An Evaluation of Algorithms for SingleEcho Biosonar Target Classification.pdfVIP

PROCEEDINGS OF WORLD ACADEMY OF SCIENCE, ENGINEERING AND TECHNOLOGY VOLUME 1 JANUARY 2005 ISSN 1307-6884 An Evaluation of Algorithms for Single-Echo Biosonar Target Classification Turgay Temel and John Hallam Abstract— A recent neuro-spiking coding scheme for feature II. BIOLOGICAL SIGNAL PROCESSING FOR ECHOES, FEATURE extraction from biosonar echoes of various plants is examined with a EXTRACTION AND PROPOSED PREPROCESSING SCHEME variety of stochastic classifiers. Feature vectors derived are employed The echo preprocessing operation consists of two stages: in well-known stochastic classifiers, including nearest-neighborhood, single Gaussian and a Gaussian mixture with EM optimization. cochlear filtering and coding, [5]. In the first stage, the Classifiers’ performances are evaluated by using cross-validation and waveform is passed through a fourth order gammatone filter bootstrapping techniques. It is shown that the various classifers with center frequency f c and –3 dB quality factor Q-3dB for perform equivalently and that the modified preprocessing modeling cochlear filters. This stage is followed by envelope configuration yields considerably improved results. extraction performed as half-wave rectification and low-pass filtering (LPF) with Keywords— Classification, neuro-spike coding, non-parametric model, parametric model, Gaussian mixture, EM algorithm.