An investigation of fusion strategies for boosting pig cough sound recognition.

• We proposed two fusion strategies to promote the accuracy of pig cough recognition. • Classifier fusion was proven to outperform feature fusion. • The proposed method achieved an accuracy of 99.20%. The recognition of pig cough sounds is an effective way to monitor pig respiratory diseases which seriously affect healthy pig breeding. Due to the complexity of the pig-housing environment, achieving high precision cough recognition by relying only on a single feature or classifier is challenging. Therefore, in this study we investigated two fusion strategies, namely feature fusion and classifier fusion, to boost classification accuracy. For feature fusion, we improved the previously proposed feature fusion algorithm and selected better acoustic and image features for fusion. We also proposed a novel classifier fusion algorithm. In the algorithm, the support vector machine (SVM) classifiers trained by the acoustic features and deep features were fused by soft voting for pig cough prediction. The sound data collected in the pig barn were used to validate the proposed methods. Our methods achieved a substantial classification rate of 97.47% and 99.20% for feature fusion and classifier fusion, respectively. The results demonstrate that our proposed fusion strategies can significantly improve the recognition accuracy of pig cough sounds. [ABSTRACT FROM AUTHOR]