Using an image retrieval system for image data management

Images of semiconductor defects are maintained in semiconductor yield-management systems to help diagnose problems that arise during the manufacturing process. A common problem in future systems is the number of images to maintain, which is increasing at an alarming rate due to the growing use of in-line and off-line imaging systems. A manufacturing-specific content-based image retrieval system, or Automated Image Retrieval (AIR) system, was developed by ORNL in coordination with International SEMATECH during 1998-1999. The system uses commercial databases to store image information and uses a customized indexing technology to rapidly retrieve similar images base don visual content. In addition to acting as a yield management tool based on storing and retrieving images, the system can be utilized as a tool for data management by helping determine when images are redundant in relation to previously stored data. Ideally this information can be used to time-stamp the data for future purging based on a variety of ratings such as 'long- term', 'mid-term', and 'short-term'. In some situations the feedback from the AIR system can even be used to omit the image entirely based on pre-existing close matches. In this paper we explore techniques for using the AIR system to assist in image data management. Experimental results are shown with simulated image data representing various degrees of image clustering of redundancy, and manufacturing image data accumulated during earlier field-testing of the AIR system at industry sites. Early results indicate substantial reductions in the size of industry databases may be achievable while continuing to maintain an adequate representation and history of the manufacturing process. To reduce the number of stored images, AIR technology can be used in place of, or as a guide for, the typical 'store for N months and purge' approach to image management. This approach will enhance the use of the image database, since the real bottleneck in such a procedure is the need to sort such massive amounts of stored data as opposed to actual disk space.