Hardware implementation of pulse mode RBFNN based edge detection system on virtex V platform

Abstract

In this paper, we have proposed a new architecture of RBFNN. Neural network efficiency in embedded systems offers the possibility of reconfiguration and the genericity of the solution. Indeed, the same integrated system can approximate any input-output function thanks to the parameters update on the chip. RBF neural networks constitute a subset of the neuronal networks, which has a great potential in reducing the size of the network. Pulse mode neural networks reduce significantly hardware resources by replacing the conventional huge multiplier by a simple frequency multiplier. As application, we approximate with the proposed RBF network, a Canny operator based edge detection, which is an important step in image processing. Acceptable edge detection approximation was done, with a mean generalization error of (4,604 %) on the Wang image database. Moreover, a design synthesis on FPGA virtex V platform was done, the results of implementation lead to an operating frequency of 445,295 MHz, which offers real time application performances.