Poriferal Vision: Using MobileNet to Classify Sponge Spicules through Transfer Learning

• Main Author: Tran, Brian
• Format: Text
• Language: unknown
• Published: SJSU ScholarWorks 2023
• Subjects: Convolutional Neural Networks (CNN); Transfer Learning; Transformations; Zero Padding; Background Padding; Other Computer Engineering; Ocean acidification
• Online Access: https://scholarworks.sjsu.edu/etd_projects/1311; https://scholarworks.sjsu.edu/context/etd_projects/article/2307/viewcontent/tran_brian.pdf

Global warming is an ongoing issue where the Earth is rapidly warming up. It negatively affects the growth of coral through ocean warming and ocean acidification. Many coral communities, home to a large variety of marine life, are expected to be severely impacted by these effects. Past evidence suggests that sponges will take over as the primary reef builders since many species of sponges have skeletons made of silica or glass which is not affected by ocean acidification. More research is needed to determine which kinds of sponge will most likely be able to thrive in today’s climate. This can be done by sampling the seabed for the target era and identifying the spicules that are present in the sample. However, classifying the spicules by hand accurately and within a reasonable amount of time is not tractable with large amounts of spicules. Transfer learning with a pre-existing convolutional neural network (CNN) can be utilized to train a model with a small spicule dataset to classify spicules. In this project, I use transfer learning with MobileNet, a pre-existing CNN, to classify seven categories of spicules. I then use image transformations, zero padding, and background padding on the data before training the model to try to improve its performance on the data. Background padding had the best performance although none of the different iterations of the model could classify all categories well at once.