TECHNOLOGY AREA(S): Air Platform, Information Systems

ACQUISITION PROGRAM: PMA-262 Persistent Maritime Unmanned Aircraft Systems

OBJECTIVE: Develop innovative techniques that combine the robustness and explainability of model-based target identification and classification approaches with the potential to optimally establish feature vector coefficients using data-driven, deep-learning artificial-intelligence approaches.

DESCRIPTION: The most robust state-of-the-art target classification and identification approaches rely on expert knowledge and model driven principles to mimic the methods used by expert human operators in manual target classification and identification. These approaches rely on the identification of a set of target features that allow one target to be confidently separated from other, different targets. This STTR topic seeks to evolve from classification ranking systems based on a distance metric with pre-assigned feature weights to a more optimal classifier in which the selection of features and their weights are determined automatically from a statistical analysis of the collected feature vectors data for each sensor. Such an approach provides the rationale for feature selection and weighting as it is based on measured feature variance vector for each sensor. Furthermore, the statistical analysis of collected feature vectors enables the calculation of accurate matching confidence within the classifier. This provides the needed explainability to quantify to what extent the operator should trust the classification results. The efficacy of the approach should improve with the accumulation of operational data. Sensor systems to be considered are imaging radar, electro-optics, imaging infrared, and electronic support measures. Applications include maritime vessel and overland vehicle classification and identification.

PHASE I: Assess feasibility of incorporating deep-learning approaches into target classification and identification beginning with a statistical analysis to determine the relative importance of features derived from templates and progressing toward the determination of feature vectors from sensor data. Outline methods and determine feasibility of evolving from a traditional approach of matching measurements to pre-defined features and weights, incorporating data-driven, deep-learning approaches to determine feature vectors from field data. This would include the association of field feature measurements for each sensor with database templates, registration and statistical analysis of all sensor feature measurements, determination of relative importance of features derived from templates, and deep-learning methods for determination of feature vectors from field data. Finally, develop an architectural plan for in-flight, operator-assisted database augmentation. Develop prototype plans to be developed under Phase II.

PHASE II: Further mature Phase I-developed algorithms and architectures for integration into the target classification and identification application. Plan requirements for and conduct data collection to support algorithms and software testing.

PHASE III DUAL USE APPLICATIONS: Perform any required modifications to the algorithm and real-time code to be hosted in the transition Program of Record as desired by the Navy. Support modeling and simulation efforts as well as software integration, field testing, and performance analysis in the specific application. The application will need to adapt to the sensor capabilities and interfaces on the specific aircraft type. Ultimately, the goal is to correctly classify a ship to find naval class from amongst all of the combatants and non-combatants of the world. Non-naval vessels can be classified to general type. Maritime activities such as in the U.S. Coast Guard, shipping monitoring, and the Department of Homeland Security—any officials that have the need to know what ship traffic exists—can benefit from this technology. The basic core of the algorithms and fusion may apply to land-based commercial vehicle tracking as well.

REFERENCES:

1. Geng, J. et al. “Deep supervised and contractive neural network for SAR image classification”. IEEE Transactions on Geoscience and Remote Sensing, Volume 55, Issue 4, April 2017. http://ieeexplore.ieee.org/document/7827114/

2. Mason, E., Yonel, B., and Yazici, B. “Deep learning for radar”. Radar Conference (RadarConf) 2017 IEEE, pp. 1703-1708, 2017, ISSN 2375-5318, http://ieeexplore.ieee.org/document/7944481/

3. Nguyen, A. et al., “Deep neural networks are easily fooled: High confidence predictions for unrecognizable images”. Computer Vision and Pattern Recognition (CVPR), IEEE, 2015. http://www.evolvingai.org/files/DNNsEasilyFooled_cvpr15.pdf

4. Zhang, L., Zhang, L., and Du, B. “Deep learning for remote sensing data: a technical tutorial on the state of the art”. Geoscience and Remote Sensing Magazine IEEE, Vol. 4, pp. 22-40, 2016, ISSN 2168-6831. http://ieeexplore.ieee.org/document/7486259/

KEYWORDS: Artificial Intelligence; Deep Learning; Data Driven; Expert System; Maritime Classification; Radar