Fine-grained plankton recognition

Abstract

This thesis focuses on fine-grained image analysis, specifically plankton recognition, utilizing attention-based methodologies. It explores principles of machine learning, cnn architecture, optimization techniques, and strategies to mitigate overfitting. The study investigates current methods in fine-grained image analysis, with a particular emphasis on plankton imaging. The main contribution of this research is the application of an attention-based recognition model, known as the ffvt, on a Plankton dataset. The ffvt combines the strengths of feature fusion and vision transformers to capture intricate visual patterns and attend to informative image regions, enabling accurate species classification. The results achieved by the ffvt on the Plankton dataset are notably impressive, exhibiting a high level of accuracy. Specifically, the model attained a weighted average accuracy of 97% when evaluated on the entire dataset without any exclusion of fine-grained classes. The thesis concludes with experimental validation of the model and a comprehensive discussion on the findings, along with suggestions for potential future work.