Methods to Improve the Robustness of Right Whale Detection using CNNs in Changing Conditions

Vickers, Will, Milner, Ben, Gorpincenko, Artjoms and Lee, R. (2020) Methods to Improve the Robustness of Right Whale Detection using CNNs in Changing Conditions. In: EUSIPCO 2020. UNSPECIFIED, pp. 106-110. ISBN 978-9-0827-9705-3

Preview

PDF (Accepted_Manuscript) - Accepted Version Download (361kB) | Preview

Abstract

This paper is concerned with developing a method of detecting right whales from autonomous surface vehicles (ASVs) that is robust to changing operating conditions. A baseline convolutional neural network (CNN) is first trained using data taken from a single operating condition. Its detection accuracy is then found to degrade when applied to different operating conditions. Two methods are then investigated to restore performance using just a single model. The first method is an augmented training approach where progressively more data from the new condition is mixed with the original data. The second method uses unsupervised adaptation to adapt the original model to the new conditions. Evaluation under changing environmental and noise conditions reveals the model produced from augmented training data to achieve higher detection accuracy across all conditions than the adapted model. However, the adapted model does not require label data from the new environment and in these situations is a more realistic solution.

Item Type:	Book Section
Faculty \ School:	Faculty of Science > School of Computing Sciences
UEA Research Groups:	Faculty of Science > Research Groups > Interactive Graphics and Audio Faculty of Science > Research Groups > Smart Emerging Technologies Faculty of Science > Research Groups > Collaborative Centre for Sustainable Use of the Seas Faculty of Science > Research Groups > Data Science and AI
Related URLs:	https://www.eurasip.org/Proceedings/Eusi...
Depositing User:	LivePure Connector
Date Deposited:	13 Nov 2020 01:22
Last Modified:	10 Dec 2024 01:12
URI:	https://ueaeprints.uea.ac.uk/id/eprint/77683
DOI:

Downloads

Actions (login required)

View Item