Investigation of acoustic and visual features for pig cough classification.

The precise detection of pig cough is a crucial step for establishing an early warning system for pig respiratory diseases. With regard to high precision pig cough recognition, feature extraction and selection are of importance. However, few studies have investigated both acoustic and visual features of pig vocalisations as input features. In this paper, we proposed a novel feature fusion method which fusing acoustic and visual features to achieve an enhanced pig cough recognition rate. We firstly extracted acoustic features from audio signals, including root-mean-square energy (RMS), mel-frequency cepstral coefficients (MFCCs), zero-crossing rates (ZCRs), spectral centroid, spectral roll-off, spectral flatness, spectral bandwidth and chroma. Then, constant-Q transform (CQT) spectrograms were employed to extract visual features involving local binary pattern (LBP) and histogram of gradient (HOG). Subsequently, a hybrid feature set was created by combining acoustic and visual features. In this stage, Pearson correlation coefficient (PCC), recursive feature elimination based on random forest (RF-RFE) and principal component analysis (PCA) were exploited for dimensionality reduction. Finally, support vector machine (SVM), random forest (RF) and k-nearest neighbours (KNN) classifiers were used to conduct a performance evaluation. It is shown that the fused acoustic features (Acoustic) combined with LBP and HOG (A-LH) achieved 96.45% pig cough accuracy. The results reveal that the fusion feature set outperforms acoustic and visual features alone. [Display omitted] • Both acoustic and visual features were investigated for pig cough. • A hybrid feature vector created by early feature fusion enhanced recognition performance of pig cough. • Feature selection and extraction were employed to choose the optimal feature set. • The proposed method reached an accuracy of 96.45%. [ABSTRACT FROM AUTHOR]