Semantic Segmentation for Aerial Imagery Using U-Nets
Technical Report,01 Sep 2018,01 Mar 2020
AIR FORCE INSTITUTE OF TECHNOLOGY WRIGHT-PATTERSON AFB OH WRIGHT-PATTERSON AFB United States
Pagination or Media Count:
In situations where global positioning systems are unavailable, alternative methods of localization must be implemented. A potential step to achieving this is semantic segmentation, or the ability for a model to output class labels by pixel. This research aims to utilize datasets of varying spatial resolutions and locations to train a fully convolutional neural network architecture called the U-Net to perform segmentations of aerial images. Variations of the U-Net architecture are implemented and compared to other existing models in order to determine the best in detecting buildings and roads. A final dataset will also be created combining two datasets to determine the ability of the U-Net to segment classes regardless of location. The final segmentation results will demonstrate the overall efficacy of semantic segmentation for different datasets for potential localization applications.
- Navigation and Guidance