Sea ice surface characterization via semantic segmentation with convolutional neural networks
Visual data provides rich information about real-world objects. Computer vision is a substantial and growing field which seeks to distill useful information from photographic imagery. The primary focus of this work centers on the application of machine-learning based computer vision algorithms, alon...
Main Author: | |
---|---|
Format: | Thesis |
Language: | English |
Published: |
Memorial University of Newfoundland
2022
|
Subjects: | |
Online Access: | https://research.library.mun.ca/15765/ https://research.library.mun.ca/15765/3/converted.pdf |
id |
ftmemorialuniv:oai:research.library.mun.ca:15765 |
---|---|
record_format |
openpolar |
spelling |
ftmemorialuniv:oai:research.library.mun.ca:15765 2023-10-01T03:59:24+02:00 Sea ice surface characterization via semantic segmentation with convolutional neural networks King, Matthew Peter Ivan 2022-10 application/pdf https://research.library.mun.ca/15765/ https://research.library.mun.ca/15765/3/converted.pdf en eng Memorial University of Newfoundland https://research.library.mun.ca/15765/3/converted.pdf King, Matthew Peter Ivan <https://research.library.mun.ca/view/creator_az/King=3AMatthew_Peter_Ivan=3A=3A.html> (2022) Sea ice surface characterization via semantic segmentation with convolutional neural networks. Masters thesis, Memorial University of Newfoundland. thesis_license Thesis NonPeerReviewed 2022 ftmemorialuniv 2023-09-03T06:50:24Z Visual data provides rich information about real-world objects. Computer vision is a substantial and growing field which seeks to distill useful information from photographic imagery. The primary focus of this work centers on the application of machine-learning based computer vision algorithms, along with minor applications of more traditional computer vision techniques. The specific task approached herein is known as semantic segmentation; the methodology by which each region of an image, at an individual pixel level, is assigned a classification from a predetermined set of possible classes. The classes considered in this work are: open water, level ice, broken ice, ridged ice, ice in flexural failure, and an ‘other’ class. Accurate segmentation of these classes is the primary objective of this work. The imagery utilized in this work were captured from two cameras mounted on different piers of the Confederation Bridge during the local ice season. Several hundred of the images collected have been manually labelled and split into training, validation, and testing subsets. These data have been used to train an ensemble of convolutional neural networks. Transfer learning is applied such that the encoder portions of the neural networks have been pretrained on the ImageNet dataset, providing them with the capability to produce meaningful feature maps while significantly reducing the training time required for the overall models to learn the semantic segmentation task at hand. Thesis Sea ice Memorial University of Newfoundland: Research Repository |
institution |
Open Polar |
collection |
Memorial University of Newfoundland: Research Repository |
op_collection_id |
ftmemorialuniv |
language |
English |
description |
Visual data provides rich information about real-world objects. Computer vision is a substantial and growing field which seeks to distill useful information from photographic imagery. The primary focus of this work centers on the application of machine-learning based computer vision algorithms, along with minor applications of more traditional computer vision techniques. The specific task approached herein is known as semantic segmentation; the methodology by which each region of an image, at an individual pixel level, is assigned a classification from a predetermined set of possible classes. The classes considered in this work are: open water, level ice, broken ice, ridged ice, ice in flexural failure, and an ‘other’ class. Accurate segmentation of these classes is the primary objective of this work. The imagery utilized in this work were captured from two cameras mounted on different piers of the Confederation Bridge during the local ice season. Several hundred of the images collected have been manually labelled and split into training, validation, and testing subsets. These data have been used to train an ensemble of convolutional neural networks. Transfer learning is applied such that the encoder portions of the neural networks have been pretrained on the ImageNet dataset, providing them with the capability to produce meaningful feature maps while significantly reducing the training time required for the overall models to learn the semantic segmentation task at hand. |
format |
Thesis |
author |
King, Matthew Peter Ivan |
spellingShingle |
King, Matthew Peter Ivan Sea ice surface characterization via semantic segmentation with convolutional neural networks |
author_facet |
King, Matthew Peter Ivan |
author_sort |
King, Matthew Peter Ivan |
title |
Sea ice surface characterization via semantic segmentation with convolutional neural networks |
title_short |
Sea ice surface characterization via semantic segmentation with convolutional neural networks |
title_full |
Sea ice surface characterization via semantic segmentation with convolutional neural networks |
title_fullStr |
Sea ice surface characterization via semantic segmentation with convolutional neural networks |
title_full_unstemmed |
Sea ice surface characterization via semantic segmentation with convolutional neural networks |
title_sort |
sea ice surface characterization via semantic segmentation with convolutional neural networks |
publisher |
Memorial University of Newfoundland |
publishDate |
2022 |
url |
https://research.library.mun.ca/15765/ https://research.library.mun.ca/15765/3/converted.pdf |
genre |
Sea ice |
genre_facet |
Sea ice |
op_relation |
https://research.library.mun.ca/15765/3/converted.pdf King, Matthew Peter Ivan <https://research.library.mun.ca/view/creator_az/King=3AMatthew_Peter_Ivan=3A=3A.html> (2022) Sea ice surface characterization via semantic segmentation with convolutional neural networks. Masters thesis, Memorial University of Newfoundland. |
op_rights |
thesis_license |
_version_ |
1778533363993280512 |