Feature Extraction and Classification of FLIR Imagery Using Relative Locations of Non-Homogeneous Regions with Feedforward Neural Networks

The classification of forward looking infrared (FLIR) imagery is explored using Gabor transform decomposition and relative locations of non- homogeneous regions combined with feedforward neural networks. A feature saliency metric is developed from a Bayesian sensitivity analysis of feedforward neural networks. This metric is then used to reduce the dimensionality of the feature vectors used to identify FLIR imagery without any degradation of classification accuracy. Several system architectures are developed using a roving window combined with a series of Gabor filters to produce feature vectors for presentation to a neural network classifier. One architecture uses the Gabor filter coefficients to learn the gestalt of known images. Several of the gestalt networks are then combined to determine the class of an unknown image. Another architecture uses centroid metrics for the cluster of Gabor resonances to feed a backpropagation network acting as a traditional Bayesian classifier. A system architecture is developed which uses the relative locations of texture regions within a user defined template to classify imagery. The relative location architecture is shown to outperform traditional matched filter classifiers. The relative location architecture is shown to be robust in solving the problems presented by crepuscular lighting conditions. Image Processing, Gabor Transforms, FLIR Image Classification, Feature Saliency, Neural Networks.