Initialized iterative reweighted least squares for automatic target recognition

Automatic target recognition is typically deployed on infrared focal plane arrays with high resolution, which could be costly. Due to the compressibility of infrared images, compressive sensing allows us to reduce the resolution requirements of a focal plane array while keeping the same target recognition ability. In this paper, we develop an iterative reweighted least squares algorithm with stochastically trained initial weights. Our simulations indicate that this method has higher automatic target recognition accuracy than conventional methods such as OMP, BP, and IRLS when applied to the U.S. Army Night Vision and Electronic Sensors Directorate (NVESD) dataset.