Evolving the mapping between input neurons and multi-source imagery

We present a mutable input field concept that allows a neural network to evolve a mapping between its input layer and a 3-dimensional 'input cube' consisting of a local window applied within multiple imagery sources, such as hyperspectral bands, feature maps, or even encoded tactical information regarding likely object location and class. This allows the net to exploit salient regions (both within and across sources) of what may otherwise be an unwieldy input domain. Small recurrent neural networks are evolved to perform object detection within airborne reconnaissance imagery that has been processed to provide 3 colour bands and 2 feature maps including one designed to identify man-made structures based on perpendicularity of edge direction. A variable input field is shown to provide faster convergence and superior detector fitness over a number of trials than a set of alternative fixed input field mappings.