A Study on Deepsea Fish Detection Using Convolutional Neural Networks

<p>This study investigates automated deep-learning methodologies for two primary tasks, firstly fish and its habitat classification and deep-sea fish detection using the DeepFish dataset. A pretrained ResNet-50 model attained a validation accuracy of 82.8% for multi-class habitat classification, demonstrating robust performance in most habitats based on 6,517 images, although exhibiting lower recall for visually similar or underrepresented classes. A binary ResNet-50 classifier achieved 99.8% accuracy in distinguishing fish from. no-fish images using approximately 40,000 images. The YOLOv5s model, trained on 4,505 images containing 15,463 bounding-box annotations, achieved a mean average precision (mAP@0.5) of 98.2% All models were trained on the SeaHawk computing cluster with four GPUs. These findings highlight promising directions for future research in video-based modeling, semi-supervised learning and efficient marine edge computing.</p>

Name thesis:degree_name

Master of Science (MS)

Level thesis:degree_level

Masters Thesis

Year dc:date.available

2025

Author dc:creator

• Madari, Poojitha Priyadarshini

Contributors dc:contributor

• Dr. Chaoyang Zhang
• Dr. Sarah Lee
• Dr. Zhaoxian Zhou

• Machine Learning
• Deep learning
• YOLOv5
• ResNet-50
• Marine
• DeepFish Dataset

Repository record dc:identifier

https://aquila.usm.edu/masters_theses/1150

OAI identifier oai:identifier

oai:aquila.usm.edu:masters_theses-2267