Resumen de Dissimilarity-based multiple instance classification and dictionary learning for bioacoustic signal recognition
In this thesis, two promising and actively researched fields from pattern recognition (PR) and digital signal processing (DSP) are studied, adapted and applied for the automated recognition of bioacoustic signals: (i) learning from weakly-labeled data, and (ii) dictionary-based decomposition. The document begins with an overview of the current methods and techniques applied for the automated recognition of bioacoustic signals, and an analysis of the impact of this technology at global and local scales. This is followed by a detailed description of my research on studying two approaches from the above-mentioned fields, multiple instance learning (MIL) and dictionary learning (DL), as solutions to particular challenges in bioacoustic data analysis. The most relevant contributions and findings of this thesis are the following ones: 1) the proposal of an unsupervised recording segmentation method of audio birdsong recordings that improves species classification with the benefit of an easier implementation since no manual handling of recordings is required; 2) the confirmation that, in the analyzed audio datasets, appropriate dissimilarity measures are those which capture most of the overall differences between bags, such as the modified Hausdorff distance and the mean minimum distance; 3) the adoption of dissimilarity adaptation techniques for the enhancement of dissimilarity-based multiple instance classification, along with the potential further enhancement of the classification performance by building dissimilarity spaces and increasing training set sizes; 4) the proposal of a framework for solving MIL problems by using the one nearest neighbor (1-NN) classifier; 5) a novel convolutive DL method for learning a representative dictionary from a collection of multiple-bird audio recordings; 6) such a DL method is successfully applied to spectrogram denoising and species classification; and, 7) an efficient online version of the DL method that outperforms other state-of-the-art batch and online methods, in both, computational cost and quality of the discovered patterns
