Reconocimiento de canto de aves basado en el análisis de componentes principales del espectrograma

Abstract

En el presente artículo se propone un método de reconocimiento automático de canto de aves. Este método está basado en el análisis de componentes principales del espectrograma de los cantos, implementado mediante la técnica de eigenfaces. Para la preparación de la base de datos se implementó un método de detección de actividad acústica por medio de un filtro de media móvil. Se construyó una base de datos con tres cantos diferentes de tres aves: Formicivora Grisea, Harpia Harpyja y Reinita Protonotaria. Las pruebas realizadas muestran la efectividad del método propuesto, obteniéndose una tasa de reconocimiento de 100%.

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

“…For the extraction of characteristics, reduction of dimensions, and classification of the edge, methods have been used: simple or complex, traditional or modern. Some of the tools used are mentioned below: Mean, Deviation, Variance, Covariance, Correlation, Euclidean Distance, Spectrographic Cross-Correlation (SPCC) (Khanna et al, 1997;Chen et al, 2020), Discriminant Function Analysis (DFA) (Chen et al, 2020), Linear Discriminant Analysis (LDA ) (Hsu et al, 2018), Principal Component Analysis (PCA) (Hsu et al, 2018;González et al, 2019), Gaussian Mixture Model (GMM) Algorithm (Lee et al, 2008), Vector Quantization (VQ) (Lee et al, 2008), Dynamic Time Warping (DTW) (Kogan and Margoliash, 1998), Hidden Markov Model (HMM) (Kogan and Margoliash, 1998), Support Vector Machines (SVM) (Fagerlund, 2007), Decision Tree (Chen and Li, 2013), Artificial Neural Network (ANN) (Sukri et al, 2020) ANN of Self Organizing Map (SOM) types (Tanttu et al, 2003), Recurrent Neural Networks (RNN) (Wan et al, 2022), Convolutional Neural Networks (CNNs) (Zhang et al, 2019), among others.…”

“…to reduce the dimensionality of the features, and finally a simple distancebased classifier is used. In González et al (2019) the Spectrogram of the song is obtained, to these the Eigenface technique is applied based on the PCA, to obtain the eigenvectors of the covariance matrix that best represents the spectrograms, the vectors that present the highest energy are taken, and the song is classified by comparing the training and test vectors. In Lee et al (2008) GMM and VQ are used to represent the MFCC of the songs in different bells (cluster) or regions (vectors), to later estimate the mean of the GMM or centroids of VQ and form the prototype vectors of a certain species, and from these, recognize the vector that best fits the song of the species to be classified using K-means.…”

“…The differences are in the pre-processing and feature extraction techniques. Kogan and Margoliash (1998) and Hsu et al (2018) used tools like DWT; Khanna et al (1997) and González et al (2019) employed spectrograms, Lee et al (2008), Zhang et al (2019) Wang et al (2022) use MFCC.…”

“…For feature extraction, we collected and analyzed the mountains and peaks of the DFT. We deliberately avoided advanced statistical techniques such as PCA, as used by González et al (2019), LDA, as used by Hsu et al (2018), or DFA, as used in the study by Chen et al (2020), to reduce the computational cost.…”