Feature Grouping Based on Graphs and Neural Networks

Abstract

An attempt is made to develop a parallel-sequential method for feature grouping with special applications to gray value segmentation, grouping of contour points, and dot pattern clustering. These problems are closely connected with other preprocessing tasks such as edge preserving smoothing and edge detection which are also dealt with. The method is based on a special graph structure called Feature Similarity Graph which is defined via an adaptive feature similarity criterion and (Voronoi) neighborhood relations between the features. A recursive edge preserving method of feature averaging which is based on the similarity criterion is presented. Via the averaging procedure nonlocal processing is implemented which is necessary for efficient noise reduction. Network structures which are similar to the Cellular Neural Networks (CNN) can be used efficiently for implementing the nonlinear algorithm. One processing element or neuron is assigned to each feature and (Voronoi) neighbored features are connected with adaptive weights depending on feature similarity. It is demonstrated with few examples that the same grouping principles can be used for different tasks of segmentation and clustering.