Spectrogram image encoding based on dynamic Hilbert curve routing

In this paper we propose an image-based biological classification system that can identify different creatures via their sounds. The overall system involves the relative spectral transform-perceptual linear prediction for spectrogram image extraction, cosine similarity measure for feature matching, dynamic Hilbert curve for spectrogram routing, and Gaussian mixture model for 1-D spectrogram classification. As an example of our approach, results for honk, dolphin, and whale classification are presented. This method works well on a wide variety of bio-sounds, especially for the highly self-repeated ones. Applications of this approach include biological signal analysis and spectrogram library establishment.