D etonation C lassification from A coustic S ignature with the R estricted B oltzmann M achine

摘要: We compare the recently proposed Discriminative Restricted Boltzmann Machine (DRBM) to classical Support Vector (SVM) on a challenging classification task consisting in identifying weapon classes from audio signals. The three considered this work (mortar, rocket, and rocket-propelled grenade), are difficult reliably classify with standard techniques because they tend have similar acoustic signatures. In addition, specificities of data available study make it rigorously classifiers, we address methodological issues arising situation. Experiments show good accuracy that could these suitable for fielding autonomous devices. DRBMs appear yield better than SVMs, less sensitive choice signal preprocessing model hyperparameters. This last property is especially appealing such where lack makes validation difficult. (10Roughly speaking, number DOF regression residuals computed as observations training set minus parameters part model. © 2012 Wiley Periodicals, Inc.)