Publication Date

Spring 5-22-2019

Degree Type

Master's Project

Degree Name

Master of Science (MS)


Computer Science

First Advisor

Philip Heller

Second Advisor

Leonard Wesley

Third Advisor

Kevin Smith


Artificial neural networks (ANN), sponge spicules, bioinformatics, computer vision, deep convoluted neural networks (CNN), FlowCAM, generative adversarial networks (GAN), global silica biogeochemical cycle, image transformations


Sponges provide nourishment as well as a habitat for various aquatic organisms. Anatomically, sponges are made up of soft tissue with a silica based exoskeleton which serves both as support and protection for the underlying tissue. The exoskeleton persists after the tissue decomposes, and microscopic parts of the exoskeleton break away to form spicules. Oceanographic studies have shown that the density of the sponge spicules is a good indicator of the sponge population in an area. This measure can be used to study sponge population dynamics over time. The spicule density is measured by imaging spicules from samples of water extracted from the oceans using an instrument called FlowCAM, which separates and photographs individual small items in a sample. It has a high processing rate, but is inefficient at computationally analyzing large numbers of photographs. Computer vision technologies, particularly deep learning using Artificial Neural Networks, and Support Vector Machines have shown to be effective in handling large scale image classification problems and are the de-facto standard in image recognition problems. Typically, these models require a large amount of data to learn the underlying distribution in datasets effectively and avoid model overfitting, which is currently a challenge to procure a vast dataset of images. To mitigate this challenge and achieve the overarching purpose of developing a high- performance classifier, we demonstrate various geometrical image transformation techniques to enhance the size of the dataset. We also show initial experimental results for training Generative Adversarial Networks for artificial synthesis of spicule images. Finally, we develop a Convolutional Neural Network and compare its performance against a Support Vector Machine for classifying images of sponge spicules training both the models on the original set of images and the newly generated set of images and achieve a test accuracy of 95% with a CNN trained on the newly generated images.