Deep Learning Platform for High Performance Machine Learning for Automatic Target Detection and Recognition
Abstract
Artificial Intelligence/Machine Learning (AI/ML) techniques are becoming increasingly important in many real-world applications. In particular, Deep Learning (DL) is proving to be a very successful set of tools able to surpass humans in solving certain highly computational tasks. Consequently, DL is the fastest growing trend in big data analysis and is increasingly the method of choice for many applications. The success of DL for object recognition using RGB images has led to an interest in using these techniques for automatic target recognition (ATR) using Infra-Red (IR) imagery. In fact, various DoD research projects (e.g., supported by the US Army and Air Force) have emphasized research in DL for ATR. In particular, the ArmyÕs Advanced Targeting and Lethality Automated System (ATLAS) program is focusing on DL for target detection, classification, and identification. Unlike traditional AI/ML techniques that design and handcraft features based on domain specific knowledge, DL exploits feature representations learned exclusively from training data. This joint learning of features and decision statistics requires huge amounts of training data and intensive computing resources. The PI has been working in DL for the past few years. He has developed algorithms that combine DL with multiple instance learning and adapted them to two main applications: (1) vehicle make and model recognition from RGB images and video; and (2) buried explosive object detection using ground penetrating radar data. Recently, the PI started working on the ATLAS program, in collaboration with the US Army Night Vision and Electronic Sensors Directorate (NVESD), developing DL algorithms for target recognition using IR images, with an emphasis on explaining and interpreting the output. This project requests the acquisition of a Deep Learning Platform for Big Data (DLP-BD) with high performance data storage, networking, and processing capabilities to enable machine learning and specifically DL for ATR. The acquisition responds to the need for an in-house instrument enabling permanent storage of sensitive data that are in the tens of terabyte-range, and providing efficient big data processing through a non-volatile memory enabled hierarchical storage system, high speed networking interconnects, and multiple processors in parallel. Existing in-house or close proximity infrastructures are either outdated in terms of processing and I/O capabilities, impose limits on data storage, or they are not compliant to hold restricted ITAR data. The requested DLP-BD will significantly augment the current research capabilities of the PI and allow his team to conduct research in DL more efficiently. It will allow his team to design and train multiple networks using very large data and gain insights in hours instead of weeks or months. This will advance their knowledge and contributions to areas that are highly relevant to the US Army. The requested platform will also contribute to the education and training through research of several graduate students in disciplines important to DoD missions.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 09, 2020
- Source ID
- W911NF2010063
Entities
People
- Hichem Frigui
Organizations
- Army Contracting Command
- United States Army
- University of Louisville