Traditional and modern processing of digital signals and images for the classification of birds from singing

Abstract

Avitourism depends on the understanding of birds and the intention of birdwatchers to see or hear a specific species. Increase of this activity strengthens the economy of communities and helps finance biodiversity conservation projects. Technological products that incorporate traditional and modern tools for signal and image processing facilitate the tracking, classification and observation of birds. This article has two approaches. The first one proposes and evaluates a lightweight classification model that uses feature vector extracted from the bird's song spectrum and is based on comparison of Euclidean distance between sample features and a set vector by species. The second approach adapts and evaluates convolutional neural network architectures for bird classification using the spectrogram of the bird's song. The methodology applied in both approaches consists of: pre-processing, feature extraction, classification and evaluation metrics. The main results are the feasibility of the proposed lightweight classification model with an accuracy of 0.8 and a loss of 2.32, and the feasibility of using convolutional neural networks with an accuracy above 0.9 and a loss of less than 1, in the ResNet50, VGG19, and InceptionV3 architectures, this using as a minimum 30 spectrograms per species during training. It is concluded that the model that best meets the needs of fewer samples and less computational resources required for training is ResNet50. Additionally, it is discussed to combine the two approaches in a hierarchical and hybrid classification model that allows to introduce in the reduction and classification layers of the neural network features of another type and of other sources.