A close relationship between light propagation and processes involved in communications became evident in the 1950s and was described in several publications. These publications demonstrated how communications theory, mainly Fourier analysis, could be applied to optics. Although most of the information acquired and processed by an average human being reaches us in an optical form, the complementary process of using optics for signal manipulation had to wait until the invention of a source for coherent light radiation-the laser.
In 1960, L.J. Cutrona et al published a paper on optical data processing and filtering systems that demonstrated the computing power of optical systems. However, for many of us, the real beginning of the era of optical signal processing and computing was marked by B. VanderLugt in 1963, when he introduced the optical correlator containing the spatial matched filter. The optical correlator is based on the fact that a lens generates, over the focal plane (Out), the Fourier transformation of a transparency placed over the input plane (In) and illuminated by coherent light. If a spatial filter (another transparency) is placed over this Fourier transform plane and another Fourier transformation performed using a second lens, you obtain, over the new output plane, the correlation between the input function and the filter function. VanderLugt demonstrated the power of the double Fourier transform system to implement a pattern recognition task.
The success of the VanderLugt correlator raised much interest and extended the technology to several applications. As a result, an intensive research effort was initiated and this research activity continued for about two decades. Unfortunately, there were two main obstacles that prevented the achievement of the expected goals. First, existing optical information was carried by incoherent light that had to be converted into a coherent image before it could be processed by the optical correlator. No satisfactory interfaces were available for this incoherent-to-coherent transformation except photographic film with all the lengthy chemical processing involved. The same problem was faced by information that was already in electronic form and had to be converted into light signals to be processed optically.
Second, the whole concept of the optical correlator was based on the matched filter borrowed from communication theory. The matched filter is optimal for detecting a signal immersed in additive, white Gaussian noise. While a good approximation for temporal signals is encountered in communication, the noise in coherent optical systems depends strongly on the signal and does not satisfy these conditions. In addition, the objective of an optical correlator is more frequently the discrimination among objects that may have many similar features, rather than the detection of a given signal immersed in additive noise.
Moreover, optical images are derived from three-dimensional objects, in contrast with the one-dimensional strings encountered in communications. Three-dimensional objects may undergo significant changes in appearance due to rotations around different axes or changes in scale due to the variation in their observation distance. The situation may be complicated even further if real, physical distortions of the objects are allowed, as is the case in handwritten character recognition.
Over the years, many procedures were developed to mitigate the fundamental difficulties indicated above. In particular, highly sophisticated filters were invented to replace the classical matched filter, and new correlator architectures were devised. Unfortunately, the progress in interface technology did not satisfy the demands, and the impressive academic achievements remained mainly in the academic domain.
Frustrated with the declining interest in optical correlators and other analog processors, many scientists turned to the investigation of possible implementations of digital computing with optics. Here, too, the speed of light and the high parallelism of optics were expected to be of great advantage and the concept of all-optical computing was introduced. But several issues were overlooked. Massive parallelism can be obtained because light does not interact with light. As a result, the most important aspect of digital processes-the control of one signal by another signal-cannot be realized with optical signals unless some interface (material or electronic) is involved.
Although a large number of optical channels can be operated in parallel, this number is limited by diffraction in uniform media and free space. If guided interconnections are used (channel waveguides or optical fibers) the limitation comes from the number of physical guides that can be accommodated and interconnected within a given region of space. It is also noteworthy that, in spite of the high speed of light, it takes a finite time for light to cross an optical system. This time delay, called pipeline delay, is quite large (of the order of nanoseconds) but can be tolerated in most applications, even in the promoted high-speed computing. However, as was shown by this author in 1987, in a regular optical system, there exists also a differential spread in the effective lengths of the various optical channels between the input and the output of the system. This differential spread leads to a very disturbing time skew that may become significant at bit rates on the order of 10 Gbits/s.
As the difficulties encountered by optical processors were being realized, digital electronics leaped forward, leaving optical technology far behind. In practical terms optical processing could not compete with electronics, and interest in optical signal processing declined. Nevertheless, the fact that nature exploits optics so extensively indicates that optics must have some attributes that cannot be matched by other media. One of these attributes is the capability of photons to solve the wave equation with any given boundary conditions. Moreover, this wave equation is solved almost instantaneously, in parallel, and with no expense of energy. Energy is dissipated only at the moment when the final result is detected. In contrast, digital computers dissipate energy for each intermediary step of a calculation, even if those intermediary calculation results are of no interest. As pointed out by H.J. Caulfield and this author, this attribute alone is an adequate incentive to pursue optical signal processing. Furthermore, with all the limitations indicated above, the parallelism available with optics is still several orders of magnitude above what can be expected for electronics in the foreseeable future.
In recent years there's been a revitalization of optical processing in a much more mature way. Recognizing the limitations and attributes of optics, researchers realized that optics should not be required to do the job alone, and optics' attributes must be complemented with those of digital electronics in sophisticated ways. For example, optics can be used for linear processes at high speed with massive parallelism and low expenses in energy, while the nonlinear calculations can be implemented digitally by electronic computers.
A good example is the advanced hybrid electro-optical correlator, which was introduced in 1990 by J. Rosen, U. Mahlab and this author, to implement complicated learning and optimization algorithms. In these systems, optics perform the correlations as parallel linear processes, while all nonlinear processes and calculations are implemented electronically in a digital computer. Containing feedback loops (electronic as well as optic), these systems continuously monitor their own performance. As a consequence, they can be trained to cope with a wide selection of problems under variable circumstances, including compensation for imperfections that may exist in any real system.
It is also interesting to note that the increasing speed of digital computers became now an attribute for the hybrid processors since in these systems the digital part is still the bottleneck.
Technologically, hybrid processing is being made practical by the rapid progress in different areas. Display technology enables the development of advanced spatial light modulators that serve as the electronic-to-optical interface. Detector technology provides the essential optical-to-electronic interface. The development of large surface-emitting laser diode arrays provides a way to control light signals by using light signals.
The concept of all-optical digital processing deserves serious reconsideration. In digital processing, optical interconnections will play an important role, and devices such as smart pixel arrays, where the basic digital processing is implemented in each pixel, will become the workhorse of image processing for artificial vision.