ABSTRACT
This chapter adopts user-defined salient image features as training examples for a specially designed model-based neural network to perform feature detection. This chapter presents a model-based feature detection neural network with hierarchical architecture (HMBNN) which directly learns the essential characteristics of user-specified features through a training process. Edge characterization represents an important subproblem of feature extraction, the aim of which is to identify those image pixels with appreciable changes in intensities from their neighbors. In simple edge detection, the user specifies a global threshold on the edge magnitudes in an interactive way to produce a binary edge map from the edge magnitude map. The result is usually not satisfactory since some of the noisy pixels may be misclassified as edge pixels. The localization of an edge within a single pixel width requires some form of Laplacian of Gaussian (LoG) filtering to detect the zero crossings, where more associated parameters have to be specified.
