Institutional Repository
Technical University of Crete
Technical University of Crete
EN
|
EL
| URI: | http://purl.tuc.gr/dl/dias/1AD8DA64-7591-443F-8617-65B075B8BB36 | ||
| Year | 2022 | ||
| Type of Item | Master Thesis | ||
| License |
|
||
| Bibliographic Citation | Ioanna-Andrianna Rodopoulou, "Coastline litter detection using deep convolutional neural networks", Master Thesis, School of Mineral Resources Engineering, Technical University of Crete, Chania, Greece, 2022 https://doi.org/10.26233/heallink.tuc.92735 | ||
| Appears in Collections |
One of the environmental problems of modern life is the ecological and aesthetic degradation of coastal zones by litter disposal and accumulation. Research conducted to monitor and evaluate pollution incidents in the coastal environment usually consists of groups of volunteers or civil servants who record, count and sort the litter. However, recent research efforts show that litter detection can become an automated process, thanks to the development of remote sensing and computer vision methods. Recording of litter by simple devices like smartphones or more sophisticated devices like drones, can be used as an input to investigate the performance of state-of-the-art object detection algorithms. In the current study, the Mask R-CNN algorithm was used to investigate its performance in detecting coastline litter and classifying it based on its material and type. Mask R-CNN is part of a broader family of deep learning object detection algorithms, the R-CNN object detection algorithm family, that are based on convolutional neural networks (CNNs), and it is able to tackle two tasks of computer vision: object detection and instance segmentation. The performance of Mask R-CNN was mainly evaluated on a domain-specific image dataset created to facilitate this study. It was also tested on an open-sourced drone dataset of litter images, since no study has investigated the use of this algorithm on this specific image dataset. The experimental results in terms of average precision showed that Mask R-CNN exhibited strong potential in litter detection and segmentation of the new dataset, but performed moderately on the drone dataset because of the small size of litter. The algorithm showcased great predictions for well-represented classes, but performed poorly on others that were either under-represented or contained objects that varied significantly in shape, indicating challenges that will need to be further addressed.
Copyright © DIAS 2013
