A Two-layer Neural Network for Robust Image Segmentation and Its Application in Revising Hydrographic Features

In this paper we use a two-layer neural network for robust image segmentation. The rst layer is a locally coupled recurrent network. Using probabilistic measures, contextual information is incorporated eeectively through underlying coupling structures, which reduces and even resolves local ambiguities in images. The second layer is a LEGION (Locally Excitatory Globally Inhibitory Oscillator Network) network. It has has been shown rigorously that the LEGION network can achieve synchronization among oscillators corresponding to one region and desynchronization among diierent oscillator groups representing diierent regions rapidly. Thus it provides a computationally eeective representational framework for image segmentation. We observe fast convergence and robustness to noise of the two-layer network, which is demonstrated using synthetic and real images. For practical use, we develop an eecient algorithm, which has been successfully applied to extracting hydrographic features from digital orthophoto images. Based on the segmentation results, we propose a framework for revising hydrographic and other geographic features by integration through adjacency graphs.