Testing the Effectiveness of Transfer Learning for Underwater Debris Classification and Object Detection

Emmen, Pepijn Floris Johannes (2024) Testing the Effectiveness of Transfer Learning for Underwater Debris Classification and Object Detection. Bachelor's Thesis, Artificial Intelligence.

	Text Thesis Final (3).pdf Restricted to Registered users only Download (6MB)
Preview	Text wolf 11 mrt.pdf Download (146kB) | Preview

Abstract

Gathering images of underwater debris is a highly expensive and challenging task. These images are required to train autonomous agents to detect, classify and remove underwater debris. This thesis aims at investigating the effectiveness of a low-cost solution utilizing transfer learning with data augmentation techniques. The focus is on two distinct fields: object detection using a Faster R-CNN model and image classification with a custom made CNN. The approach involves leveraging easily accessible and inexpensive above-water images of debris, thereby reducing the reliance on costly underwater datasets. The methodology includes training a baseline model directly on the underwater debris dataset to set a performance benchmark. Subsequently, a model is trained on the above-water debris dataset and then fine-tuned on the underwater debris dataset, with strategic freezing of certain layers. The study showed a significant improvement (p < 0.05) in classification metrics for transfer learning models compared to the respective baseline. Furthermore, no significant improvements (p > 0.05) were found for object detection models utilizing transfer learning compared to the respective baseline. The results highlight how transfer learning can be an effective tool if leveraged with carefully designed pretraining datasets and data augmentation techniques in complex underwater classification tasks.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Wolf, B.J.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	11 Mar 2024 13:46
Last Modified:	11 Mar 2024 13:56
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/32039

Actions (login required)

View Item