PhD Interns Help DST Research Surge Ahead
Home » Case Studies » PhD Interns Help DST Research Surge Ahead
PhD INTERNS HELP DST RESEARCH SURGE AHEAD
A bevy of Australia’s brightest PhD students have begun undertaking short-term 3-5 month internships at Defence Science and Technology Group (DST) in win-win partnerships.
DST gains access to their specialist know-how and enthusiasm (and their team of Academic Mentors), while students are able to build essential industry research skills at the frontline of Defence and national security innovation. Last year, DST agreed to support 100 APR interns over the next four years.
DAYSI DIVES INTO THE DEEP LEARNING END
RMIT PhD student, Ana “Daysi” Ruvalcaba Cardenas, was matched with DST to explore object recognition using deep learning on low-resolution time-of-flight data.
For Daysi, it was an interesting challenge working with data from highly-sensitive single photon avalanche detector chips. Compared to more traditional academic work, one of the challenges was that imaging outdoors and at long range adds significant atmospheric disturbance to the recording and greatly complicates the processing.
“The purpose of my project is to investigate deep learning techniques with this new type of data to identify different objects and localise them in a certain area. I’ve implemented and reviewed several techniques with different types of filtering, with good results,”
Daysi Cardenas, PhD Intern at DST
The strength of deep learning for image recognition arises from training the system on a large number of images, but faced with only a tiny SPAD dataset, Daysi had to graft SPAD image recognition capability onto a pre-existing neural framework that had been trained on millions of non-SPAD images.
“I took another framework, eliminated certain classification layers and added new ones. By doing that I was able to train the network with the new SPAD data to classify our objects,”
Daysi Cardenas, PhD Intern at DST
The initial results demonstrate the success of Daysi’s technique. After some training her network was able to correctly classify different types of chairs, UAVs and planes.
This article was first published in DST Connections Issue 237.
