Lightweight and efficient convolutional neural networks for recognition of dolphin dorsal fins

Gianvito Losapio, Rosalia Maglietta, Tiziano Politi, Ettore Stella, Carmelo Fanizza, Karin Hartman, Roberto Carlucci, Giovanni Dimauro, Vito Renò

Abstract

The study of cetaceans is of vital importance to infer biological information useful to drive sustainable action plans aimed at preserving the marine environment and its biodiversity. In a recent study, we developed a novel algorithm for the detection of dorsal fins in the context of a fully automated pipeline for the photo-identification of Risso's dolphins. A lightweight convolutional neural network (CNN) architecture was proposed to recognize fins among cropped images, filtering the inputs for the photo-identification algorithm. In this paper, we compare the performances of that custom CNN to another extremely efficient architecture: Shufflenet. Training an efficient classifier is a key effort to speed up the first part of the photo-identification pipeline, enabling the feasibility of large scale ecological studies. The experiment confirms that both architectures provide a robust feature extraction capability for the problem in hand, even with a significantly smaller number of parameters with respect to other popular state-of-the-art CNNs.

Download

IMEKO-TC19-MetroSea-2020-13.pdf

DOI

-

IMEKO TC

TC19 - Environmental Measurements

Back to the proceedings