Region labelling with a multi-layered perceptron

The correct labelling of segmented images has been a difficult and generally unsolved problem in computer vision for many years. Unlike conventional labelling methods, it is possible to design a neural network implementation which only uses contextual and statistical features that are directly measured from an image, and which does not require any initial labelling at all. The article describes the preliminary results of a study into the possibility of using such a neural network to label the regions obtained from real images of natural outdoor scenes, by first training it on such scenes