Stottler Henke is developing the Automated Maritime Entity Recognition (AMER) system, which will be able to classify ships, by extracting features from the input data that is common to all ship classes such as size, speed, roll frequency, and silhouette profile. Through only featuring elements shared by all ship classes, the architecture will be able to incorporate transfer learning, as AMER 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. Our technology comes from our prior work in ship classification from ExPATSS and SIFTIC on ISAR imagery.
By utilizing transfer learning, AMER will be able to employ an ISAR Simulator to quickly and cheaply generate the large training data set needed for machine learning approaches. In the Phase I effort, AMER developed a hierarchical convolutional neural net that was able to achieve 100% accuracy across all levels of the classification hierarchy, 92% accuracy at the most fine classification level on unseen view angles, and 99% accuracy at the most fine classification level when introducing +/- 10m range and +/- 0.5Hz Doppler sensor inaccuracies to just the test data.
Disclaimer: This material is based upon work supported by the Naval Sea Systems Command under Contract No. N68335-20-G-1026. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of NAVSEA.