposed algorithm for iris detection has shown an improvement of 41.4% with DoG as the processing method compared to the conventional method. 37.3% and 13.5% of improvements have been noticed on the proposed algorithm with histogram equalization and Logabout. A reduction of 14.7% and 9.8% false positive rate for DoG and Logabout has been reported while 6.1% increment of false positive rate is reported for histogram equalization. The experiment proves that the proposed method has achieved higher iris detection rate and lower false positive rate compared to the conventional iris detection method. It performs better in Middle East individuals even with the presence of beard and moustache. The previous method [4] which is based on the statistical method and template matching has limited robustness when facing samples which vary from the selected template in terms of faces, orientation and illumination. However, the scarf, eyebrows and imperfectly detected face region as shown in Fig. 2 have been the problems that contribute to the false positive detection of the proposed algorithm.

ACKNOWLEDGMENT The authors gratefully acknowledge the support from Universiti Tunku Abdul Rahman and Extol MSC Sdn Bhd. REFERENCES
[1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] W. J. Chew, L. M. Ang and K. P. Seng, Automatic model based face feature detection system, Proceedings of International Symposium on Information Technology, 2008, pp. 1-6. K. Tsuyoshi and M. Rizon, Iris detection using intensity and edge information, Pattern Recognition, 2008, vol. 36, pp. 549-562. M. Balasubramaniam, S. Palanivel, and V. Ramalinga, Real time face and mouth Recognition using radial basis function neural networks, Expert Systems with Applications, 2008, pp. 1-8. T. Y. Chai, M. Rizon and M. Shazri, Real time detection of face and iris, Wseas Journal Transactions on Signal Processing, 2009, vol. 5, no. 6, pp. 209-218. P. Wang, M. B. Green and Q. Ji, Automatic eye detection and its validation, IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops, vol. 3, 2005, pp. 164. Q. Ji, H. Wechsler, A. Duchowski and M. Flickner, Special issue on eye detection and tracking, Computer Vision and Image Understanding, Elsevier Science Inc., vol. 98, no. 1, Apr. 2005. S. Ziauddin and M. N. Dailey, A robust hybrid iris localization technique, 6th Int. Conf. on Electrical Engineering/Electronics, Computer, Telecommunications and Information Technology, 2009. V. C. Subbarayudu and M. V. N. K. Prasad, A novel iris recognition system, 3rd Int. IEEE Conf. on Signal-Image Technologies and Internet-Based System, 2008, pp. 883-887. L. Ma, T. Tan, Y. Wang and D. Zhang, Efficient iris recognition by characterizing key local variations, IEEE Transactions on Image Processing, vol. 13, 2004, pp. 739-750. M. Nabti and A. Bouridane, An effective iris recognition system based on wavelet maxima and gabor filter bank, 9th Int. Symposium on Signal Processing and Its Applications, 2007, pp. 1-4. M. A. Mohamed, M. A. Abou-El-Soud and M. M. Eid, Automated algorithm for iris detection and code generation, Int. Conf. on Computer Engineering and Systems, 2009, pp. 475-481. M. Shamsi, P. B. Saad, S. B. Ibrahim and A. R. Kenari, Fast algorithm for iris localization using daugman circular integro differential operator, Int. Conf. of Soft Computing and Pattern Recognition, 2009, pp. 393-398. H. Zheng and F. Su, An improved iris recognition system based on gabor filters, IEEE Int. Conf. on Network Infrastructure and Digital Content, 2009, pp. 823-827. R. D. Labati, V. P. Fellow and F. Scotti, Neural-based iterative approach for iris detection in iris recognition systems, IEEE Symposium on Computational Intelligence in Security and Defense Applications, 2009, pp. 1-6. P. Viola and M. J. Jones, Robust real-time face detection, International Journal of Computer Vision, 2004 vol. 57, no.2, pp. 137-154. X. Tan and T. Bill, Enhanced local feature sets for face recognition under difficult lighting conditions, IEEE Int. Workshop on Analysis and Modelling of Faces and Gestures, 2007, pp. 235-249. H. Liu, W. Gao, J. Miao and J. Li, A novel method to compensate variety of illumination in face detection, Joint Conf. on Information Sciences, 2002, pp. 692-695. A. Graps, Introduction to Wavelets, IEEE Computational Science and Engineering, 1995, vol. 2, no. 2, pp. 50-61. S. G. Mallat, A theory for multi-resolution signal decomposition: the wavelet representation, IEEE Transactions on Pattern Analysis and Machine Intelligence, 1989, vol. 11, no. 7, pp. 674-693. C. M. Bishop, Neural networks for pattern recognition, Oxford University Press, 1996. M. Shazri, N. Ramlee, Chai Tong Yuen, "Wavelet PCA/LDA Neural Network eye detection", IEEE Conference on Cybernetics and Intelligent Systems (CIS), 2010, pp. 48 - 51.

[13] [14] Figure 2. Sample images for failed iris detection. [15]

CONCLUSION

From the experiments, we can conclude that the proposed algorithm which fuses the techniques from [4] with Wavelet Transform, PCA and Neural Network has shown a better recognition result for iris detection on the Middle East face database. The algorithm has achieved 90.5% successful detection with no false positive being detected. Logabout has been proven as the most effective lighting compensation method via the experiments. In future, the proposed algorithm will be tested with a larger Middle East face database which covers a variety of individuals, facial expressions, head orientations and lighting conditions for face recognition system.
[16] [17] [18] [19] [20] [21]