Refining Algorithms in Correlation Filter Design for Target Detection

In automatic target recognition (ATR), correlation filters are widely used to detect target signature variations. In this paper we concentrate on a particular case: target pose angle. For the traditional maximum average correlation height (MACH) filter method, only a few special angles can be used due to the limitation of the training data and the requirement on efficiency for real-time applications. In order to improve performance and save computation time, we propose an approximation approach to the filter designs. Based on a band-width assumption on the filter functions, we derive an optimal number of the filters required to achieve the same performance obtained with a much larger filter bank. Furthermore, we develop a refining algorithm for the filter designs based on the sinc-function approximation to the filter bank. This allows for a classification result over the continuum of a design parameter rather than the discrete possibilities represented by standard filter bank implementation. The filters we designed here are easy to compute and have good performance. Our method allows us to use the smallest number of MACH filters without losing performance to even gain fidelity in classification.