Target detection in infrared and SAR terrain images using a non-Gaussian stochastic model

Automatic detection of targets in natural terrain images is a difficult problem when the size and brightness of the targets is similar to that of the background clutter. The best results are achieved by techniques that are built on modeling the images as a stochastic process and detection as a problem in statistical decision theory. The current paper follows this approach in developing a new stochastic model for images of natural terrain and introducing some novel detection techniques for small targets that are based on hypothesis testing of neighborhoods of pixels. The new stochastic model assumes the observed image to be a pointwise transform of an underlying stationary Gaussian random field. This model works well in practice for a wide range of electro-optic and synthetic aperture radar (SAR) natural images. Furthermore the model motivates the design of target detection algorithms based on hypothesis tests of the likelihood of pixel neighborhoods in the underlying Gaussian image. We have developed a suite of detection algorithms with this model, and have trialled them on ensembles of real infra-red and SAR images containing small artificially inserted targets at random locations. Receiver operating characteristics (ROCs) have been compiled, and the dependence of detection statistics on the target to background contrast ratio has been explored. The results show that for the infrared imagery the model-based algorithms compare favorably with the standard adaptive threshold detector and the generalized matched filter detector. In the case of SAR imagery with unobscured targets, the generalized matched filter performance is superior, but the model-based algorithms have the advantage of not requiring prior information on target statistics. While all algorithms have similar poor performance for infrared images with low contrast ratios, the new algorithms significantly outperform existing techniques when there is good contrast. Finally the advantages and disadvantages of applying such techniques in practical detection systems are discussed.