Neural Filtering for Fusing TV and Thermal Images

The image fusion problem is formulated as a least-squares estimation problem, using the tv and thermal images as the measurements and the corresponding "perfect" image as the signal to the estimated. To overcome the difficulties asociated with the conventional estimation-theoretic approach, a neural filter is proposed. The neural filter, which was recently developed at the Maryland Technology Corporation, is a recurrent multilayer perceptron systhesized from experimental data without the need for a mathematical model for the data. It was analytically proven to be optimal for the given architecture and to converge to the least-squares estimator as the filter size increases. General purpose neural network chips (eg Intel's 80170NX, 45mm by 45mm, 1.5W in active mode) are commercially available for prototyping real-time (3 microseconds/layer) neural filters for image fusion. therefore, the production model, which will be one or more special purpose chip(s), is expected to be small and energy-frugal, making it ideal for military and commercial applications.