Recognition of woodpecker calls using a convolutional deep neural network

Woodpeckers calls are readily recognizable on spectrograms and this opens the door for their identification from images through a convolutional Deep Neural Network (DNN). We built a dataset of 12154 images, half woodpecker calls (9 classes) and half noise, from Xeno-Canto and private recordings. We experimented with two approaches: a) training a small net (2 convolutional layers, 2 dense layers) from scratch using the theano framework and b) re-training legacy image nets (over 150 layers) using Pytorch. The larger nets successfully differentiate the calls from noise and identify them with an accuracy greater than 90%.