Coral classification in unconstrained underwater imagery

Abstract

Classification of coral reefs in uncontrolled underwater images plays a crucial role in the examination of marine biodiversity and efforts for conservation. This study proposes a sophisticated methodology employing deep learning techniques to enhance the precision and effectiveness of coral classification. More precisely, personalized ResNet models combined with multi-head attention and one-shot learning were created, alongside the utilization of YOLOv8 for object detection and segmentation tasks. The effectiveness of the approach was assessed on the Eilat, Eilat2, and RSMAS datasets, displaying notable enhancements in classification accuracy. ResNet models, which were pretrained on the Eilat2 dataset, displayed exceptional accuracy in the identification of coral species. YOLOv8 was employed for segmentation purposes, effectively outlining individual coral formations. The amalgamation of these advanced deep learning techniques substantially improved the precision and dependability of coral classification. The findings of this study contribute to the progression of coral reef conservation initiatives and investigations into marine biodiversity by offering robust tools for monitoring coral well-being and diversity. The prosperous application of instance segmentation, supported by Convolutional Neural Networks (CNNs) and Vision Transformers, showcases a high potential for generalization and robust performance in practical underwater imagery.