Artificial neural networks with gamma kernels for automatic target detection

The quadratic gamma detector (QGD) has been used successfully to discriminate man made objects from background clutter in synthetic aperture (SAR) radar imagery. It implements a linear discriminant function based on quadratic terms of the image intensity of a pixel under observation and of its surroundings. This paper extends the QGD to a multilayer perceptron with a first layer built from 2D gamma kernels. We call this nonlinear extension the NL-QGD. We compare the performance of the QGD and NL-QGD based on receiver operating curve of targets in SAR data. The NL-QGD shows improved performance with respect to the QGD for most of the values of the detection probability. However, there is a deterioration of performance at 100% target detection, and the system performance is highly dependent upon the stopping criterion