Shark Detection Classification

Abstract

In recent years, collaborative work between previous CS4624 capstone groups and Dr. Francesco Ferretti's team contributed to the development of the SharkPulse project. With the goal of enhancing shark conservation and elevating public awareness through the collection and analysis of global, crowd-sourced shark sightings data, SharkPulse developed a data / machine-learning pipeline to detect and classify sharks from a given image. This report presents the improvement of the machine-learning pipeline previously established in "Shark detection and classification with machine learning" (Jenrette et al.). The improvements to the pipeline increased classification accuracy as well as species breadth. Mainly, the existing classification architectures are replaced with Transformers (ViTs). The updated shark identifier achieves an accuracy of 96%, the updated genus classifier, an accuracy of 72%, and the updated genus-specific species classifiers, an average accuracy of 74%. This updated classification system is able to classify 27 genera and 51 species. A framework for automating data-collection, model training, and maintenance is also introduced. Potential future work is discussed, including integrating the model into the SharkPulse platform.