ABSTRACT
Matched filter [1,2] and many of its variations [3-11] have been proposed for image recognition. In image recognition, it is often desired to design filters or algorithms that are invariant to deterministic distortions of the target and are robust to input noise. This means that the filters need to recognize a target that is contained in the noisy input scene, and this target may be a distorted version of the reference target. For instance, the target may be a rotated or a scaled version of the reference target placed in background noise. These filters are referred to as distortion tolerant filters. The most common approaches to design these filters are based on training the filter with a set of distorted versions of the reference targets [12-16]. What distinguishes such filters from one another is the metric or method used to optimize the filter.
