LBP-based bird sound classification using improved feature selection algorithm.

Local binary pattern (LBP)-based features for bird sound classification were investigated in this study, including both one-dimensional (LBP-1D) and two-dimensional (LBP-2D) local binary patterns. Specifically, the discrete wavelet transform was first used as a pooling method to generate multi-level features in both time (LBP-1D-T) and frequency domain (LBP-1D-F) signals. To obtain richer time–frequency information of bird sounds, uniform patterns (LBP-2D) were extracted from the log-scaled Mel spectrogram. To fully exploit the complementarity of different LBP features, a hybrid fusion method was implemented. Next, neighborhood component analysis (NCA) was employed as a feature selection method to remove redundant information in the fused feature set. In order to reduce the running time of NCA and improve the classification accuracy, an improved feature selection method (DSNCA) was proposed. Finally, two machine learning algorithms: K-nearest neighbor and support vector machine were used for classification. Experimental results on 43 bird species of North American wood-warblers indicated that LBP-2D achieved a higher balanced-accuracy than LBP-1D-T and LBP-1D-F (86.33%, 81.05% and 70.02%, respectively). In addition, the highest classification accuracy was up to 88.70%, using hybrid fusion. [ABSTRACT FROM AUTHOR]