Optical systems shine at pattern matching

Database-searching techniques are currently used in many security applications. But searching image databases is not trivial. Not only are images large, they also require that computationally intensive operations be performed on them. Traditionally, a good way to perform pattern recognition with images has been to apply the Fourier or fast Fourier transform (FFT).

Optical systems shine at pattern matching

Database-searching techniques are currently used in many security applications. But searching image databases is not trivial. Not only are images large, they also require that computationally intensive operations be performed on them. Traditionally, a good way to perform pattern recognition with images has been to apply the Fourier or fast Fourier transform (FFT).

In such methods, images to be compared are transformed into the spatial-frequency domain by the FFT. After the power spectrum of the FFT is determined, an inverse FFT provides correlation information that indicates whether the two images match. Because the FFT is a repetitive algorithm, it is well suited to single -instruction multiple-data (SIMD) architectures.

Despite this, researchers at the National Research Council?s Institute for National Measurement Standards (Ottawa, Ont., Canada) have found that when compared to optical systems, SIMD architectures are still orders of magnitude slower (see table). In their research, Simon Boothroyd and his colleagues benchmarked an optical spatial light modulator (SLM) system, a DECmpp 12000 SIMD machine from Digital Equipment Corp. (DEC; Maynard, MA), with 211 processors and an AlphaServer 1000 from DEC running at 233 MHz.

Using the SLM, target and reference images were input simultaneously and their joint Fourier transform processed optically to identify similar patterns between the two images. On the DECmpp 12000 SIMD machine, the FFT algorithm was programmed as a callable subroutine. On the AlphaServer, a complied C matrix program from The Math Works? (Natick, MA) MATLAB package was used. To test each of the machines, 8-bit gray-scale images were sequentially compared to a database of images.

OBecause the computation time for the Fourier transform and image correlation is independent of the size of the input image, the optical system performed significantly faster than the DECmpp processor for large images,O says Boothroyd. OAnd, although SIMD machines with 214 processors are available, they would still fall below the performance of the optical system.O

The present optical system operates at 30 matches per second, but is being increased to 1000 frames per second, according to Boothroyd.

More in Boards & Software