AMER: High Accuracy Ship Target Classification from Multi-Modal Imagery

Stottler Henke is developing the Automated Maritime Entity Recognition (AMER) system to hierarchically classify ships in IR ISAR imagery.  AMER is designed to provide the U.S. Navy with an automatic target recognition system usable for both single target identification and for surveying multiple objects in a scene for situational awareness. 

If working from an ISAR image, AMER takes as input the ISAR magnitude and phase data and then fuses this data with the combined decibel scale conversion into a 3-channel image.  If working from an IR image, AMER uses a combination of computer vision techniques to segment the ship from the background and then classifies a silhouette mask of the target.

AMER combines a convolutional neural network with a hierarchical classification loss function to classify a data example across the different levels of the class hierarchy simultaneously.  In the Phase II effort across 23 fine ship classes, AMER has proven to be 100% accurate across all fine ship classes and it is also robust to noise, held out angles, and held out classes.

A major finding of the AMER project is the ability to use transfer learning to train a machine learning model on simulated data and test on real data.  By using computer vision to featurize both simulated and real data, and effectively projecting both data sets into the same frame of reference, the AMER system is able to classify real test data with 96% accuracy across the set of target ship classes when only trained on simulated imagery.