Information Theoretic Detection of Objects Embedded in Cluttered Aerial Scenes.

摘要

This dissertation presents the theoretical development and numerical implementation of a minimum cross-entropy target detection algorithm. The procedure is based on the solution of a nonlinear constrained cross-entropy minimization problem and requires information in the form of raw image moments. The detection rule involves both preprocessing and real-time computations. The preprocessing requires the selection of a set of target templates and the solution of the constrained cross-entropy minimization problem for the selected target templates. The real-time processing requires the computation of image moments and a set of dot product operations. An orthonormal set of information functions is developed and numerical methods of converting raw image moments into the expected values of the information functions are given. Numerical techniques for image moment computation and a solution scheme for the nonlinear set of constraints are developed and implemented. The procedure is applied to a test set of 100 images and the detection algorithm error probability is projected and related to the salient performance determining factors.