Seabed Image Segmentation Using Machine Learning and Mean Field Annealing
Open Access
- Author:
- Witlin, Isaac
- Graduate Program:
- Electrical Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- February 24, 2025
- Committee Members:
- John Doherty, Professor in Charge/Director of Graduate Studies
David Jonathan Miller, Thesis Advisor/Co-Advisor
Daniel C. Brown, Committee Member
John F Doherty, Committee Member - Keywords:
- Machine Learning
U-Net
Mean Field Annealing
Seabed Image
Backprojection
Wavelets
Texton
Segmentation
Semantic Segmentation
Unsupervised Segmentation
Seabed Composition
Synthetic Aperture Sonar (SAS)
Simple Linear Iterative Clustering (SLIC)
Convolutional Neural Networks (CNNs) - Abstract:
- This thesis examines the use of Mean Field Annealing (MFA) techniques applied to the task of segmenting underwater acoustic imagery into known categories of seabed types. Established techniques including wavelets, textons, and Convolutional Neural Networks are applied along with the use of MFA to improve unsupervised segmentation results. An explainer on Synthetic Aperture Sonar (SAS) data is provided. Detailed explanations of the various image processing techniques used for unsupervised seabed segmentation are provided as well. Experiments are performed to examine optimal selection of MFA hyper-parameters, and final accuracy of results are defined and provided. MFA produced a smoother segmentation mask with slightly higher accuracy over existing techniques. MFA can be used in tandem with other segmentation techniques to train and fine tune a neural network. That network can then be used to ingest never before seen images and segment them in real time. More accurate unsupervised segmentation improves the ability to train neural networks to accurately segment never before seen SAS imagery. This serves to reduce the cost and complexity of SAS survey results and helps quickly identify seabed composition of a chosen area of operation.