STOTTLER HENKE ASSOCIATES, INC — Department of Defense SBIR Phase I: SCO182-008

Description

Stottler Henke proposes the Naval ISAR Target RecOgnition (NITRO) system, leveraging our related past work in ship classification from ISAR imagery. NITRO will be able to maintain low space, weight, and power (SWaP) constraints by utilizing recent advancements in mobile processors, such as AMD’s APU chips (a single chip CPU-GPU hybrid) or Intel’s CPU-FPGA hybrid chips. NITRO will incorporate the fruits of our previous Navy development efforts ExPATSS and SIFTIC in order to classify ships. NITRO will extract features from the input data that is common to all ship classes, such as size, speed, roll frequency, and silhouette profile. By only featurizing elements shared by all ship classes, the architecture will be able to incorporate transfer learning. NITRO will classify the incoming ISAR data stream and extracted features using the computer vision technique “part-based models” coupled with deep learning through a convolutional neural net. These techniques have both been implemented and applied to ISAR images in Stottler Henke’s ExPATSS and SIFTIC projects. By utilizing transfer learning, NITRO will be able to employ an ISAR Simulator to quickly and cheaply generate the large training data set needed for machine learning approaches.