SPECIALIZED HARDWARE

For various reasons it may be faster or more convenient to do some of the computations in pattern recognition with specialized hardware. Here are some special cases.

COHERENT OPTICS

Using laser illumination and appropriate optics, we can perform both the forward and the backward Fourier transform optically. So the output is an image of the input. That leaves us half way in between with a spatial display of the Fourier transform. We can insert a physical mask in that plane to do recognition that is invariant to input image shifts. Under some circumstances, this may offer advantages over doing the transformations and filtering in a digital computer using software. We have done much work in this field, but the people who use our work most often do so in the digital domain. Here are some summary publications. For the latest advance see the discussion in training methods.

• Selected Papers on Fourier Optics, SPIE Milestone Series, Vol. Ms. 105, 610 pps., 61 papers, edited by Mustafa A. G. Abushagur and H. John Caulfield (1995)
• "Optimization Methods for Pattern Recognition," Optical Pattern Recognition, Joseph L. Horner & Bahram Javidi, Eds., SPIE Optical Engr. Press, Bellingham, Washington, (1991), J. Shamir, Joseph Rosen, Uri Mahlab and J. Caulfield

DSPs – DIGITAL SIGNAL PROCESSORS

These are specialized hardware on a special chip. Because that chip can do manu useful subroutines very fast, it has many uses. http://www.amazon.com/exec/obidos/ASIN/0070540047/qid=1019241831/sr=8-3/ref=sr_8_67_3/103-7785010-9361431

SPECTRAL CORRELATOR

This is one case in which optics has a clear and dramatic advantage over electronics (Computer or DSP). We do the discrimination in the optical domain before detection. We simply detect the answer. There are profound advantages including these:

• One detetction event replaces 100s, because no spectrum ever need to be computed. This means that we can use a very sensitive detector rather than an array of less sensitive one. But, more importantly, we do not need to share light with muutiple detectors. For a given integration time, our signal becomes much greater as does the signal-to-noise ratio. It also means that the data can be taken very fast. Applying the kinds of sophisticated pattern recognition methods described here should allow us to achieve the discrimination capability of hyperspectral imaging in a system that is much cheaper, much more sensitive, much faster, much more robust, and much less expensive than a hyperspectral imaging system.
• The spectrometer part of the system does not need to be as good (and expensive) as it would in a conventional system for the same purpose. We train the filter on adjacent pieces of the spectral display, but we do not interpret thes as spectra. They are just numbers.

The founding papers of this field follow.

· "Direct Optical Computation of Linear Discriminants for Color Recognition," Opt. Eng., Vol. 23, No. 1, pg. 16‑19 (1984) H. J. Caulfield and P. F. Mueller.

· "Spectral Correlator Technologies," Opt. Eng., Vol. 24, No. 6, pg. 996 (1985) D. S. Frankel, M. Camac, H. J. Caulfield, C. Gozewski, and C. E. Kolb.