The TOPS optical correlator program will demonstrate the capability for optical processors to perform target recognition and tracking tasks. This paper describes the demonstration to be performed. Operational considerations such as system complexity, reprogrammability, and throughput are discussed. Growth of these type systems and long term prospects are examined. 2. INTRODUCTION Optical processing is a maturing technology that has extensive application to processing of sensor information in real-time. The applications are often limited by the imagination of the process designer and his ease of use of the technology. The technology used in the application will, moreover, be replaced by newer technology in the near future. Thus, a gap has formed between the application and the processor developments. To bridge the gap the TOPS (Transition of Optical Processing to Systems) program will provide the user with explicit detail of example uses of optical processors that are built with off-the-shelf optical components. This will provide a snapshot in time of the capability and application of the optical processors. This will provide an update to close the application gap. The application stressed in this paper is pattern recognition. Pattern recognition and its uses in systems such as automatic target recognizers has been slow to mature. This is partly due to the limitations on the computation environment and not to concepts of the recognition process. Indeed the ideas of pattern recognition are understood at the functional as well as the operational level. The general problem is known to be very difficult from the computer science perspectivethe problem is a random problem. In general, then, good implementation solutions should have large complexity thus accommodating a wide range of problems from the simplest to the more complex. This is the promise of the optical processors. This capability must be achieved without large cost to the useri.e. small size, low power consumption , light weight, and low monetary cost. In this paper we will address these and other relevant issues using a military problem as a demonstration of the technology. Other pattern recognition uses are just as relevant as will be obvious to system designers from other problem disciplines. 3. PROCESSOR NOMENCLATURE The TOPS pattern recognition processor is a special purpose machine with interfacing through two ports. The ports are input spatial light modulator (SLM) devices and a detector array output devices, which are configured to provide ease of interface to a sensor at the input and digital 0-81 94-0866-2/921$4.OO SPIE Vol. 1701 Optical Pattern Recognition II! (1992) / 11 electronics at the output. The processor itself has an internal computation engine with program memory, scratch memory, and data memory. The processor engine is an optical correlator . The internal control and data flow is designed around this engine. To illustrated the control and flow we describe the basics of the correlator. The classical four focal length correlator is shown in Figure 1. The collimator portion of the system is equivalent to the power supply of the equivalent electronic system with the additional capability to switch the computation on/off by correspondingly switching the laser. The program store is both in the optics hardware, i.e. the Fourier lens, and the SLM memory, which is the memory of the SLM at the filter plane. The data memory for additional filters is a high speed electronic memory, which is controlled from a host system. This control allows the selection and sequencing from a bank of filters which are loaded to the high speed memory at system initiation. This host provides an operational environment that is easily reconfigured for different problems. 4. FUNCTIONAL CAPABTLITY Pattern recognition can be described as a vector space problem. Let (f} represent the class of patterns to be recognized. The correlation filter that recognizes these patterns is h. The correlation gives the relation I f(x+t)h(x)dx = with h=f then R(t)= e where e = f® f The functional operation achieved is that of the dot product, with the product being executed at each shift variable t. This operation is the basic operation of the popular form for artificial neural networks. Indeed the execution of a multilayered neural network by iteration through the TOPS correlator can be simulated. Thus, the pattern recognition system can execute the recognition of f by searching through the set {f) for the one that matches, i.e. gives e as an output. If (f} is large the comparisons for recognition become large and the problem must be partitioned. This partitioning can be achieved by the methods of the SDF. 5. DEMONSTRATIONPROBLEM The demonstration problem posed for the optical correlation based pattern recognition is one where an military target (tank) must be acquired, recognized, and tracked from an overhead, down-looking aspect. The scenario basics are shown in Figure 2, where the field-of-view (FOV) footprint is shown. The images of the target will have arbitrary rotation angle in the correlator reference frame. The closure will cause a scaling in the images and will in the end game, resulting in overfill of the FOV. This results in the processing filter constraints shown in Table 1. In addition to the FOV cases discussed, the FOV of the sensor is larger than the demo processor input spacebandwidth. This requires the searching of the FOV with a sequential search, one such method shown in Figure 3. This establishes an operational consideration that must be handled both by the input formatter and the filter control software. Finally, the system will switch modes as the terminal homing phase is reached. This is accomplished by simple control logic in the host. 12 / SPIE Vol. 1701 Optical Pattern Recognition III (1992) Operation of the system will entail certain costs, which will be directly related to capability limits. We will consider throughput, complexity, and programmability as our measures of performance capability and power, weight, volume, and hardware complexity as our costs. 6. PROGRAMMABiLITY Two issues are pertinent to the programmability namely the algorithm complexity problem and energy/system complexity cost to reprogram the system. These two issues are not optimized for this demonstration and are achieved by the following: filters are designed in an offline work station and has all the programming cost with the development of the filter software. The designed filters are stored in electronic memories and shipped to an SLM for execution, resulting in execution costs and additional energy from the electronic memories and the electronic communications. These costs will be high for more complex problems but are manageable for this demonstration.