Abstract

Past work in optical pattern recognition of rotated targets emphasized smeared Fourier plane filters or synthetic discriminant functions. These are reduced information content filters. Our aggregate phase-only filter detects multiple targets whose Fourier transforms have minimal overlap, which utilizes unoccupied areas of the Fourier plane and which reduces smearing. This retains effective target detection for all targets. Using multiple targets which are rotated versions of the same image, we demonstrate that an aggregate filter formed from widely rotated targets performs better than one formed from slightly rotated versions. The strong overlap of the Fourier transforms of the similar targets leads to choosing compromise values for the phase in the Fourier plane filter. The Fourier transforms of the dissimilar targets had less overlap and permitted choosing the phase of each pixel in the detection filter to match only one of the targets. The correlation peak amplitude has also been predicted as a function of overlap in the Fourier plane for the aggregate filter with two, three, four, and five simultaneous targets. Rotation invariance is obtained by using several aggregate filters in parallel, each of which detects several widely separated angular positions of the target.

© 1988 Optical Society of America