Automatic Recognition of Marine Creatures using Deep Learning

The identification of marine species is a challenge for people all over the world, and the situation is not different for Mauritians. It is of utmost importance to create an automated system to correctly identify marine species. In the past, researchers have used machine learning to address the issue of marine creature recognition. The manual feature extraction part of machine learning complicates model creation as features have to be extracted manually using an appropriate filter. In this work, we have used deep learning models to automate the feature extraction procedure. Currently, there is no publicly available dataset of marine creatures from the Indian Ocean. We created one of the biggest datasets used in this field, consisting of 51 different marine species collected from the Odysseo Oceanarium in Mauritius. The original dataset has a total of 5,709 images and is imbalanced. Image augmentations were performed to create an oversampled version of the dataset with 171 images per class, for a total of 8,721 images. The MobileNetV1 model trained on the oversampled dataset with a split ratio of 80% for training and 10% for validation and testing was the best performing one in terms of classification accuracy and inference time. The model had the smallest inference time of 0.10 seconds per image and attained a classification accuracy of 99.89% and an F1 score of 99.89%.