Extended Functions of Multiple Instances for target characterization

摘要

An extension of the Function of Multiple Instances (FUMI) algorithm for target characterization is presented. FUMI is a generalization of Multiple Instance Learning (MIL). However, FUMI differs significantly from standard MIL and supervised learning approaches because only data points which are functions of class concepts are available. For applicability to hyperspectral data, this paper addresses the problem in which data points are convex combinations of target and non-target concepts. The presented method, eFUMI, extends previous methods to allow for further unspecificity in training labels while estimating target and non-target concepts, the number of non-target concepts, and the weight associating each concept to each data point. For eFUMI, training data need only binary labels indicating whether a spatial area in an input image contains or does not contain some proportion of target material; the specific locations or target proportions for training data are not needed. After learning the target concept, target detection can be performed on test data. Results showing sub-pixel target detection on simulated and real Hyperspectral data are provided.