Overcomplete Dictionaries for Scene Classification in Coded-Aperture Spectral Imaging

I. INTRODUCTION Traditional spectral imaging sensors entail the acquisition of high-dimensional data that is used for the discrimination of objects and features in a scene. Recently, a novel architecture known as coded aperture snapshot spectral imaging (CASSI) system has been proposed for the acquisition of compressive spectral image data of a scene with just a few coded focal plane array (FPA) measurements [1]. In this work, a supervised classification approach of hyperspectral images directly from a small set of optimal CASSI compressive measurements, without first reconstructing the full datacube, is proposed. The classification method is based on finding a sparse representation of each observation pixel in an overcomplete dictionary learned directly from the training samples. It provides a more discriminative sparse representation of the datacube, and therefore a better classification accuracy can be obtained [2]. An estimated sparse vector is obtained from the optimal CASSI measurements by solving a sparsity-constrained optimization problem, and is then used to directly determine the class of the unknown pixel. Optimal CASSI compressive measurements are obtained when optimal coded apertures in the optical system are used [3]. The set of optimal coded apertures are designed such that the CASSI sensing matrix satisfies a Restricted Isometry Property (RIP) with high probability [4]. Simulations results illustrate the performance of the proposed classifier using overcomplete learned dictionaries and optimal coded apertures.