Automatic Labeling of Multiple Target Synthetic Aperture Radar (SAR) Imagery for Automatic Target Recognition (ATR)

OUSD (R&E) MODERNIZATION PRIORITY: Artificial intelligence/machine learning (AI/ML), autonomy TECHNOLOGY AREA(S): Information systems; sensors; electronics OBJECTIVE: Develop novel algorithms for labeling multiple target classes in Synthetic Aperture Radar (SAR) imagery to expedite training of SAR Automatic Target Recognition (ATR) algorithms. DESCRIPTION: The focus of this research will be the automatic labeling of multiple-target target classes in SAR imagery for deep learning based SAR ATR. A critical first step for AI/ML-based target classification involves providing a large amount of labeled data to train deep neural networks (DNN). As of now, there is no automated approach to label the training data (i.e., multiple target input SAR imagery). Currently, after SAR data collection and image formation, data labeling is conducted manually. As a result, the development and deployment of AI/ML-based algorithms can be greatly delayed. For each data collection or mission, labeling thousands of images manually is costly in terms of time and money. Hence, research should be conducted to develop novel algorithms to expedite the labeling process. Currently, some research efforts attempt to apply active learning techniques to label single targets in SAR imagery (i.e., an image chip). One approach is a graph-based technique that labels a few images covering multiple target types, learns the features, and applies these features to label unlabeled imagery. This technique, sSemi-supervised learning (SSL), shows some success. Researchers also tried other approaches such as Core-set to label non-radar imagery. The critical issue is that if a SAR image contains multiple targets multiple targets, which can be vehicles, bright vegetation, buildings, or unknown clutter, making labeling each detected object a difficult problem. Moreover, if the targets are heterogeneous in size, detection and labeling is further complicated. The goal of this research is to develop automated labeling algorithms that can label multiple classes of target without a human in-the-loop. Research in this area requires significant experience in SAR imaging (2D, 3D imaging), clutter reduction, constant false alarm rate (CFAR) based detection, region-based detection, only-look-once detection, and deep learning-based target classification. An understanding of SAR datasets such as the AFRL Gotcha radar data, Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset, and implementation of various deep learning techniques to MSTAR datasets will also be helpful. As needed, the government will work with the performer to find relevant synthetic and measured datasets. PHASE I: Research, develop, and demonstrate an automated algorithm framework to label multiple target classes in radar imagery. Provide a baseline technique to automatically label multiple target classes in the open (separable targets surrounded by benign clutter such as cut grass, desert, etc.) in SAR imagery. PHASE II: Implement automated SAR image labeling for complex targets scenarios. Evaluate performance of labeling and performing ATR on targets embedded in complex clutter (e.g., trees, etc.) using relevant datasets. Demonstrate an end-to-end ATR system that include the automated labeling of multiple target classes in SAR imagery, training, testing and ATR classification. PHASE III DUAL USE APPLICATIONS: Adapt algorithms from Phase II to other mission relevant datasets. REFERENCES: Z. Meng, E. Merkurjev, A. Koniges, and A. L. Bertozzi, Hyperspectral image classification using graph clustering methods, Image Processing On Line, vol. 7, pp. 218 245. 2017. K. Miller, H. Li, and A. L. Bertozzi, Efficient graph-based active learning with probit likelihood via Gaussian approximations, arXiv preprint arXiv:2007.11126, 2020. C. Garcia-Cardona, E. Merkurjev, A. L. Bertozzi, A. Flenner, and A. G.Percus, Multiclass data segmentation using diffuse interface methods on graphs, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 36, no. 8, pp. 1600 1613, 2014. U. K. Majumder, E. P. Blasch, and D. A. Garren, Deep Learning for Radar and Communications Automatic Target Recognition. Norwood, MA, USA: Artech House, 2020. O. Sener, Silvio Savarese, Active learning for convolutional neural networks: A core-set approach, ICLR 2018. KEYWORDS: Automatic target recognition (ATR), synthetic aperture radar (SAR), data labeling, active learning, deep learning, artificial intelligence/machine learning (AI/ML), constant false alarm rate (CFAR)