Evolving network architectures with custom computers for multi-spectral feature identification

This paper investigates the design of evolvable FPGA circuits where the design space is severely constrained to an interconnected network of meaningful high-level operators. The specific design domain is image processing, especially pattern recognition in remotely sensed images. Preliminary experiments are reported that compare neural networks with a recently introduced variant known as morphological networks. A novel network node is then presented that is particularly suited to the problem of pattern recognition in multi-spectral data sets. More specifically, the node can exploit both spectral and spatial information, and implements both feature extraction and classification components of a typical image processing pipeline. Once trained, the network can be applied to large image data sets, for at the sensor to extract features of interest with two orders of magnitude speed-up compared to software implementations.