A model-based neural network for edge characterization

Abstract

In this paper, we investigate the feasibility of characterizing significant image features using model-based neural network with modular architecture. Instead of employing traditional mathematical models for characterization, we ask human users to select what they regard as significant features on an image, and then incorporate these selected features directly as training examples for the network. As a first step, we consider the problem of the characterization of edges, which are usually regarded as significant image features by humans. Unlike conventional edge detection schemes where decision thresholds have to be specified, the current NN-based edge characterization scheme implicitly represents these decision parameters in the form of network weights which are updated during the training process, and which thus allow automatic generation of the final binary edge map without further parameter adjustments. Experiments have confirmed that the resulting network is capable of generalizing this previously acquired knowledge to identify important edges in images not included in the training set. In particular, one of the important attributes characterizing the current approach is its robustness against noise contaminations: the network can be directly applied to noisy images without any re-training and alteration of architecture, as opposed to conventional edge detection algorithms where re-adjustment of the threshold parameters are usually required.