Using Convolution Neural Networks to Classify Coral Species

School Name

South Carolina Governor's School for Science & Mathematics

Grade Level

12th Grade

Presentation Topic

Computer Science

Presentation Type

Mentored

Oral Presentation Award

1st Place

Abstract

The research was done with Convolutional Neural Networks (CNN), a method designed in the 1990’s and has been successfully used over the past decade. The goal of the research was to learn how to create a CNN could correctly identify different coral species at a high rate of efficiency. The program that created would then be uploaded into an aquatic robot which navigates underwater areas on its own to catalog life. To create a test model, a CNN was created to differentiate between cats and dogs. It was able to do so at 98 percent accuracy, and so that became the working model. The next step of the project was to understand how to differentiate between the coral species properly using a relatively small dataset of photos. This process was done by taking the images and individually splicing them into hundreds of smaller, more pixelated photos for the computer to later group. Unfortunately, the when the spliced images were fed back into the clustering algorithm it would only ever return a blank white image. The reason for this is still unknown. Theoretically, however, if this problem were to be later solved, the image would then be reassembled by reversing the affects of the splicing, thereby putting the image back together like a jigsaw puzzle. From there the labeled points would be used for later training by the computer. Then the trained model would be uploaded into the aquatic robot to differentiate between coral species.

Location

Founders Hall 140 A

Start Date

3-30-2019 11:30 AM

Presentation Format

Oral Only

Group Project

No

COinS
 
Mar 30th, 11:30 AM

Using Convolution Neural Networks to Classify Coral Species

Founders Hall 140 A

The research was done with Convolutional Neural Networks (CNN), a method designed in the 1990’s and has been successfully used over the past decade. The goal of the research was to learn how to create a CNN could correctly identify different coral species at a high rate of efficiency. The program that created would then be uploaded into an aquatic robot which navigates underwater areas on its own to catalog life. To create a test model, a CNN was created to differentiate between cats and dogs. It was able to do so at 98 percent accuracy, and so that became the working model. The next step of the project was to understand how to differentiate between the coral species properly using a relatively small dataset of photos. This process was done by taking the images and individually splicing them into hundreds of smaller, more pixelated photos for the computer to later group. Unfortunately, the when the spliced images were fed back into the clustering algorithm it would only ever return a blank white image. The reason for this is still unknown. Theoretically, however, if this problem were to be later solved, the image would then be reassembled by reversing the affects of the splicing, thereby putting the image back together like a jigsaw puzzle. From there the labeled points would be used for later training by the computer. Then the trained model would be uploaded into the aquatic robot to differentiate between coral species.