A Novel Approach to Coral Fish Detection And Classification in Underwater Footage Based on Convolutional Neural Network

Abstract

Tracking and identifying fish species is crucial to understanding marine ecosystem and its role in the world. In this paper, a cost-effective coral fish detection and identification method is proposed. Using the up-to-date models of Convolutional Neural Network (CNN), this paper is able to analyze an underwater footage and highlight all the detectable fish in color frames, and then identifying the species name among the detectable fish. Fish object detection was employed using Open Images Dataset and Tensorflow Object Detection. The paper further explores CNN with squeeze and excitation for fish classification. The proposed model was evaluated on fish4-knowledge dataset and achieved 100% validation accuracy after 50 epochs, better than AlexNet and RetNet, indicating that the solution is robust and practical. In addition, a dataset for coral fish classification in specific location was built using different sources. The model achieved 100% validation accuracy on the proposed dataset.