Advanced Machine Learning Fusion of Radar and EO/IR/LIDAR for Ship Classification and Identification

We propose a two-prong machine learning approach that simultaneously uses two complementary techniques, deep learning CNN and manifold learning, to exploit the automatic feature and regularities discovery of deep learning to fuse the multiple sensor data and the sparsity representation of the data in manifold learning to fuse the raw sensor data as represented by their highly compressed lower dimensional manifolds. This two-prong approach, combines with the baseline handcrafted features used to augment the features discovered by the deep learning CNN algorithm, will provide unprecedented robust ship classification and potentially identification performance. For operationally utility, we will leverage industry commercial off the shelf (COTS) multi-core graphical processing units (GPUs) processors such as those already developed by NVIDIA and Intel specifically for deep learning implementations. Moreover, by combining the mathematical concept underlying manifold learning and compressive sensing, the multiple sensor data can be represented, fused, and used to classify ships by way of their class-specific coefficients (i.e., lower dimensional manifold) without loss of information. The sparseness of the data essentially allows lossless compression by greater than 90% allowing the compressive manifold algorithm to be designed for execution on low-power mobile processors such as GPUs or field-programmable gate arrays (FPGAs).