Publicação
Usage of convolutional neural networks for identifying marine mammal individuals
| Resumo: | Identifying marine mammals is a common practice performed by whale-watching crew members. Typically, an experienced marine ecologist is the one who can identify not just the taxa, but also the individual. This process is however always done in the aftermath of data sampling, where the goal is to use photo identification and match the dorsal fins of individuals spotted at the different spatio-temporal scales. This dissertation provides the pipeline and addresses the chal lenges in the usage of Convolutional Neural Networks (CNNs) to discriminate marine mammal individuals, in this case (pilot whales) based on the dorsal fins. The dissertation uses as input the 1138 images dataset containing over 856 individuals, and through three experiments addresses the issues when discriminating such a high number of classes. In the first experiment, the dissertation studies the role of synthetic data augmentation in boosting model performance. In second, the dissertation benchmarks the existing state-of-the-art convolutional neural network architectures. In third, the dissertation focuses on discriminating other features from dorsal fins to identify indi viduals (scratches, nicks, roundness, wideness). The dissertation outlines the issues and proposes the guidelines for the next effort in discriminating marine mammal individuals. |
|---|---|
| Autores principais: | Gouveia, Jorge Miguel Vieira |
| Assunto: | Convolutional neural networks Deep Learning Marine mammals Photo identification Object detection Image classification Redes Neuronais Convolucionais Deep Learning (DL) Mamíferos marinhos Identificação por fotos Deteção de objectos Classificação de imagens Engenharia Informática . Faculdade de Ciências Exatas e da Engenharia |
| Ano: | 2023 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade da Madeira |
| Idioma: | inglês |
| Origem: | DigitUMa - Repositório da Universidade da Madeira |
| Resumo: | Identifying marine mammals is a common practice performed by whale-watching crew members. Typically, an experienced marine ecologist is the one who can identify not just the taxa, but also the individual. This process is however always done in the aftermath of data sampling, where the goal is to use photo identification and match the dorsal fins of individuals spotted at the different spatio-temporal scales. This dissertation provides the pipeline and addresses the chal lenges in the usage of Convolutional Neural Networks (CNNs) to discriminate marine mammal individuals, in this case (pilot whales) based on the dorsal fins. The dissertation uses as input the 1138 images dataset containing over 856 individuals, and through three experiments addresses the issues when discriminating such a high number of classes. In the first experiment, the dissertation studies the role of synthetic data augmentation in boosting model performance. In second, the dissertation benchmarks the existing state-of-the-art convolutional neural network architectures. In third, the dissertation focuses on discriminating other features from dorsal fins to identify indi viduals (scratches, nicks, roundness, wideness). The dissertation outlines the issues and proposes the guidelines for the next effort in discriminating marine mammal individuals. |
|---|