Percorrer por autor "Millan Batista, Andres Alejandro"
A mostrar 1 - 1 de 1
Resultados por página
Opções de ordenação
- Comparison of computer vision and deep learning techniques for identifying marine mammal individualsPublication . Millan Batista, Andres Alejandro; Radeta, Marko; Mendonça, FábioThis thesis addresses the difficulties of the process for identifying individual marine mammals, as well as their automation and considerations when working on this specific task. Although it has been established and evolving for more than 40 years, this process is traditionally performed using photo identification, a process that is time-consuming and relies on domain expertise for accurate annotation. This approach poses substantial hurdles, particularly since deep learning methods, which provide a promising alternative, require a large volume of annotated images to achieve high model accuracy. However, such data can be cumbersome to acquire for elusive individuals. To ad dress these limitations and solve the identification of marine mammal individuals, this dissertation develops and benchmarks diverse algorithms and pipelines to automate and facilitate their classi f ication. The first employs computer vision and data science techniques to process images without relying on manual annotation. The second leverages deep learning models that incorporate anno tated data for training, as well as image identification and object classification, with the intention of speeding up this process. Validation of these pipelines is performed by comparing the predic tions against ground-truth annotations, providing a measure of model performance. Additional tests assess the efficiency of pipelines in identifying previously unseen individuals, evaluating both the accuracy and time required to recognize them. These methodologies are thus a link between manual and automated identification, providing scalable solutions for marine mammal research and conservation efforts. Every step achieved in this process allows the researchers to reduce the working time when it comes to the whole process, allowing them to focus on more specific tasks, such as classifying individuals directly, rather than having to focus on every step of the image pro cessing. It can also be eventually adapted to become a platform that would allow the entirety of the process to be automated, leaving the scientist with a process that would be as simple as placing an image through the program and receiving a highly likely result, making their work exponentially easier over the course of hundreds of images.