Abstract

Pattern recognition methodology is extremely important for robotics vision applications especially in the present era of automation. Perhaps one of the most well-known and important class of techniques of pattern recognition is the Vander Lugt matched filter correlator¹ and its related methods^2-3. In the optical implementations of the matched filter correlator, the technique involves the storage of the Fourier transform, via a thin lens, of the amplitude and phase of an image in a recording medium and later addressing the stored information by the Fourier transform of a new input. When the inverse Fourier transform of the multiplication of the two Fourier transforms are taken, cross-correlation between the new input and the stored is obtained. The cross—correlation intensity is a measure of the similarity between the two images. In the digital implementations, the Fourier transform operation is accomplished sequentially by electronics instead of the parallel transformation of a thin lens. Although the matched filter method is effective in recognizing an input image with the advantage of shift invariant, the question of whether the process emulates biological vision process is difficult if not impossible to answer.

© 1992 Optical Society of America