Hardware implementation of RAM-based neural networks for tomographic data processing

The hardware discussed in the paper employs field programmable logic devices to interface with memory components and, unlike previous implementations, utilises dynamic RAM without compromising performance. The network offers the opportunity to estimate process parameters without recourse to image reconstruction. Tests reveal speedups of 17, 22 and 6 for image reconstruction, void fraction estimation and flow regime classification, respectively, from electrical capacitance tomography data.