Mathematical Morphology Applied to Feature Extraction in Music Spectrograms

Mathematical Morphology has proven to be a powerful tool for extracting geometric information from greyscale images. In this paper, we demonstrate its application to spectrograms (two-dimensional greyscale images of sound) of music excerpts. The sounds of musical instruments exhibit particular shapes when represented as a spectrogram. These shapes are determined by the sound characteristics. In general, musical sounds contain three different components: the attack component, appearing as vertical lines; the sustain component, appearing as horizontal lines; and the stochastic component, appearing as a landscape of hills and holes. In this paper we propose a pipeline of morphological operators to separate these three components. This separation allows us to build a new sound similar to the input one.