Date of Award

2022-12-01

Degree Name

Master of Science

Department

Computer Science

Advisor(s)

Olac Fuentes

Abstract

Holographic cameras show potential as a sensor to monitor oil spills. Holographic cameras record the light interference from particles in a volume of space, producing an image called a hologram. Processing these holograms is known as hologram reconstruction. It produces a representation of particles located in three-dimensional space. These cameras can record precise shapes and sizes of particles in a volume of water. However, it is very time-consuming and resource-intensive to process the images. Most algorithms that perform particle analysis require the hologram reconstruction step. The well-documented hybrid method is one such algorithm. Machine learning is one possible technique that shows potential for avoiding the costly hologram reconstruction altogether while extracting particle statistics. We tested two different machine learning models, U-Net with a resnet34 backbone and YOLOv5, and compared their performance. We collected a dataset of real-world hologram images of oil particles. We then trained our machine learning models on this real-world data with hybrid method derived labels. We used average precision and average recall to compare model performance. Compared to the hybrid method, we demonstrated a processing time reduction of 99\%. YOLOv5 with GPU acceleration proved to be the superior model in all metrics. Machine learning is a promising technique for processing these hologram images rapidly. This allows quick-response agencies such as oil spill response teams the ability to use holographic cameras to assess oil spills.

Language

en

Provenance

Received from ProQuest

File Size

46 p.

File Format

application/pdf

Rights Holder

Daniel Cruz

Share

COinS