A Comparison of Target Detection Algorithms Using DSIAC ATR Algorithm Development Data Set

The problem of reliably detecting targets in infra-red imagery at acceptably low false alarm rates continues to be of significant interest to the research community. The challenge is that, unlike common applications that use RGB cameras, the amount of labeled training data is very limited. Further, infra-red phenomenology differs significantly from that in the visible band, meaning that algorithms trained on RGB data cannot be readily incorporated into infra-red applications.

There has been a tremendous surge in deep learning/convolutional neural nets (CNNs) in the field of computer vision at large. Hence, there is significant interest in determining if similar performance gains can be achieved by applying these techniques in the infra-red domain.

In this presentation from SPIE Defense + Commercial Sensing, Abhijit Mahalanobis of University of Central Florida, presents preliminary results of infra-red target detection using the well-known Faster R-CNN network using a publicly available MWIR data set released by NVESD.

"This work is looking at the feasibility of using convolutional neural nets on a data set that was released by Night Vision Lab several years ago," said Mahalanobis. "This has actually been out there for a while. I think one of the challenges we have in the ATR community is a good data set with a lot of reliable ground truth, right? The commercial world is very fortunate they don't have this issue with searching for data sets. But in the ATR world, this has always been a challenge for us."

SPIE Fellow Abhijit Mahalanobis is an assistant professor at University of Central Florida (UCF). Mahalanobis came to UCF from Lockheed Martin, where he was a Senior Fellow. His primary research areas are in Systems for Information processing, Computational Sensing and Imaging, and Video/Image processing for information exploitation and ATR.

Read the full manuscript on the SPIE Digital Library.