Neural networks for step edge detection
F.-Y. Liao, Mitchell Middler, Weilin Lin · 2002
Summary form only given. Previously proposed neural networks were modified to enhance their capability of edge detection. The previous network consists of two separate modules: the prospective edge selection module and the final edge selection module. The prospective edge selection module consists of two or more networks, working in parallel, that locate prospective edges in the image. A separate network is required to detect the edges in each desired orientation. The prospective edge map resulting from the preliminary networks is then fed into the final network to remove spurious edges and select the desired edges. The proposed modification allows the networks to ignore small changes of pixel values in a region and thus avoid overdetection. In addition, only the edge pixels detected by the preliminary network have the chance to retain their status (as edge pixels) in the final network. The inclusion of this constraint not only avoids the occurrence of inconsistency between the preliminary and final networks but also speeds up the computation significantly.>