A Gaussian-mixture based approach to spatial image background modeling and compensation

In an optical inspection instrument, there is an undesirable image background which is often due to the nonuniform illumination characteristics of the system. The background however may also involve other, hard-to-model effects such as stray light. In the present paper, we report on our efforts to achieve robust elimination of smooth image backgrounds so as to achieve improved inspection of flat patterned media. We consider a uniform two-dimensional array of bivariate Gaussian functions on the image plane and consider the optimal approximating model to the smooth image background signal. The representation and associated algorithm effectively captures the background while being minimally effected by the high frequency pattern on the inspection surface. The set of linear weights of the Gaussian kernel offers a compact representation of the background and is used to eliminate the background for further processing (e.g., defect detection) of the surface image. Performance results are illustrated on a representative problem of TFT-LCD panel inspection for finding production defects. This process involves a sub-pixel resolution pattern subtraction scheme and therefore is sensitive to background variations, effectively forming a good case study.