Neural network based classification using automatically selected feature sets

For many vision-based inspection tasks, clear measurable features inherent in a visual image are sufficient to allow classification of the image content. Sometimes, however, the classification can only be made on the basis of subtle relationships within the image, making heuristic selection of suitable feature sets difficult. An example of such an inspection task is the classification of integrated circuit solder joints on surface mount printed circuit boards. A procedure for automatically selecting feature sets which best represent the distinction between different classes of solder joints, based on a modified version of the Karhunen-Loeve transform applied to images, is summarized. The results of applying artificial neural network classification techniques to coefficients produced by this procedure in the above inspection task are presented. The use of `momentum´ effects and network pruning procedures, and the interpretation of the function of internal network nodes are also discussed