David Malmgren-Hansen: For military surveillance purposes consistent image quality is imperative. The time-of-day as well as harsh weather conditions typically has a big impact on the image quality of classical optical image capturing techniques.
This of course makes them sub-optimal for surveillance purposes. Synthetic Aperture Radar (SAR) imaging is, as opposed to regular image capturing techniques, quite robust to these varying conditions, making SAR imaging a valuable alternative for acquiring surveillance information.
Using SAR images for object classification purposes this invariance removes many difficulties, but introduces new challenges. The background for this is NATOs Allied Ground Surveillance project (AGS), where high-altitude Unmanned Aerial Vehicles (UAVs) are planned to capture SAR images of both land based and maritime vehicles/vessels to provide better military intelligence.
This project addresses the need to develop methods for object classification on images from different image modalities. The main focus will be put on modern airborne Synthetic Aperture Radar (SAR) images, but the found methods are to be applied in a broader range of image modalities with classification and possibly identification goal. Furthermore the project aims at adding knowledge about how to optimize the classification results with respect to the parameters and help in finding the optimal way of capturing images to obtain the best classification.
Effective start/end date 01/09/2014 → 13/12/2017
Published as PhD report: Convolutional Neural Networks - Generalizability and Interpretations
Supervisor: Allan Aasbjerg Nielsen
Section for Visual Computing