Urban acoustic classification based on deep feature transfer learning

Abstract

Urban acoustic classification (UAC) plays a vital role in smart city engineering, urban security, noise pollution analysis, etc. Convolutional neural networks (CNNs) based feature transfer learning have been shown competitive performance in many applications but little attention has been paid to UAC. In this study, a novel UAC algorithm exploiting the deep CNNs based feature transfer learning and the deep belief net (DBN) based classification is developed. The spectrogram is first adopted for the urban acoustic stream representation. Then, three deep CNNs pre-trained on ImageNet database are applied as feature extractors. The extracted features are concatenated and fed to a DBN for classifier learning. To achieve a good generalization performance, three restricted boltzmann machines (RBM) trained by the contrastive divergence algorithm (CD) followed by a back-propagation (BP) based fine parameter tuning is adopted in DBN. The proposed UAC is evaluated on a real acoustic database, including 11 categories of acoustic signals recorded from the urban environment. Performance comparisons to many state-of-the-art algorithms are presented to demonstrate the superiority of the proposed method.