COMPACT BILINEAR DEEP FEATURES FOR ENVIRONMENTAL SOUND RECOGNITION

Abstract

Environmental sound recognition (ESR) has extensive various civilian and military applications. Existing ESR methods generally tackle this problem by employing various signal processing and machine learning methods. Herein, an ESR paradigm based on feature extraction from pre-trained deep convolutional neural networks (CNN), the derivation of higher-order statistics by compact bilinear pooling and normalisation. In particular, we consider two deep ImageNet architectures for deep feature extraction, and the Random Maclaurin (RM) to produce the compact bilinear features. A support vector machine (SVM) with homogeneous mapping is used in the classification stage. Two publicly available environmental sound datasets are used to verify the efficacy of the approach namely, ESC-50 and ESC-10. We compare the proposed method with various previous state-of-the-art methods. Presented results indicate the suitability of the higher-order statistics of DEEP SPECTRUM representations for ESR classification tasks.