Towards Sustainable Oceans: Deep Learning Models for Accurate COTS Detection in Underwater Images
Afonne, Joseph, Chinenye (2024-02-22)
Afonne, Joseph, Chinenye
22.02.2024
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi-fe202402228305
https://urn.fi/URN:NBN:fi-fe202402228305
Tiivistelmä
Object detection is one of the main tasks in computer vision, which includes image classification
and localization. The application of object detection is now widespread as it powers various applications such as self-driving cars, robotics, biometrics, surveillance, satellite image analysis, and in healthcare, to mention just a few. Deep learning has taken computer vision to a different
horizon. One of the areas that will benefit immensely from deep learning computer vision is the
detection of killer starfish, the crown-of-thorns starfish (COTS). For decades, this killer starfish
has dealt a big blow to the Great Barrier Reef in Australia, the world’s largest system of reefs, and in other places too. In addition to impacting negatively environmentally, it affects revenue
generation from reef tourism. Hence, reef managers and authorities want to control the populations of crown-of-thorns starfish, which have been observed to be the culprits. The deep learning
technique offers real-time and robust detection of this creature more than earlier traditional
methods that were used to detect these creatures.
This thesis work is part of a competition for a deep learning approach to detect COTS in
real time by building an object detector trained using underwater images. This offers a solution
to control the outbreaks in the population of these animals. Deep learning methods of Artificial
Intelligence (AI) have gained popularity today because of its speed and high accuracy in detection and have performed better than the earlier traditional methods. They can be used in
real-time object detection, and they owe their speed to convolutional neural networks (CNN).
The thesis gives a comprehensive literature review of the journey so far in the field of computer
vision and how deep learning methods can be applied to detect COTS. It also outlines the
steps involved in the implementation of the model using the state-of-the-art computer vision
algorithm known for its speed and accuracy – YOLOv8. The COTS detection model was trained
using the custom dataset provided by the organizers of the competition, harnessing the powers
of deep learning methods such as transfer learning, data augmentation, and preprocessing
of underwater images to achieve high accuracy.
Evaluation of the results obtained from the training showed a mean average precision of
0.803mAP at IoU of 0.5-0.95, acknowledging the detector model’s versatility in making accurate
detection at different confidence levels. This supports the hypothesis that when we use pre trained model, this enhances the performance of our model for better object detection tasks.
Certainly, better detection accuracy is one way to detect killer starfish, the crown-of-thorns starfish (COTS), and help protect the oceans.
and localization. The application of object detection is now widespread as it powers various applications such as self-driving cars, robotics, biometrics, surveillance, satellite image analysis, and in healthcare, to mention just a few. Deep learning has taken computer vision to a different
horizon. One of the areas that will benefit immensely from deep learning computer vision is the
detection of killer starfish, the crown-of-thorns starfish (COTS). For decades, this killer starfish
has dealt a big blow to the Great Barrier Reef in Australia, the world’s largest system of reefs, and in other places too. In addition to impacting negatively environmentally, it affects revenue
generation from reef tourism. Hence, reef managers and authorities want to control the populations of crown-of-thorns starfish, which have been observed to be the culprits. The deep learning
technique offers real-time and robust detection of this creature more than earlier traditional
methods that were used to detect these creatures.
This thesis work is part of a competition for a deep learning approach to detect COTS in
real time by building an object detector trained using underwater images. This offers a solution
to control the outbreaks in the population of these animals. Deep learning methods of Artificial
Intelligence (AI) have gained popularity today because of its speed and high accuracy in detection and have performed better than the earlier traditional methods. They can be used in
real-time object detection, and they owe their speed to convolutional neural networks (CNN).
The thesis gives a comprehensive literature review of the journey so far in the field of computer
vision and how deep learning methods can be applied to detect COTS. It also outlines the
steps involved in the implementation of the model using the state-of-the-art computer vision
algorithm known for its speed and accuracy – YOLOv8. The COTS detection model was trained
using the custom dataset provided by the organizers of the competition, harnessing the powers
of deep learning methods such as transfer learning, data augmentation, and preprocessing
of underwater images to achieve high accuracy.
Evaluation of the results obtained from the training showed a mean average precision of
0.803mAP at IoU of 0.5-0.95, acknowledging the detector model’s versatility in making accurate
detection at different confidence levels. This supports the hypothesis that when we use pre trained model, this enhances the performance of our model for better object detection tasks.
Certainly, better detection accuracy is one way to detect killer starfish, the crown-of-thorns starfish (COTS), and help protect the oceans.