Optical Neural Network Classifier Architectures

Abstract

We present an adaptive opto-electronic neural network hardware architecture capable of exploiting parallel optics to realize real-time processing and classification of high-dimensional data for Air Force Hostile Target Identification (HTI). This architecture utilizes a grayscale-input radial basis function neural network based on a previously demonstrated binary-input version. The greyscale-input capability broadens the range of applications for the classifier by allowing it to handle 8 bit input data. We characterized a key component of this system, a variable phase retarder, and found that the phase uniformity changed less than 7% with applied voltage. An optical wavelet transform preprocessor is also discussed. The preprocessor produces a reduced feature set of multiwavelet images to improve training times and discrimination capability of the neural network. The design uses a joint transform correlator (JTC) to provide cross correlations of multiple input images. We present experimental results for a JTC which used four input images generated with a spatial light modulator. We then propose using wavelet functions as input images to perform a multiwavelet feature extraction. The results from the retarder characterization and optical wavelet transform work were to be used in a software simulation of the neural network system to determine its feasibility. However, this work remains unfinished as this project was canceled due to budget cuts.