Recognition of woodpecker calls using a convolutional deep neural network

Recent research on recognizing bird calls has changed its focus from audio features to images. Indeed, many bird calls, including the most iconic woodpecker calls, can be visually identified on spectrograms. The intention is then to feed the images to a convolutional Deep Neural Network (DNN) and let it detect the important visual patterns and recognize the bird species. We thus designed a network with two convolutional layers for the image analysis and two dense layers to converge towards a diagnosis. The base images are standardized to 1 sec in duration and target the dominant frequency range for woodpeckers, 1--3.5 kHz. This is shorter than the typical woodpecker call, hence the images only capture high-energy moments within the calls. This approach focuses on recognizing syllables first, leaving call structure as a next step. The dataset is derived from Xeno-Canto and private recordings and comprises 12154 images, half of them woodpecker calls (9 classes). The perspicacity of the network is improved using dropout and data augmentation. We also expect the DNN to benefit from being co-trained on broader tasks such as distinguishing birdcalls from other sounds. However, early results speak against the straightforwardness of DNNs. The net assigns the highest probability to the correct class in only 36.4% of the cases. Considering the top three suggestions, the accuracy reaches 77.8%, standing above 90% for four calls (Jynx torquilla, D. medius, D.martius flight call, Dryobates minor) while Picus canus plummets at 14.3%.