Detection and classification of non-stationary signals using sparse representations in adaptive dictionaries

Thumbnail Image


Publication or External Link






Automatic classification of non-stationary radio frequency (RF) signals is of particular interest in persistent surveillance and remote sensing applications. Such signals are often acquired in noisy, cluttered environments, and may be characterized by complex or unknown analytical models, making feature extraction and classification difficult. This thesis proposes an adaptive classification approach for poorly characterized targets and backgrounds based on sparse representations in non-analytical dictionaries learned from data. Conventional analytical orthogonal dictionaries, e.g., Short Time Fourier and Wavelet Transforms, can be suboptimal for classification of non-stationary signals, as they provide a rigid tiling of the time-frequency space, and are not specifically designed for a particular signal class. They generally do not lead to sparse decompositions (i.e., with very few non-zero coefficients), and use in classification requires separate feature selection algorithms. Pursuit-type decompositions in analytical overcomplete (non-orthogonal) dictionaries yield sparse representations, by design, and work well for signals that are similar to the dictionary elements. The pursuit search, however, has a high computational cost, and the method can perform poorly in the presence of realistic noise and clutter. One such overcomplete analytical dictionary method is also analyzed in this thesis for comparative purposes. The main thrust of the thesis is learning discriminative RF dictionaries directly from data, without relying on analytical constraints or additional knowledge about the signal characteristics. A pursuit search is used over the learned dictionaries to generate sparse classification features in order to identify time windows that contain a target pulse. Two state-of-the-art dictionary learning methods are compared, the K-SVD algorithm and Hebbian learning, in terms of their classification performance as a function of dictionary training parameters. Additionally, a novel hybrid dictionary algorithm is introduced, demonstrating better performance and higher robustness to noise. The issue of dictionary dimensionality is explored and this thesis demonstrates that undercomplete learned dictionaries are suitable for non-stationary RF classification. Results on simulated data sets with varying background clutter and noise levels are presented. Lastly, unsupervised classification with undercomplete learned dictionaries is also demonstrated in satellite imagery analysis.