Sara Vanaki shares advice for future scholars and describes her project using computer vision to identify creatures in sea floor photos.
NOAA has collected millions of images of the seafloor to learn about the organisms that live there. But how do you look at millions of photos and identify what is in them? Sara Vanaki, a 2021 Hollings scholar, spent her summer internship with NOAA Fisheries taking on this challenge. She created a machine learning program that can analyze the images and identify sand dollars and sea stars.
![Two images are side-by-side. Each image of the sea floor looks like sand with shell fragments and some darker circles that are identifiable as sand dollars if you zoom in. In the first photo, there are squares around some of the dark circles, but many were not "seen" and identified by the computer. The confidence numbers are almost all under .10. In the right photo, all of the dark circles are identified as sand dollars and the confidence values are much higher. (Image credit: Sara Vanaki) Two images are side-by-side. Each image of the sea floor looks like sand with shell fragments and some darker circles that are identifiable as sand dollars if you zoom in. In the first photo, there are squares around some of the dark circles, but many were not "seen" and identified by the computer. The confidence numbers are almost all under .10. In the right photo, all of the dark circles are identified as sand dollars and the confidence values are much higher.](/sites/default/files/styles/landscape_width_1275/public/2022-08/Sara-Vanaki-ComputerVision_headshot_combined.png?h=151db1d8&itok=rmGvDIBR)
These images showcase the improvement in Sara’s machine learning program. The image on the left is from her team’s first time running the program. The squares indicate an identified sand dollar and the number next to the square indicates how confident it is in the identification. After going through several hundred such images and checking machine annotations, they were able to improve the machine learning program. The image on the right shows how accurate the machine was after improvements. (Image credit: Sara Vanaki)