Neural-fuzzy Feature Detector : A New Approach

A novel scheme for developing, at low computational cost, neural-fuzzy classifiers based on large-scale, model-based exemplars is outlined. The new method extends the approach that Bezdek applied to train a neural net (NN) Sobel edge classifier by training the NN on the complete population of 3x3 binary image prototypes scored to fuzzy values by a classical operator. We first show that, replacing the fuzzy values of edgeness of the exemplaers, by crisp defuzzified values vastly improved computational speed. A complexity analysis proves however that for operators based on larger windows, the use of complete binary exemplars sets will be computationally intractable. In the new scheme the NN classifier is trained over a hybrid set { selected binary image exemplars with crisp outputs | sampled pixels within a realistic image, these pixels being crisply scored by use of a classic operator.} We demonstrate the scheme by deriving a 5x5 neural fuzzy Plessy operator, far superior to the classic Plessy.