Abstract:The major objective of this system is to classify imageobjects into two classes which represent good anddefective indications of the objects. Several featuresare extracted from the indications based ongeometrical, morphological, and gray scale intensityinformation of the image objects. Most of the extractedfeatures of the two classes are clustered andnonseparable when projected onto a 1-D feature spaceusing the Fisher's linear discriminant method. To getbetter results, a multilayer neural network system thatincreases the class separation distance is used. Thenetwork, which is based on the structure proposed byWebb, consists of an input layer, an output layer, andone hidden layer. The inputs to the input layer arefeature vectors, the elements of which are extractedfeatures of the image objects. The transformation ofthe input feature vector to the hidden units isnonlinear, while the transformation of the output ofthe hidden layer to the output layer is linear. Thenonlinear transformation from the input layer to thehidden units is such that the resulting pattern fordiscrimination is easier than the original pattern. Thetransformation from the hidden to the output layer islinear so that the square error at the output of thenetwork system can be minimized. By choosing a suitablenonlinear transformation and number of hidden units,this network structure is applicable to the 2-classdiscrimination problem to achieve better classseparation. The neural network system for classifyingthe two classes of objects is described. This method ofclassification is very important in many industrialinspection applications. Comparison of the numericalresults of the neural network approach with theclassical pattern recognition approach, the Fisher'slinear discriminant, is made and presented as well.!
展开▼
