Evolutionary Feature Generation for Content-Based Audio Classification and Retrieval

Many commonly applied audio features suffer from certain limitations in describing the data content for classification and retrieval purposes. To remedy this drawback, in this paper we propose an evolutionary feature synthesis (EFS) technique, which is applied over traditional audio features to improve their data discrimination power. The underlying evolutionary optimization algorithm performs both feature selection and feature generation in an interleaved manner, optimizing also the dimensionality of the synthesized feature vector. The process is based on multi-dimensional particle swarm optimization (MD PSO) with two additional techniques: the fractional global best formation (FGBF) and simulated annealing (SA). The experimented classification and retrieval performances over a 16-class audio database show improvements of up to 11% when compared to the corresponding performances of the original features.