A GAUSSIAN-MIXTURE BASED APPROACH TO SPATIAL IMAGE BACKGROUNDMODELING AND COMPENSATION

摘要

In an optical inspection instrument, there is an undesirablernimage background which is often due to the nonuniformrnillumination characteristics of the system. The backgroundrnhowever may also involve other, hard-to-model effects suchrnas stray light. In the present paper, we report on our efforts tornachieve robust elimination of smooth image backgrounds sornas to achieve improved inspection of flat patterned media. Wernconsider a uniform two-dimensional array of bivariate Gaussianrnfunctions on the image plane and consider the optimalrnapproximating model to the smooth image background signal.rnThe representation and associated algorithm effectivelyrncaptures the background while being minimally effected byrnthe high frequency pattern on the inspection surface. Thernset of linear weights of the Gaussian kernel offers a compactrnrepresentation of the background and is used to eliminate thernbackground for further processing (e.g., defect detection) ofrnthe surface image. Performance results are illustrated onrna representative problem of TFT-LCD panel inspection forrnfinding production defects. This process involves a sub-pixelrnresolution pattern subtraction scheme and therefore is sensitivernto background variations, effectively forming a goodrncase study.