Identificação de espécies de pássaros utilizando espectrogramas e dissimilaridade

• Ano de defesa: 2017
• Autor(a) principal: Zottesso, Rafael Henrique Dalegrave
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Estadual de Maringá; Brasil; Departamento de Informática; Programa de Pós-Graduação em Ciência da Computação; UEM; Maringá, PR; Centro de Tecnologia
• Palavras-chave em Português: Sistemas de reconhecimento de padrões; Pássaros; Identificação de espécies de pássaros; Padrões (informática); Sistemas de reconhecimento; Espectrogramas; Processamento de sinais; Dissimilaridade; Brasil.; Bird species identification; Dissimilarity; Pattern recognition; Spectrogram and texture; Signal processing; Brazil.; Ciências Exatas e da Terra; Ciência da Computação
• Link de acesso: http://repositorio.uem.br:8080/jspui/handle/1/2528

This work presents a proposal for bird species identification using spectrograms and dissimilarity approach, in a database with a high number of species. The database is composed by audio recordings pre-selected by the LifeClef 2015 Bird Task that can be easily found on Xeno-canto website. In this work, eight subsets of data were created from this database, in order to diversify the amount of species and the duration of the audio samples in our tests, selecting only bird songs and discarding the bird calls. All audio samples used were preprocessed to reduce the impact of noise, removing other sources of sounds, and to detect points of interest with greatest relevance. Then, to transform the audio samples in images, there was a task to generate spectrograms, which went through the zoning process in order to enhance local information from each region created. Three texture descriptors were used to perform feature extraction: Local Binary Pattern (LBP), Local Phase Quantization (LPQ) and Robust Local Binary Pattern (RLBP). In the model-dependent approach these features were directly classified. In the dissimilarity approach it was needed to compute dissimilarity vectors (positive and negative), to further apply the classification scheme. Both cases used a classification through the SVM, allowing the application of combination rules to reach a final decision. After a series of experiments, it was perceived that the dissimilarity approach presented superior results in relation to a model-dependent approach and the literature.