North Atlantic right whale localization and recognition using very deep and leaky Neural Network

Loading...
Thumbnail Image
Date
2016
ORCID
Advisor
Referee
Mark
Journal Title
Journal ISSN
Volume Title
Publisher
Vysoké učení technické v Brně, Fakulta strojního inženýrství, Ústav matematiky
Altmetrics
Abstract
We describe a deep learning model that can be used to recognize individual right whales in aerial images. We developed our model using a data set provided by the National Oceanic and Atmospheric Administration. The main challenge we faced when working on this data set is that the size of the training set is very small (4,544 images) with some classes having only 1 image. While this data set is by far the largest of its kind, it is very di cult to train a deep neural network with such a small data set. However, we were able to overcome this challenge by dividing this problem into smaller tasks and by reducing the viewpoint variance in the data set. First, we localize the body and the head of the whale using deep learning. Then, we align the whale and normalize it with respect to rotation. Finally, a network is used to recognize the whale by analyzing its callosities. The top-1 accuracy of the model is 69.7% and the top-5 accuracy is 85%. The solution we describe in this paper was ranked 5th (out of 364 teams) in a challenge to solve this problem.
Description
Citation
Mathematics for Applications. 2016 vol. 5, č. 2, s. 155-170. ISSN 1805-3629
http://ma.fme.vutbr.cz/archiv/5_2/ma_5_2_kabani_el_sakka_final.pdf
Document type
Peer-reviewed
Document version
Published version
Date of access to the full text
Language of document
en
Study field
Comittee
Date of acceptance
Defence
Result of defence
Document licence
© Vysoké učení technické v Brně, Fakulta strojního inženýrství, Ústav matematiky
Collections
Citace PRO