Audio-only Bird Species Automated Identification Method with Limited Training Data Based on Multi-Channel Deep Convolutional Neural Networks

XIE Jiang-jian DING Chang-qing; LI Wen-bin; CAI Cheng-hao (1 School of Technology,Beijing Forestry University, Beijing, 100083, P. R. China. 2 School of Nature Conservation,Beijing Forestry University, Beijing, 100083, P. R. China.) Abstract 1. Deep convolutional neural networks (DCNN) have achieved breakthrough performance on bird species identification tasks based on spectrogram features, but a huge number of labeled samples are needed to train an excellent DCNN model. However, it is difficult to collect enough samples for certain bird species. For practical uses of bird species identification, it is significant to study a method which only requires a small sample set. 2. Transfer learning is an available solution to train deep learning models with limited samples. Based on the parameter transfer mode, we design a bird species identification model that uses the VGG-16 model (pretrained on ImageNet) for feature extraction, then a classifier consisting of two fully-connected hidden layers and a Softmax layer is attached. We take the vocalization signals of eighteen bird species which were recorded in Beijing Song-Shan National Nature Reserve as an example, comparing the performance of the transfer learning model with the original VGG16 model. The results show that the former has higher train efficiency, but lower mean average precisions (MAP). 3.To improve the MAP of the transfer learning model, we investigate two fusion modes to form multi-channel identification models. Then we evaluate the models on our own sample sets. We find that the result fusion mode outperforms the feature fusion mode, and the best MAP reaches 0.9998. The number of model parameters is 13110, which is only 0.0082% of the VGG16 model. Also, the size demand of sample is decreased. 4. The type and duration of spectrogram may affect the performance of identification. We choose three kinds of time frequency transformation methods, including Short Time Fourier Transform, Mel-frequency Cepstrum Transform and Chirplet Transform, to calculate the spectrogram. Then the spectrogram is segmented to the duration of 100ms, 300ms and 500ms. Chirplet spectrogram improves training efficiency and the MAP, which is more suitable for the feature representation of bird vocalization. We choose three durations, 100ms, 300ms and 500ms for comparison, the result reveals that the 300ms duration is the best. The duration should be determined based on the time domain characteristic of bird vocalization. Tweetable Abstract: Based on the transfer learning, we design a bird species identification model that uses the VGG-16 model (pretrained on ImageNet) for feature extraction, then a classifier consisting of two fully-connected hidden layers and a Softmax layer is attached. We compare the performance of the proposed model with the original VGG16 model. The results show that the former has higher train efficiency, but lower mean average precisions(MAP). To improve the MAP of the proposed model, we investigate the result fusion mode to form multi-channel identification model, the best MAP reaches 0.9998. The number of model parameters is 13110, which is only 0.0082% of the VGG16 model. Also, the size demand of sample is decreased.