Pattern recognition with composite correlation filters designed from noisy training images

Abstract

Correlation filters for target detection are usually designed under the assumption that the appearance of a target is explicitly known. Because the shape and intensity values of a target are used, correlation filters are highly sensitive to changes in the target appearance in the input scene, such as those of due to rotation or scaling. Composite filter design was introduced to address this problem by accounting for different possibilities for the appearance of the target within the input scene. However, explicit knowledge for each possible appearance is still required. In this work, we propose composite filter design when an object to be recognized is given in noisy training images and its exact shape and intensity values are not explicitly known. Optimal filters with respect to the peak-to-output energy criterion are derived and used to synthesize a single composite filter that can be used for distortion invariant target detection. Parameters required for filter design are estimated with suggested techniques. Computer simulation results obtained with the proposed filters are presented and compared with those of common composite filters.