Multiple-instance multiple-label learning for the classification of frog calls with acoustic event detection

Frog call classification has received increasing attention due to its importance for ecosystem. Traditionally, the classification of frog calls is solved by means of the single-instance single-label classification classifier. However, since different frog species tend to call simultaneously, classifying frog calls becomes a multiple-instance multiple-label learning problem. In this paper, we propose a novel method for the classification of frog species using multiple-instance multiple-label (MIML) classifiers. To be specific, continuous recordings are first segmented into audio clips (10 seconds). For each audio clip, acoustic event detection is used to segment frog syllables. Then, three feature sets are extracted from each syllable: mask descriptor, profile statistics, and the combination of mask descriptor and profile statistics. Next, a bag generator is applied to those extracted features. Finally, three MIML classifiers, MIML-SVM, MIML-RBF, and MIML-kNN, are employed for tagging each audio clip with different frog species. Experimental results show that our proposed method can achieve high accuracy (81.8% true positive/negatives) for frog call classification