Institutional Repository of Vilnius University
Vaizdų atpažinimas ir apdorojimas taikant dirbtinius neuroninius tinklus /
Abstract
dc:descriptionEven though, today neural networks are pretty good at performing a lot of object detection tasks, the main problem still remains -- neural network learning. As computer hardware evolves and becomes faster, this problem can seem to be less and less important. However, one of the main tasks in the neural network learning process cannot be fundamentally improved. It is image labelling. Image labelling is a very long and tedious process and, unfortunately, it cannot be automated. But instead of trying to improve image labelling process, we can eliminate it entirely if we gather images differently. For example, if a computer itself generates images, it can label them just as easily and fast. In this work we analysed the object detection accuracy of a "Faster R--CNN" neural network model when training it with real, synthetic and data from both of these sets using transfer learning. Our goal was to determine which parameters during synthetic data generation using Unity engine has the biggest impact to the accuracy of our neural network. In this study the data was generated with different number of background images and 3D models. Also the number of generated data was changed and, lastly, noise was added to it. We concluded that the biggest impact to the accuracy of our neural network can be made by changing the number of background images, 3D models and the number of images in a training dataset.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2019
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Baliukynas, Tautvydas,
- Contributors dc:contributor
-
- Eglinskas, Mindaugas
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD81919292&prefLang=en_US