A new approach to optical information processing based on neural network models with application to object recognition

An optical information processing system based on simplified models of the Human Visual System (HVS) is studied. An associative Content Addressable Memory (CAM) model reintroduced recently by Hopfield as a simplified model of biological neural networks is analyzed and implemented using incoherent optical hardware. The parallelism and massive interconnectivity of neural net models makes optics an attractive means for their implementation. We study and experiment with several variations of the model: binarization of connectivity strengths, relaxation of the thresholding curve, dilute coded memory, and double layered hetero-associative memory. The binarization of connectivity strength or memory matrix is vital for optical implementations. A binary optical mask is much simpler to fabricate than a continuous gray scale mask, and can also be replaced by commercially available programmable binary spatial light modulators. The relaxation algorithm is introduced to simulate the asynchronous operation of a neural network efficiently on digital computers. The algorithm improves the performance of an artificial neural net when a unipolar binary memory matrix is employed. Dilute coded memory is found to greatly enhance the storage capacity of associative storage at the expense of reduced error correction capability. The architecture for optical implementation were extended to 2-D input and output vectors employing a partitioned 4-D memory matrix to make artificial neural networks more compatible with 2-D image and feature space data. The ability to use unipolar binary connections simplifies greatly electronic and opto-electronic realization of neural nets and paves the way for the implementation of layer nets. To overcome the hardware imperfections, a new optical mask generation technique, involving in situ fabrication, is also demonstrated. A simple scheme to incorporate double layered net structure in optical implementations was devised and demonstrated. Finally refinements and modifications in an existing concept of a Fourier Transform camera to perform generalized transforms and other operations useful in global and local feature extraction in natural scenes were described. The combination of the optical CAM and the Transform camera can provide a valuable research tool for the study of smart sensing and automated recognition as applicable to pattern recognition and computer vision.