In this paper, we utilize time-frequency image representations of sound signals for feature extraction in an audio surveillance application. Starting with the conventional spectrogram images, we consider a new feature which is based on image texture analysis. It utilizes the gray-level co-occurrence matrix, which captures the distribution of co-occurring values at a given offset. We refer this as the spectrogram image texture feature. Texture analysis is carried out in subbands and experimented on a sound database containing ten classes with each sound class containing multiple subclasses. The proposed feature was seen to be more noise robust than two commonly used cepstral features, mel-frequency cepstral coefficients and gammatone cepstral coefficients, the spectrogram image feature (SIF), where central moments are extracted as features, and a variation of SIF with reduced feature dimension. In addition, we achieved a significant improvement in classification accuracy for the three time-frequency image features by utilizing a gammatone filter-based time-frequency image, referred as cochleagram image, for feature extraction instead of the spectrogram image. A combination of cepstral and cochleagram image features also gave improvement in the classification performance.
|Number of pages||11|
|Journal||IEEE Transactions on Information Forensics and Security|
|Publication status||Published - 1 Dec 2015|
- Audio surveillance
- gammatone filter
- gray-level co-occurrence matrix
- support vector machines