Fundamentals of Target Classification Using Deep Learning

In this presentation from SPIE Defense + Commercial Sensing, Irene L. Tanner of University of Central Florida, examines the application of deep learning for automated target recognition (ATR) using a shallow convolutional neural network (CNN) and infrared images from a public domain data provided by US Army Night Vision Laboratories.

"The motivation for this project comes from the fact that infrared images can be used to train algorithms to classify targets in scenes with natural backgrounds," says Tanner. "We know that infrared imaging sensors provide images of heat signatures of targets at any time of day, and this has important functionality for aiding operators in surveillance and situational awareness."

Tanner notes that previous approaches have used quadratic correlation filters, auto-encoders, and similar technologies. But over the past few years, deep learning has really come to the forefront, and has been shown to be very good at classification.

The research team's main goal was, given a scene, to either detect a target or, given an image patch from the scene, to distinguish between target or clutter right there. The first goal was to discriminate between target and clutter -- further goals were to discriminate between tracked versus wheeled vehicles, and so on.

"In the future, we want to improve our detection algorithm, whether that be optimizing the parameters of the difference of Gaussian kernel or using a standard detection algorithm like Faster R-CNN," says Tanner. "And we would also like to include additional vehicle classifiers, such as civilian versus military vehicles or tracked versus wheeled vehicles."

Irene L. Tanner is an undergraduate research assistant in the Center for Research in Computer Vision at University of Central Florida.

Read the full manuscript on the SPIE Digital Library.