Time-frequency features for pattern recognition using high-resolution TFDs

This paper presents a tutorial review of recent advances in the field of time-frequency ( t , f ) signal processing with focus on exploiting ( t , f ) image feature information using pattern recognition techniques for detection and classification applications. This is achieved by (1) revisiting and streamlining the design of high-resolution quadratic time frequency distributions (TFDs) so as to produce adequate ( t , f ) images, (2) using image enhancement techniques to improve the resolution of TFDs, and (3) defining new ( t , f ) features such as ( t , f ) flatness and ( t , f ) entropy by extending time-domain or frequency-domain features. Comparative results indicate that the new ( t , f ) features give better performance as compared to time-only or frequency-only features for the detection of abnormalities in newborn EEG signals. Defining high-resolution TFDs for the extraction of new ( t , f ) features further improves performance. The findings are corroborated by new experimental results, theoretical derivations and conceptual insights. A streamlined methodology for designing high resolution quadratic TFDs using separable, directional and adaptive kernels.A formulation of new (t, f) features by translation from time-domain only features or frequency-domain only features.A review of (t, f) image processing techniques for resolution enhancement, de-noising and improved classification.A review of multi-component IF estimation techniques as a performance criterion to compare time-frequency distributions.Experiments that illustrate the above points in EEG seizure detection and classification using a large medical database.