Deep Learning for Radar Target Recognition

USA
The United States Air Force Research Laboratory

Presentation Menu

Abstract

The focus of this course will be recent research results, technical challenges, and directions of Deep Learning (DL) based object classification using radar data (i.e., Synthetic Aperture Radar / SAR data). First, we will present an overview RF ATR research in the past (i.e., template-based approach conducted under DARPA MSTAR (Moving and Stationary Target Acquisition and Recognition) program and limitations of this approach. Then will provide an overview of various machine learning (ML) theories applied to RF data. Finally, we will demonstrate implementations and performance analysis of DL-based ATR on SAR data using Amazon’s Web Service (AWS) cloud computing, Google Collaboratory GPUs, and/or TensorFlow. Datasets to be used include AFRL public released MSTAR, CVDome, SAMPLE and North Eastern University RF signals data.

Description

Objective

It is evident that significant research efforts have been devoted to applying DL algorithms on video imagery. However, very limited literature can be found on technical challenges and approaches to execute DL algorithms on radio frequency (RF) data. We will present hands-on implementation of DL-based radar object classification using PyTorch and TensorFlow tools. Unlike passive sensing (i.e., video collections), Radar enables imaging ground objects at far greater standoff distances and all-weather conditions. Existing non-DL based RF object recognition algorithms are less accurate and require impractically large computing resources. With adequate training data, DL enables more accurate, near real-time, and low-power object recognition system development. We will highlight implementations of DL-based (i.e., Convolution Neural Networks (CNN)) SAR object recognition algorithms in graphical processing units (GPUs) and energy efficient computing systems. The examples presented will demonstrate acceptable classification accuracy on relevant SAR data. Further, we will discuss special topics of interest (adversarial machine learning, transfer learning) on DL-based RF object recognition as requested by the researchers, practitioners, and students.

Learning Outcomes

  1. The student will understand various machine learning algorithms
  2. The student will be able to identify object features in radar imagery
  3. The student will be able to construct a machine learning system to detect and classify targets from radar imagery
  4. The student will be able to compare technical challenges involving radar and video image classification
  5. The student can differentiate benefits of DL-based RF object classification as compared to existing algorithms (i.e., template-based approach)
  6. The student would become aware of software tools and data applicable to their research interests

Intended Audience

Engineer/Researcher interested in applying Deep Learning on radar or electro-optical data for developing object recognition/self-driving car/autonomous/expert systems.

Course Level: The materials for this course will be intermediate; some understanding on machine learning will be useful but not a requirement.

Course Materials: I will be using our recently published (July 2020, Artech House) Textbook “Deep Learning for Radar and Communications ATR”.

Outline and time frame of the Short Course

Course Length: The course could be made for Half-day (4 hours of instruction) or Full-Day (8 hours of instruction).

Lecture Outline

  1. Radio Frequency ATR: Past, Present, and Future: 20 min
  2. Mathematics for Machine Learning / Deep Learning: 20 min
  3. Review of ML Algorithms: 30 min
  4. Deep Learning Algorithms: 30 min
  5. RF Data for ML Research: 15 min
  6. DL for Single Target Classification: 25 min
  7. DL for Many Targets Classification: 20 min
  8. RF Signals Classification: 15 min
  9. RF ATR Performance Evaluation: 25 min
  10. Emerging ML Algorithms for RF ATR: 35 min

Instructor Short Course History

Previous Offerings: I taught (on-site) this course at IEEE Radar Conference 2022 (NYC), 2021 at Atlanta (Video), RadarConf. 2019 (Boston). At Boston, I had the largest tutorial attendees (42 people) among all the tutorials. I also presented this tutorial at SPIE DCS 2018. The authors also presented a tutorial on “SAR Signal and Image Processing” at IEEE RadarConf 2010 (Washington D.C.). We also been selected for this tutorial offerings at RadarConf. 2020 (DC), and Tri-Service Radar Symposium 2020. I also provided four DL presentations. Currently, I am invited to offer this Short course at IEEE Bangalore, India (July 8, 2022).