Abstract

We have developed an image classifier that uses data from a photon-counting camera as input to a feedforward neural network. The photon-counting camera consists of a photocathode, a stack of microchannel plates, a resistive anode, and position-computing electronics. Incoherently illuminated images are used, and the x- and y-coordinates of the photoevents are passed to a microcomputer. A fixed number of photoevents are collected and are input to a three-level feed­forward neural network, implemented in the microcomputer. The three-level network is trained beforehand using a modification of the backpropagation method, which takes into account the photon-limited nature of the input images. Experimental results are presented for the classification of printed alphabetic characters of different typefaces. Theoretical calculations are presented that describe the statistical nature of the photon-limited images and predict the classification performance of the neural network. Good classification results are obtained with <1000 detected photoevents. Some advantages of using photon-limited images for feedforward neural networks are discussed. One advantage is the network’s invariance to changes in object illumination level, which arises from the practice of collecting a fixed number of photoevents.

© 1991 Optical Society of America