Cochleagram image feature for improved robustness in sound recognition

In this paper, we use the cochleagram image of sound signals for time-frequency analysis and feature extraction, instead of the conventional spectrogram image, in an audio surveillance application. The signal is firstly passed through a gammatone filter which models the auditory filters in the human cochlea. The filtered signal is then divided into small windows and the energy in each window is added and normalized which gives the intensity values of the cochleagram image. We then divide the cochleagram image into blocks and extract central moments as features. Using two feature vector representation methods, the results show significant improvement in overall classification accuracy when compared to results from literature employing similar feature extraction and representation techniques but using spectrogram images. The most improved results were at low signal-to-noise ratios.