One-dimensional local descriptors for signal feature representation

Abstract

Feature representation, which plays an important role in several signal analysis tasks, is the process of extracting discriminative information from the signals. For example, feature representation of physiological signals such as electroencephalogram (EEG), electrocardiogram (ECG) and electromyogram (EMG) is an integral part of the automated diagnostic systems. In addition, feature representation of speech signals is an integral part of speech classification problems such as voiced-silence-unvoiced speech classification and gender identification. This thesis aims to develop novel and effective techniques for signal feature representation.Two-dimensional (2D) local descriptors are quite effective for image representation and have been explored for various image analysis tasks, including face recognition, gender recognition and texture classification. Motivated by the success of 2D-local descriptors, in this thesis, we focus our efforts on the development of effective one-dimensional (1D) local descriptors for signal feature representation. In addition, this thesis presents novel signal classification approaches using the proposed 1D local descriptors. The contributions of this thesis can be segregated into six parts. In the first two works, novel local descriptors for feature representation of EEG signals have been developed. Specifically, 1D-Gabor magnitude based local binary patterns, generalized multi-level local patterns (GMLP) and histogram of local variance for feature representation of EEG signals are proposed and subsequently, novel approaches for classification of epileptic EEG signals have been developed. In the third and fourth works, we introduce 1D local descriptors with reduced feature vector length for feature representation of ECG signals. In particular, we propose1D-local binary patterns based composite feature set for detecting ECG changes in epileptic patients and symmetrically weighted local binary patterns for obstructive sleep apnea detection. In fifth work, we propose two 1D-phase descriptors namely, 1D-local phase patterns and 1D-local phase quantization for detecting changes in Gabor phase response and demonstrate their relevance to OSA detection. Finally, a noise robust voiced and non-voiced speech classification approach using empirical wavelet transform and GMLP has been developed.This thesis experimentally demonstrates that the proposed 1D local descriptors are very effective in feature representation of EEG, ECG and speech signals. It also shows that 1D local descriptor-based classification approaches developed in this study are very efficient for signal classification. More importantly, most of the proposed approaches achieve the state-of-the-art performance.