- Title
- Modelling false positive reduction in maritime object detection
- Creator
- Nkele, Nosiphiwo
- Subject
- Computer vision Neural networks (Computer science)
- Date
- 20xx
- Type
- Thesis
- Type
- Masters
- Type
- MSc (Computer Science )
- Identifier
- http://hdl.handle.net/10353/17168
- Identifier
- vital:40862
- Description
- Target detection has become a very significant research area in computer vision with its applications in military, maritime surveillance, and defense and security. Maritime target detection during critical sea conditions produces a number of false positives when using the existing algorithms due to sea waves, dynamic nature of the ocean, camera motion, sea glint, sensor noise, sea spray, swell and the presence of birds. The main question that has been addressed in this research is how can object detection be improved in maritime environment by reducing false positives and promoting detection rate. Most of Previous work on object detection still fails to address the problem of false positives and false negatives due to background clutter. Most of the researchers tried to reduce false positives by applying filters but filtering degrades the quality of an image leading to more false alarms during detection. As much as radar technology has previously been the most utilized method, it still fails to detect very small objects and it may be applied in special circumstances. In trying to improve the implementation of target detection in maritime, empirical research method was proposed to answer questions about existing target detection algorithms and techniques used to reduce false positives in object detection. Visible images were retrained on a pre-trained Faster R-CNN with inception v2. The pre-trained model was retrained on five different sample data with increasing size, however for the last two samples the data was duplicated to increase size. For testing purposes 20 test images were utilized to evaluate all the models. The results of this study showed that the deep learning method used performed best in detecting maritime vessels and the increase of dataset improved detection performance and false positives were reduced. The duplication of images did not yield the best results; however, the results were promising for the first three models with increasing data.
- Format
- 85 leaves
- Format
- Publisher
- University of Fort Hare
- Publisher
- Faculty of Science and Agriculture
- Language
- English
- Rights
- University of Fort Hare
- Hits: 699
- Visitors: 731
- Downloads: 105
Collections
UFH Department of Computer Science
| Thumbnail | File | Description | Size | Format | |||
|---|---|---|---|---|---|---|---|
| View Details | SOURCE1 | Dissertation- Nkele (201203131)-Computer Science.pdf | 3 MB | Adobe Acrobat PDF | View Details |