IMDEA Networks Institute Publications Repository

Very Small Neural Networks for Optical Classification of Fish Images and Videos

Paraschiv, Marius and Padrino, Ricardo and Casari, Paolo and Fernández Anta, Antonio (2020) Very Small Neural Networks for Optical Classification of Fish Images and Videos. In: Global OCEANS 2020: Singapore – U.S. Gulf Coast, 5-14 October 2020, Virtual Conference.

[img] PDF (Very Small Neural Networks for Optical Classification of Fish Images and Videos) - Published Version
Download (2440Kb)

Abstract

The task of visual classification, done until not long ago by specialists through direct observation, has recently benefited from advancements in the field of computer vision, specifically due to statistical optimization algorithms, such as deep neural networks. In spite of their many advantages, these algorithms require a considerable amount of training data to produce meaningful results. Another downside is that neural networks are usually computationally demanding algorithms, with millions (if not tens of millions) of parameters, which restricts their deployment on low-power embedded field equipment. In this paper, we address the classification of multiple species of pelagic fish by using small convolutional networks to process images as well as videos frames. We show that such networks, even with little more than 12,000 parameters and trained on small datasets, provide relatively high accuracy (almost 42% for six fish species) in the classification task. Moreover, if the fish images come from videos, we deploy a simple object tracking algorithm to augment the data, increasing the accuracy to almost 49\% for six fish species. The small size of our convolutional networks enables their deployment on relatively limited devices.

Item Type: Conference or Workshop Papers (Paper)
Subjects: UNSPECIFIED
Divisions: UNSPECIFIED
Depositing User: Antonio Fernandez
Date Deposited: 04 Dec 2020 12:31
Last Modified: 04 Dec 2020 12:31
URI: http://eprints.networks.imdea.org/id/eprint/2253

Actions (login required)

View Item View Item