A method of feature extraction from an image for quality analysis

We propose a method for feature extraction from a two-dimensional pattern with an application to a semiconductor's quality management problem. Given a distribution pattern of defected chips on a wafer, our method generates a linguistic description such as "There is a medium-size dense lump of defected chips at the upper right corner of the wafer." In addition to the use as a report for a quality manager, this description is used as an index to retrieve wafer data having similar patterns because similar patterns tend to suggest that the failures are caused by similar causes. Our method generates a linguistic description by the following two steps; First, we generate a dendrogram of defected chips using a hierarchical clustering method. Then, using the dendrogram, we identify an optimum granularity of cluster distribution, and each cluster is associated with a linguistic description. Through experiments, the description generated by our method is found to agree well with expert engineers judgement.