Spatial multiple instance learning for hyperspectral image analysis

Abstract

Standard multiple instance learning (MIL) techniques are capable of learning when there is a lack of target information (including size, shape, and even location); however, this is attained at the cost of the utility of spatial information. This is unfortunate because in many image analysis applications, there is a substantial amount of observable spatial information. The research presented in the following investigates appropriate methods to incorporate spatial information into the MIL framework while maintaining the benefits of the MIL paradigm. The proposed Spatial Multiple Instance Learning (S-MIL) method is applied to a hyperspectral data set for the purposes of landmine detection.