Multilayer backpropagation network for flexible circuit recognition

This paper presents an industrial application of the multilayer backpropagation neural network with a modified learning rule, in recognizing transparent flexible membrane printed circuits independent of the position, orientation and scale. We give a new learning algorithm which reduces the complexity of the multilayer backpropagation network by pruning insignificant weights and chooses the best size to suit the underlying complexity of the recognition problem. This new learning algorithm is also compared with other network redundancy reduction techniques and tested on 3-bit parity problem. In this particular application, moment invariant features were chosen to train the multilayer backpropagation network. As the circuits have regular shape with a limited number of corners, a fast corner based moment invariance estimation algorithm is employed. This algorithm is almost hundred times faster than standard occupancy array based algorithm for shapes with a small number of corners