الفهرس 
Thesis Organization
Related WorkOnly 14 pages are availabe for public view
 |  | from | 108 |  |  |
| | | from | 108 | | |
Abstract
Arabic Sign Language ArSL is widely used in the Arabian countries due to its facilities to communicate with Hearing Impaired HI individuals. ArSL Recognition becomes vital for communication among HI persons and many technologies were employed to serve the recognition purpose and a lot of researches had been conducted to study either static or dynamic gestures.
Sign Language Recognition breaks the barrier between deaf and normal people. As the only way to communicate with HI persons is acting the meaning of words by hands and body, this way will deliver the meaning for either the normal or HI people. After increasing of Sign Language usage and importance, translation system for such languages become an essential need and also the requirement for a standard dictionary for ArSL.
The main purpose of the thesis is to develop an Arabic Sign Language Recognition ArSLR system which translates ArSL to Arabic words using deep techniques. The dataset which is used in our system is taken from the standard Arabic Sign Language dictionary published in 2005 [1]. The words are represented in 40 postures those were captured and collected from different signers and in different environments. The dataset was recorded with a digital camera from different angles.
The machine learning hand gesturing model is developed to recognize ArSL using two integrated deep models; Convolutional Neural Network CNN and Deep Belief Network DBN, as deep techniques have recently shown a significant gain for building hieratical architecture for unlabeled data to learn, make usage of the self-learning attitude, and modeling the feature information learned from each network.
IV
The system was tested with 25 words that are represented in 40 postures. The data is represented by images that are taken by a digital camera with resolution 1280x720. The system achieves an average accuracy 90% on the observed trained data and 86% on the unseen postures. It also achieves an average accuracy of 77% for gestures that are represented as sequences of trained postures. However, misclassification of postures in the sequence has been observed due to images closeness in multiple words’ sequence which is lead to confusion between more than one gesture that happens to result in lower accuracy rate. The system is tested on hardware with Intel Core i5-2520M CPU, 2.5 GHz, 3.78 GB memory RAM, Intel HD Graphics 3000, and Windows8 64 bit operating system.
After the new generation of input sensors, ArSL has to make use of the sensors’ advantages. New modern sensors are used and tested for the recognition purpose. Also, their effects on the accuracy rate are observed to achieve the best model that results in high accuracy rate with low computational power. A system based on Leap Motion data input is developed to show the impact of the new data format. Also, the integration between Leap Motion sensors with number of other input sensors are fed to a system to maximize each sensor’s advantage and increase the recognition accuracy. And one additional system is developed to include a new features set other than hand gestures to enlarge the features space used and so the accuracy obtained.