Mosaic-Based Deep Learning Approach for Bird Species Identification and Behavior Recognition
- TETERJA, DMITRIJ
- José García Rodríguez Codirector/a
- Jorge Azorín López Codirector
Universidad de defensa: Universitat d'Alacant / Universidad de Alicante
Fecha de defensa: 19 de julio de 2024
- Manuel Pérez Malumbres Presidente/a
- Andrés Fuster Guilló Secretario
- Alexandra Psarrou Vocal
Tipo: Tesis
Resumen
Embracing life’s diversity across various biological levels, including within and between species and ecosystems, biodiversity offers ecological, economic, and cultural benefits. Conservation, essential for responsible natural asset management, ensures ecosystem resilience. Protected Natural Areas (PNAs) play a critical role in sustaining biodiversity, and providing environmental, social, and economic benefits. Our innovative approach integrates comprehensive PNA perspectives, employing data-centric technology to monitor biodiversity composition and animal species behaviour. The focal point of our thesis is addressing challenges in avian species identification and behaviour. Our mosaic-based method, which combines image mosaicing and deep learning, offers a reliable and accurate solution for bird identification and behaviour recognition, encompassing even chicken behaviour. This technology, implemented on edge computing devices, contributes to avian population monitoring and conservation efforts. The significance of knowing the species present, their abundances, and their behaviour within the parks cannot be overstated for effective management. This thesis provides a valuable tool for park managers by offering ready-to-use information. This includes the ability to identify bird species, monitor the activities of birds in specific areas, and detect potential issues with birds, enabling proactive problem-solving strategies. This could be achieved by deploying a new state-of-the-art dataset with annotated behaviour videos by skilled ornithologists, so this knowledge can be used with automated systems based on deep neural networks (DNNs). The implementation of our innovative approach enhances the efficiency of classical DNNs, adapting input data without altering the model architecture. It is compatible with transfer learning (TL) and is well-suited for scenarios involving both bird species identification and bird behaviour recognition, in addition to chicken behaviour recognition, on edge computing devices, accommodating their limited resource availability.