Adaptive detection of stationary optical and IR targets using correlated scenes

The detection of stationary optical and infrared targets immersed in a heavy clutter background has become of increasing interest in recent years. Even if the target signature is completely known, standard correlation techniques generally fail due to the poor detection signal-to-noise ratio. Since velocity related techniques often cannot give sufficient detection solutions, new signal processing methods for clutter rejection are needed. ❧ In this dissertation, a new adaptive method for target detection is developed, which achieves clutter rejection by the use of correlated observations with lesser signal intensity. First, a basic optimal algorithm is developed for detecting a known target with dual correlated images. It is discovered herein that after the removal of a space-varying mean, the covariance matrix of many IR and optical images can be approximated by a diagonal matrix. This special structure is the key element in the derivation of several fast detection algorithms. These new algorithms furnish efficient solutions to the multiple correlated scenes (different frequency bands) problem as well as to the sequential imaging situation (recursive algorithm). ❧ Another important related problem is the detection of known targets in images distorted by impulse or glint-type noise. A new optimal decision rule is developed for images distorted by this kind of noise, which is called the disregard rule. The disregard rule ignores pixels in the detection process if their values (above or below the mean) exceed some computed threshold. This rule is compared to several other decision rules and filtering techniques by using the Asymptotic Relative Efficiency method. The new disregard rule was found to be superior to the other nonlinear decision rules including limiter and median filtering. ❧ Also discussed is the problem of non-negligible target intensity in the reference scene. An optimal algorithm is derived for this case and its expected loss is computed. It is shown that mismatching to the target intensity in the reference scene does not induce significant loss as long as the images are highly correlated. ❧ Finally the assumed Gaussian image model with a nonstationary mean and a more slowly varying covariance is tested in detail. The plausibility of this model was discussed before [2,3] using qualitative plotting techniques. In this dissertation, statistical test for goodness-of-fit are applied to test the accuracy of the Gaussian model. ❧ Throughout the work, computer simulations based on actual optical and IR image data served for checking the theoretical results. Simulation results generally agree well with the developed theory: The image model is verified for many IR and optical subimages and it is shown that the new detection algorithm yield significant improvement in the signal-to-noise ratio - up to 1 0 db and more-over conventional correlation methods. It is also shown that these algorithms can be used, in many cases, without significant loss even if the target intensity in the reference scene is not negligible. Also, the new disregard algorithm is shown indeed to be optimal for detecting a weak target in images distorted by impulse noise