Tez özetleri Astronomi ve Uzay Bilimleri Anabilim Dalı

Application of Hybrid Techniques to Face Recognition Problem

Human distinguishing problem is a long term study which is still developing with newly proposed methods. Till present day many methods that use distinguishing aspects of biological features are stated in order to solve this problem. Today some of these methods are used at customs, airports, banking operations and places that require security measures according to variety of their necessities. Generally, human recognition problem is processed with methods like iris recognition, fingerprint recognition, face recognition and hand/face vein recognition. By the way, in comparison to other stated methods face recognition method can gain necessary input data set easier, allowing it to be commonly used in subjects like personnel or criminal identification with security cameras.

In general definiton face recognition process is a union of sub-processes. These sub-processes are; extraction of face image from raw input image which is going to be searched (Face Detection), removing the physical noise which hardens recognition from extracted image, calculation of face image’s deterministic features (Feature Extraction) and comparison of calculated features with deterministic features of existing face data set (Classification).

In this study, Eigenfaces method which is based on PCA and one of well known face recognition methods in literature, is used in Feature Extraction. At following basic step, Classification, Neural Networks (NN) and Support Vector Machines (SVM) are used as 2 different classification methods and effects of using these hybrid techniques in face recognition process are examined. Results of hybrid techniques are compared with results of Eigenfaces method and due to increasing number of poses in training set better recognition rates with PCA-SVM hybrid technique are obtained against Eigenfaces technique. For Eigenfaces method difference in eigenspace dimension, for first hybrid technique (PCA-NN) difference in number of neurons in hidden layer and difference in training error and for second hybrid technique (PCA-SVM) difference of used Kernel Functions are studied with their effects on recognition performance.

In prepared application software, the input raw face image is converted into grayscale format and processed in face detection step. After these steps, user can choose to use only Eigenfaces method or other hybrid techniques (PCA-NN or PCA-SVM) for face recognition.

In first part of this study, general information about biometrical systems and face recognition are given. In second part commonly agreed methods in literature for face recognition, in third part information about this study’s main subjects Eigenfaces, Artificial Neural Networks and Support Vector Machines, in fourth part results that are obtained with application software, in fifth part comments on results and consideraton about future work are explained.