| المستخلص: |
Poor alertness is a major problem encountered by many individuals nowadays. This problem is critical in situations where the safety and health of the individual are jeopardized. Not only safety and health are affected, also financial losses are incurred by the society, companies, and governments. As loss of alertness is not easily recognized by individuals immersed in their daily activities, several methods were proposed to automatically detect shifts in alertness states using physiological and behavioral cues. Sparse representation-based classification is a newly developed method which demonstrated its efficiency in a number of applications [4, 5]. The core idea is to classify a sparse coefficient vector obtained from representing a feature vector (extracted from the EEG signal) using an over-complete dictionary. The sparse representation based classifier comprises three parts: dictionary selection/learning, sparse coding (sparse coefficient vector computation) and class assignment. A number of measures of sparsity, such as -(pseudo) norm, -norm, have been proposed in the literature. One of the latest proposed measures is the Gini index. This index was shown to satisfy a number of desirable properties. Its usage was, so far, was mostly restricted to sparse signal representation. The aim of this thesis is to investigate the potential use of Gini index in a sparse representation-based classification and compare its performance to those using - and -norms as measures of sparsity. More specifically, the classification problem is to automatically identify the human alertness state using features extracted from the electroencephalogram (EEG). The proposed method uses Gini index in all the three stages mentioned above (dictionary learning, sparse coding and class assignment). To the author's best knowledge, this is the first time such usage of Gini index has been used. The obtained result clearly show the superiority of Gini index-based method over those using -norm and -norm in classifying human alertness states.
|
