Music Thumbnailer: Visualizing Musical Pieces in Thumbnail Images Based on Acoustic Features

This paper presents a principled method calledMusicThumbnailer to transform musical pieces into visual thumbnail images based on acoustic features extracted from their audio signals. These thumbnails can help users immediately guess the musical contents of audio signals without trial listening. This method is consistent in ways that optimize thumbnails according to the characteristics of a target music collection. This means the appropriateness of transformation should be defined to eliminate ad hoc transformation rules. In this paper, we introduce three top-down criteria to improve memorability of thumbnails (generate gradations), deliver information more completely, and distinguish thumbnails more clearly. These criteria are mathematically implemented as minimization of brightness differences of adjacent pixels and maximization of brightness variances within and between thumbnails. The optimized parameters of a modified linear mapping model we assumed are obtained by minimizing a unified cost function based on the three criteria with a steepest descent method. Experimental results indicate that generated thumbnails can provide users with useful hints as to the musical contents of musical pieces.