Application of a multilayer network in image object classification
Yonghong Tang, William G. Wee, Chia Yung Han · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 1991
The major objective of this system is to classify image objects into two classes which represent good and defective indications of the objects. Several features are extracted from the indications based on geometrical, morphological, and gray scale intensity information of the image objects. Most of the extracted features of the two classes are clustered and nonseparable when projected onto a 1-D feature space using the Fisher''s linear discriminant method. To get better results, a multilayer neural network system that increases the class separation distance is used. The network, which is based on the structure proposed by Webb, consists of an input layer, an output layer, and one hidden layer. The inputs to the input layer are feature vectors, the elements of which are extracted features of the image objects. The transformation of the input feature vector to the hidden units is nonlinear, while the transformation of the output of the hidden layer to the output layer is linear. The nonlinear transformation from the input layer to the hidden units is such that the resulting pattern for discrimination is easier than the original pattern. The transformation from the hidden to the output layer is linear so that the square error at the output of the network system can be minimized. By choosing a suitable nonlinear transformation and number of hidden units, this network structure is applicable to the 2-class discrimination problem to achieve better class separation. The neural network system for classifying the two classes of objects is described. This method of classification is very important in many industrial inspection applications. Comparison of the numerical results of the neural network approach with the classical pattern recognition approach, the Fisher''s linear discriminant, is made and presented as well.© (1991) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.