Pattern recognition algorithms for target detection in the presence of environmental degradation of the input image

Date of Completion

January 1999

Keywords

Engineering, Electronics and Electrical

Degree

Ph.D.

Abstract

The presence of turbulence and aerosols between the target and the observer degrades the detection and classification performance of electro-optical sensors and detectors. The first key contribution of this dissertation is the development of a filtering algorithm that takes into account the environmental degradation, the background nonoverlapping noise, the non-stationarity of the scene, non-target objects and additive system noise. In this model, both the background noise that is spatially nonoverlapping with the target and the noise that is additive to the target and the input image are considered. The criterion used to design the filter is to minimize the mean squared error between the filter output and a desired indicator function located at the target position in the presence of the noise. We will present computer simulation results for a number of blurred noisy input images, and the performance of the proposed filter will be determined. The second contribution of this dissertation is the extension of the MMSE (Wiener) filter using a set of training images to achieve distortion-invariant with respect to different target aspects. We will also test the composite filter's discrimination against undesired objects and tolerance to out-of-plane rotational distortions. The detection performance of this MMSE algorithms have been found to be superior to filters that are optimal with respect to noise statistics but do not take into account the effects of environmental conditions. The third contribution of this dissertation is the use of the Receiver Operating Characteristics curve to compare processors used for pattern recognition. Finally, we will derive the performance of the MMSE filters, the Matched Spatial Filter and the Phase-Only Filter for blurred and un-blurred scenes under the Gaussian noise assumption. This is the first unifying work to illustrate the dependence of filter performance as a function of the target parameters and the noise parameters. It is now possible to predict filter performance without doing any computer simulations. Computer simulations that are typically used to show superior performance are never conclusive since one can always select a scene in which the filter performs better for some particular performance criterion. ^

Share

COinS