Simon Atileh

Using transfer learning to assist in the detection of starfish in the great barrier reef

Started Sept. 21, 2022

Abstract or project description

Transfer Learning is a consistently successful technique for adapting Deep Learning models in Computer Vision to domain-specific problems. (You Only Look Once) YOLOv5 is a massive image segmentation and object detection model that has been continuously refined into its current form since 2015 by a team at the University of Washington Paul G. Allen School for Computer Science (UDub). The original form of this Deep Net is trained on the Common Objects in Context (COCO) dataset. The Great Barrier Reef is the world’s largest coral reef and is under threat because of the overpopulation of one particular starfish—the coral-eating crown-of-thorns starfish (COTS). Scientists, tourism operators, and reef managers established a large-scale intervention program to control COTS outbreaks to ecologically sustainable levels. The state-of-the-art technique (SOTA) for surveying COTS is a manual technique performed by a snorkel diver called "Manta Tow." While towed by a boat, the expert visually assesses the reef, stopping to record variables observed every 200 meters. While generally effective, this method faces clear limitations, including operational scalability, data resolution, reliability, and traceability. The Great Barrier Reef Foundation established an innovation program to place underwater cameras that collect thousands of reef images and require AI technology that could drastically improve the efficiency and scale at which reef managers detect and control COTS outbreaks. To scale up video-based surveying systems, Australia’s national science agency, CSIRO, has teamed up with Google to develop innovative machine learning technology that can analyze large image datasets accurately, efficiently, and in near real-time. We are improving this technology by transferring the successful YOLOv5 model to this problem through pre-training, adversarial training, fine-tuning, and meta-learning.