A Novel and Fast Approach for Scalable Matrix Completion

In Phase I, we developed a novel matrix completion framework for recovering randomly missing pixels in images. Extensive simulations were performed using actual images from LADAR (18 images) and electro-optical images (23 images). We achieved high performance reconstruction even for missing rates as high as 99.9%. In Phase II, we will extend our algorithm to deal with missing data in hyperspectral images and missing links in social networks. Medium to high missing rates will be considered. The goal is to improve target detection and change detection performance by using hyperspectral images and to recover missing links in social network data. We will also address some theoretical issues in our algorithm such as how to choose the local window size. In addition, it should be noted that matrix completion algorithms are computationally intensive. So in Phase II, we will develop fast processing prototypes by using graphical processor unit (GPU), multi-core CPUs, and DSP. This will provide a number of fast processors for different applications. In the Phase II option, we will focus on a Navy application such as automatic target recognition (ATR) or social network. We will perform real-time or near real-time demonstrations and eventually integrate our prototype into a Navy program.