View Item 
        •   Utrecht University Student Theses Repository Home
        • UU Theses Repository
        • Theses
        • View Item
        •   Utrecht University Student Theses Repository Home
        • UU Theses Repository
        • Theses
        • View Item
        JavaScript is disabled for your browser. Some features of this site may not work without it.

        Browse

        All of UU Student Theses RepositoryBy Issue DateAuthorsTitlesSubjectsThis CollectionBy Issue DateAuthorsTitlesSubjects

        Pre-processing to Mitigate Inherent Problems of Underwater Data Used for Object Detection

        Thumbnail
        View/Open
        Thesis_Pascal_Verkade_Final.pdf (21.04Mb)
        Publication date
        2022
        Author
        Verkade, Pascal
        Metadata
        Show full item record
        Summary
        Ever since deep learning found its footing in the general field of AI, it has been growing increasingly larger. Deep learning technologies are nowadays commonly used in real-world applications. One field in which these applications are used is in the field of environmental protection and water management. This research is based on a project that aims to implement object detection on underwater data. As a first step, this research looks into the underwater project data for a plant detection task. The data suffers from various problems inherent to underwater data making it difficult to even label. The goal of this research is to find out how this data can be pre-processed to mitigate these problems before training a deep learning model. Various image enhancement techniques are tried on the data and through subjective analysis, RGHS is selected as the best solution for the problems of this data. Several data augmentations are used to improve the trainingset even further. A YOLOv5 model trained on the enhanced data showed better results than one trained on unenhanced data. And despite the fact that the data was more problematic than most other (public) datasets, the model performed reasonably well and similar to models trained on other data. These results suggest that the pre-processing techniques applied on the problematic data had a positive effect, improving the data and making it more usable.
        URI
        https://studenttheses.uu.nl/handle/20.500.12932/43281
        Collections
        • Theses
        Utrecht university logo