Feedforward Neural Networks for Image Classification

The application of artificial neural networks to image classification is investigated. A particular network architecture is considered: the three-level feedforward neural network. A specific configuration for image classification is described, in which each artificial neuron or unit in an input level of the network represents a pixel of a two -dimensional image. The concept of optimum binary coding is introduced, in which the classification of an image is coded in the units of a hidden level. This coding is shown to greatly reduce the computation required to run the network compared to a comparable two-level network. Computer simulations demonstrating the ability of three-level feedforward neuron networks to classify printed characters are reported. A new optical implementation of a feedforward neural network is described, in which the interconnection strengths between the input and hidden units of a three -level network are stored holographically. This approach benefits from the powerful abilities of optics to perform computational tasks on two-dimensional images quickly and in parallel. The classification of printed characters using a hybrid system, consisting of an optical processor and digital electronics, is reported. A new application of neural networks is reported: the classification of photon-limited images. The statistics of photon-limited images are analyzed, and a method of calculating the expected classification performance of a network is described. The theoretical results make it possible to predict the number of photoevents needed in a photon-limited image to achieve a given level of confidence in the classification. A photon-counting camera is used to acquire photon-limited images for input to a three-level feedforward neural network implemented in a microcomputer. Experimental results for the classification of printed characters are presented that agree with the theoretical predictions.