Neural Networks for Classifying Images of Wood Veneer. Part 2

A decision tree using smaller more specialised modular neural networks for the classification of wood veneer by an automatic visual inspection system was presented in Part 1 [1]. A key process in the design of a modular neural network is the use of "normalised inter-class variation" in the selection of the most appropriate image features to be used for its particular specialised classification task. At the root of the decision tree is a single large (holistic) neural network that initally attempts to classify all of the image classes which include clear wood and 12 possible defects (13 classes). The initial design uses 17 features of the acquired image of the wood veneer as inputs. The selection (or more correctly pruning) of inputs for this large neural network used not only "normalised inter-class variation", but also "normalised intra-class variation" in the features and their "correlation" within the same class. This results in the elimination of 6 inputs. The revised smaller 11 input neural network results in a substantial reduction in classification time, for the computer implementation used here, and at the same time the classification accuracy is improved. This is the root of the decision tree described in the previous paper.