Environmental Noise Classification for Multimedia Libraries

In the modern information society, multimedia libraries are increasingly essential core components of the information systems managing our digital assets. The effective and efficient management of large amounts of multimedia information involves the extraction of relevant features from unstructured multimedia documents, images, videos, and sound recordings, as well as the organization, classification, and retrieval of these multimedia documents. A particularly important aspect is the opportunity to combine a variety of diverse features. In this paper we are interested in a feature rarely considered in such systems: the environmental noise. We design, implement, present, and evaluate an experimental multimedia library system for video clips and sound recordings in which scenes are indexed, classified and retrieved according to their environmental noise. Namely, after adequate training, the system is able distinguish between such scenes as traffic scenes, canteen scenes, and gunfight scenes, for instance. We show how we improved existing techniques for the classification of sound to reach an accuracy of up to 90% in the recognition of environmental noise.