Unsupervised training of structuring elements
Stephen S. Wilson · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 1991
A very robust form of pattern recognition results from using mathematical morphology on binary images that have been segmented into various edge directions. These edge direction images are then transformed by a set of structuring elements to derive a set of feature images. Morphological transformations on the feature images by a final set of structuring elements result in the location or classification of the desired objects. Thus, there are two types of structuring element sets. The training of the final set of structuring elements can be provided by a method similar to supervised Hebbian learning used in neural networks, where the goal is to unambiguously locate specified objects. The first set of structuring elements is similar to a hidden layer in neural networks and is more difficult to train. A technique of unsupervised competitive learning is used. The definition of orthogonality, or uniqueness within the set of features defined by the first layer of erosions is crucial to successful training. These features must be independent and span the space of possible features; otherwise, information that may be critical to the final layer of erosions will be lost. This paper will concetrate on a technique for unsupervised learning of hidden layers of structuring elements where a training image is inspected with very little input as to what will be done with the image.