Kin-Weng Chan

Date of Award


Document Type


Degree Name

Master of Science


Department of Electrical and Computer Engineering

First Advisor

Steven C. Gustafson, PhD


Multicolor infrared imaging missile-warning systems require real-time detection techniques that can process the wide instantaneous field of regard of focal plane array sensors with a low false alarm rate. Current technology applies classical statistical methods to this problem and ignores neural network techniques. Thus the research reported here is novel in that it investigates the use of radial basis function (RBF) neural networks to detect sub-pixel missile signatures. An RBF neural network is designed and trained to detect targets in two-color infrared imagery using a recently developed regression tree algorithm. Features are calculated for 3 by 3 pixel sub-images in each color band and concatenated into a vector as input to the network. The RBF network responds with a value of unity to feature vectors representing missiles and with zero to vectors representing background. Images are thresholded prior to application to the trained RBF network to narrow the field of interest of the RBF network and increase missile detection speed. The RBF network-based technique then generates potential target locations and probabilities that the locations correspond to missiles. Results show that the RBF network-based technique operates in near teal-time and detects 100% of the missiles in data that was not used in training Receiver operating characteristic (ROC) curves show that overly high classification thresholds can exceed the RBF network response for a true missile and result in non-detection. However, these ROC curves also show that adaptive control of the classification threshold on the RBF network output can reduce the number of false alarms to zero.

AFIT Designator


DTIC Accession Number