Arm Tabanlı Gömülü Sistemlerde Kulak Tanıma Sisteminin Gerçeklenmesi

Kaçar, Ümit
Süreli Yayın başlığı
Süreli Yayın ISSN
Cilt Başlığı
Fen Bilimleri Enstitüsü
Institute of Science and Technology
Günümüzde teknolojinin gelişmesi ile beraber biyometrik sistemler önem kazanmıştır. Şifre, pin gibi insanın unutabileceği ve çalınması kolay güvenlik sistemlerin yerini yavaş yavaş biyometrik sistemler almaya başlamıştır. Parmak, iris, avuç içi ve yüz gibi birçok biyometrik sistem üzerinde çalışmalar yapılmıştır. Özellikle son zamanlarda temassız biyometrik sistemlerin önemi daha fazla artmıştır. Temassız biyometrik sistem olan yüzün jest, mimik, makyaj vb. sebeplerden dolayı değişikliğe uğraması tanıma oranını azaltmaktadır. Bu sebeple başka bir temassız tanıma sistemi olan kulak tanıma sistemi ön plana çıkmaya başlamıştır. Kulağın yüze göre düşük uzaysal çözünürlüğe sahip olması, renk dağılımın daha düzenli olması, ışık değişimine göre daha az etkilenmesi, arka plan görüntüsünün belli olması ve jest, mimik, makyaj vb. sebeplerden dolayı değişikliğe uğramaması gibi üstünlükleri vardır. Kulak tanıma sistemleri için bilgisayar ortamında birçok algoritma geliştirilmiş ve başarılı sonuçlar elde edilmiştir. Ancak gömülü sistemde gerçekleştirilen kulak tanıma sistemi çok azdır. Bu tez çalışmasında kulak tanıma sistemi, görüntü işleme için kullanılan diğer gömülü sistemlere göre daha küçük hafıza ve daha yavaş işlemci hızına sahip olan ARM tabanlı STM32 Cortex-M4 mikromedia kartı ile gerçekleştirilmiştir. Kartın avuç içine sığacak kadar küçük olması, kullanıcı için arayüz oluşturabilme özelliği, düşük güç tüketimi, dokunmatik ekrana sahip olması ve prototip bir yapıda olması ticari uygulamalar için de ideal bir durum sağlamıştır. Kulak tanıma sistemi için PCA ve DCVA olmak üzere iki yöntem kullanılmış ve birbirleri ile karşılaştırılmıştır. Bu yöntemlerin gömülü sistemlerde gerçeklenmesi için en önemli problem özdeğer ve özvektör çözümüdür. Bunun için de Jacobi ve QR olmak üzere iki algoritma kullanılmıştır. Jacobi algoritmasının QR algoritmasından daha hızlı ve doğru bir şekilde özdeğer ve özvektör problemini çözdüğü gösterilmiştir. Yöntemlerin karşılaştırılması önce Matlab programında yapılmış sonra diğerine göre daha iyi yöntem olan Jacobi algoritması kullanarak PCA ve DCVA uygulaması gömülü sistemde gerçekleştirilmiştir. Test sonuçlarında Matlab programı ile bulunan değerlere çok yakın değerler bulunmuştur. Böylece sistemin doğruluğu da gösterilmiştir. Ayrıca gömülü sistemde gerçekleştirilen kulak tanıma sistemi gerçek kulak görüntüsü ile test edilmiş ve sistemin başarılı sonuçlar verdiği görülmüştür.
Biometrics is an automated recognition of individuals based on their behavioral and biological characteristics. A biometric characteristic is a biological or behavioral property of an individual that can be measured and from which distinguishing, repeatable biometric features can be extracted for the purpose of automated recognition of individuals. Nowadays, biometric systems gained importance, together with the increasing technology. Biometrics systems are best alternative to old-fashioned systems which use passwords to verification, because it is possible to encounter the problems such as forgotten passwords or unauthenticated attempts. There has been numerous studies on biometric systems such as finger, palm print/vein, iris and face recognition. There are several advantages and disadvantages of each biometric system. It is one of the most important criteria in design of a biometrics system that our system should realise the difference of an authorised and unauthorised attempts in an optimal way. With the rapid advancement of the technology, production costs in electronic manufacture have fallen and it could be said that prices in sensors and other electronics equipments reached to an equilibrium level. However, the imaging equipments which are used in retina and iris recognition systems are still fairly expensive. On the other hand, systems which are based on face and ear recognition do not necessitate person’s contact with any kind of sensors, they are so called “contactless” and do not cause hesitation by users. For these reasons, face and ear biometrics are more preferable to other methods. Biometric accuracy of face recognition systems could be affected by the deformation by gesture, mimics or makeup. Ear recognition systems do not have these drawbacks. Looking into face and ear recognition systems, it is clear to see that ear recognition has a superiority in terms of spatial resolution, uniform color distribution, robust to luminance, constant and clear backgroud, and being nonsensitive to the factors such as gesture, mimics, and makeup. Ear biometric is one of the biometrics that changes the least. Ear images are not affected by emotional expression, illumination, aging, pose, and alike. Because of its static structure, easy collection of the data, and the small dimension of the ear image, The most important study on ear biometric field was made by Iannarelli in 1989. He conducted a comprehensive research using about 10,000 ears which are chosen arbitrarily. He concluded that ear has a distinctive shape in each person. In this research, he defined 12 different measurements to use ear in classification tasks. After that, Burge and Burger proposed that ear can be used as biometric feature. Burge and Burger realized that Iannarelli System is not suitable for machine vision applications. Because it is hard to align the image correctly and localize the first anatomical point. If the first point is not located correctly, the system fails and the measurements change. So, Iannarelli System is not suitable for realizing as an embedded system. There have been many algorithms which are applied to ear recognition tasks and evaluated in comparison with each other. However, these studies are restricted to computer applications and there exists only a few embedded implementations of an ear recognition system. In this work, we prefered ARM-based “STM32 Cortex-M4” micromedia card, even if it has scarse memory and slow processing speed in comparison with other commercial ARM-based boards which can be used for image processing purposes. The reasons why we used this card are that its size is comperatively small, it is easy to build interfaces for our application. Besides, low power consumption, touch screen, and its being in a prototype structure are other reasons which make this board preferable for our application. Two algorithms, Principal Component Analysis (PCA) and Discriminative Common Vectors Approach (DCVA), are used for ear recognition purposes, and the performance of these two algorithms are evaluated. PCA depends on eigenvector method designed to model linear variations in high dimensional data. PCA can be expressed as the oldest statistical tool to analyse the multivariable datasets and also to reduce dimension of dataset. One can also present the patterns of data like the way of their similarities or differences. DCVA is proposed to extract the common vectors of the classes in a training set by eliminating the differences of the samples in each class. A common vector for each individual class is obtained by removing all the features that are in the direction of the eigenvectors corresponding to the nonzero eigenvalues of the scatter matrix of its own class. The most important problem is finding eigenvalue and eigenvector to be able to realize these algorithms on embedded systems. In order to solve this problem, Jacobi and QR algorithms are used. Jacobi iteration method is thought to be the easiest algorithm to find the eigenpairs of a real symmetric matrix. In this algorithm, it is guaranteed that there will be a unique solution for all real symmetric matrices under the condition no zero element on the main diagonal of the matrix. The QR algorithm is one of the best known method to find the eigenpairs of a matrix. QR algorithm still maintains the throne of the solution of non-symmetric matrix. However, this situation can not be said for the symmetric matrix. QR algorithm is also dependent on the QR decomposition. Gram-Schmidt and modified Gram-Schmidt for QR decomposition, two methods have been proposed. The classical Gram Schmidt algorithm can be called as a projection and normalization method. Modified Gram Schmidt (variant of Schwarz-Rutishauser) is more stable than classical method, especially, when the matrix is ill-conditioned. It is proved that Jacobi algorithm is faster and more accurate than QR algorithm to solve the eigenvalue and eigenvector problem. The comparison of these methods fırst performed on Matlab, after that Jacobi algorithm which is better one is implemented on embedded systems by performing PCA and DVCA. The test results of embedded systems are very close the results of Matlab. Thus, the accuracy of the system is proved. In DCVA, Classical Gram-Schmidt method is proved to be numerically unstable again. However, the norm of error did not affect the recognition rate. As a result, ear recognition system is able to tolerate the numerical errors. In addition to that, the ear recognition system was tested with real ear images. uCAM-TTL camera module were used for this. The connection between the camera and micromedia card provided with UART communication. However, the actual image of the ear should be in a bright environment. Because the PCA and DCVA methods are sensitive to the dark environment. Besides, the ear of person should be properly directed to camera. Exposure shift could cause inaccuracy in our results. After tackling this undesirable issue, the recognition rate of our designed system is quite good.
Tez (Yüksek Lisans) -- İstanbul Teknik Üniversitesi, Fen Bilimleri Enstitüsü, 2013
Thesis (M.Sc.) -- İstanbul Technical University, Institute of Science and Technology, 2013
Anahtar kelimeler
Temel bileşenler analizi, Gömülü sistemler, Kulak Tanıma, Ayırt Edici Ortak Vektör Yaklaşımı, Jacobi Algoritması, QR Algoritması, ARM Tabanlı, STM32 Cortex-M4, PCA, DCVA, Principal components analysis, Embedded systems, Ear Recognition, Discriminant Common Vector Approach, Jacobi Algorithm, QR Algorithm, ARM based on, STM32 Cortex-M4, PCA, DCVA