The last two decades of the last century were a very intensive period for the research in optical processing and optical pattern recognition. All the aspects of these processors were investigated and the research progressed remarkably.
One key element in an optical correlator is the reference filter, and important part of the research concentrated on it. The correlation is shift-invariant but is scale-variant and orientation-variant. Therefore several solutions, using for example, Synthetic Discriminant Function (SDF) were proposed to overcome this drawback . Beside the classical matched filter, several other improvements have been presented . A large amount of work has been carried out to enhance the discrimination of the target in a complex scene .
The architecture of the JTC was also studied extensively, particularly by Javidi who proposed several improvements such as the nonlinear JTC.
A very large number of processors were constructed taking advantage of the progress of SLMs and of the theoretical work on the filters and on the architectures. Some of these processors stayed in the laboratories while some others were tested for real applications. Regarding the large number of optical processors that were constructed during this period of time, it is impossible to list them all in the frame of this paper. A book, written in French, by Tribillon gives a very complete state of the art of the optical pattern recognition in 1998 . The book edited in 1999 by Yu and Yin give also a complete overview on the topic . Therefore, you will find here only, some examples of the optical processors developed between 1980 and 2004.
In 1982, Cleland et al. constructed an optical processor for detecting tracks in a high-energy physics experiment. This incoherent processor was using a matrix of LEDs as input plane and a matrix of kinoforms as processing plane. It was used successfully in a real high-energy physics experiment in Brookhaven .
The Hough transform is a space-variant operation for detecting the parameters of curves . This transform can take fully advantage of the parallelism of the optical implementation. In 1986, Ambs et al. constructed an optical processor based on a matrix of optically recorded holograms . This implementation was improved ten years later with the use of a large scale DOE composed of a matrix of 64 by 64 CGHs with 4 phase levels fabricated by lithographic techniques . Several other optical implementations of the Hough transform were published. Casasent proposed several different optical implementations for example one using an acousto-optics cell . A coherent optical implementation of Hough transform has been discussed by Eichmann and Dong , where the 2D space-variant transfer function is implemented by successively performing 1D space-invariant transforms by rotating the input image around its center point and translating a film plane for recording. Another implementation for coherent or incoherent light was proposed by Steier and Shori where they use a rotating Dove prism to rotate the input image, and the detection is achieved by a linear detector array. Today the Hough transform is widely used in image processing for detecting parametrical curves, but the implementation is electronic.
Yu et al. proposed several optical processors for pattern recognition using different types of input SLMs . For example an adaptive joint transform correlator for autonomous real-time object tracking , an optical disk based JTC .
Pu et al. constructed a robot that achieved real-time navigation using an optoelectronic processor based on a holographic memory .
Thomson-CSF in France, in the frame of a European project, constructed and tested successfully a compact photorefractive correlator for robotic applications. The size of the demonstrator was 600 m00 mm, it was composed of a mini-YAG laser, a liquid crystal SLM and an updatable holographic BSO crystal . This correlator was also used for finger print identification .
Guibert et al. constructed an onboard optical JTC for real-time road sign recognition that was using a nonlinear optically addressed ferroelectric liquid crystal SLM in the Fourier plane .
A miniature Vander Lugt optical correlator has been built around 1990 by OCA (formerly Perkin-Elmer). This correlator was composed of a Hughes liquid crystal valve, a set of cemented Porro prisms and a holographic filter. The purpose of this processor was to demonstrate this technology for autonomous missile guidance and navigation. The system was correlating on aerial imagery and guided the missile to its preselected ground target. The processor was remarkable by its rugged assembly; it was 105 mm in diameter, 90 mm long, and weighted 2.3 kg .
In 1995, OCA constructed a prototype of an optical correlator that was fitting in the PCI slot of a personal computer and was able to process up to 65 Mbyte of image data per second . This processor was intended to be commercialized.
The Darpa in the USA launched in 1992 a project named TOPS (Transitioning of Optical Processing into Systems) associating some universities and about ten important companies potential users and developers of the technology.
BNS presented in 2004 an optical correlator using four kilohertz analog spatial light modulators. The processor was limited to 979 frames per second by the detection camera. However, the rest of the correlator was capable of 4,000 frames per second .
The Jet Propulsion Laboratory (JPL) developed several optical processors for real time automatic target recognition . The University of Sussex constructed also an all-optical correlator and a hybrid digital-optical correlator .
It should also be noted that several optical correlators were available commercially but it is not sure that it was a commercial success since most of them are no longer commercialized. For example, in 2000, optical correlators were commercialized by INO and BNS . In 2001, Parrein listed in her PhD thesis 10 optical correlators that were available .
Optical processors were also designed for many other operations such as matrix operations , or for systolic array processing and neural network processors .