Yırtıcı, Tolga <br/><br/>
IMPROVED TURKISH SIGN LANGUAGE RECOGNITION USING CONVOLUTIONAL NEURAL NETWORKS / 
TOLGA YIRTICI; SUPERVISOR: ASST. PROF. DR. KAMİL YURTKAN 
 - 114 sheets;  31 cm.  Includes CD 
<br/><br/>Thesis (PHD) - Cyprus International University. Institute of Graduate Studies and Research Computer Engineering Department 
<br/><br/>Includes bibliography (sheets 106-114) 
<br/><br/> ABSTRACT<br/>This thesis started with idea of creating a robust system for Turkish Sign Language<br/>recognition. Fingerspelling of Turkish Sign Language alphabet is chosen for these<br/>purposes. Turkish Sign Language alphabet consists of 29 letters just like in speaking<br/>language. Alphabet letters can be used to form a word. Two different systems designed<br/>in this manner, one with a region-based object detection method and other is an<br/>information content-based feature selection. The designed systems are employed with<br/>AlexNet architecture using transfer learning. AlexNet is a pre-trained Convolutional<br/>Neural Network that utilized for classification problems. The novel object detection<br/>method is tested with three different algorithms and achieved the best result of 0.997<br/>mean Average Precision and 0.9982 accuracy rate. The information content-based<br/>feature selection method with employed the same AlexNet architecture, used a novel<br/>feature selection algorithm and achieved more than 80% accuracy rate. Both of the<br/>systems are trained and tested on the dataset created for this study in a studio. 
<br/><br/><label>Subjects--Topical Terms: </label><br/>Object-oriented methods (Computer science)--Dissertations, Academic<br/>Transfer learning (Machine learning)--Dissertations, Academic