Computational imaging is a means of reducing the number of measurements needed to sense and reconstruct an image of a sparse scene, whether it be an image in space,^{1}^{,}^{2} time,^{3} or delay.^{4} One method for performing computational imaging is to measure the multiplication of the image of a scene with a series of projection masks. The inner product of the image and one of the projection masks constitutes a single measurement. For sparse scenes, the number of projection measurements required for highly faithful reconstruction can be significantly less than the number of pixels in the image.^{5} Typical compressive sensing^{5} approaches exploit the sparsity of a scene by making measurements on an orthogonal basis (i.e., random projection masks). Feature specific imaging (FSI)^{2} exploits sparsity in the object features of interest by directly measuring in the sparsity basis. Both are methods to reduce the number of measurements required for reconstruction. The goal of FSI is to faithfully reconstruct the full image of the objects in a scene, or at least the object features of interest in each scene, with a number of projection measurements that is significantly less than the number of pixels in the image. In FSI, there are multiple ways to choose a set of basis vectors relevant to a set of desired object features. One method is to develop a basis from a large number of training scenes. Note that the training set typically does not include the exact scene(s) to be sensed or the exact object features of the scene to be sensed, but is, instead, an extensive set of scenes that contain similar object features of interest. For example, a training set of faces with a wide range of different eyes, noses, mouths, and chins, with or without hats and glasses can be used to faithfully reconstruct the face of a person not used in the training set.^{6} Though the image of the actual face to be reconstructed is not in the training set of faces, it is assumed that the various features in the actual face are present in one or more of the faces used for training and, thus, all the critical features in the actual face can be reconstructed. From a mathematical standpoint, the utility of the training set is to capture the second-order statistics of the possible objects or object features of interest. The basis vectors are optimized and ordered to minimize the number of projections needed to reconstruct the objects or object features of interest in a scene. For example, supposing that the principle components of the training set were to be used for the measurement basis, one could choose the order of the principle components based on their respective eigenvalues and introduce compression by rejecting those principle components whose eigenvalues are below some threshold.