多次元空間におけるモデルベースド欠陥検出

抄録

A highly sensitive inspection algorithm is proposed that extracts defects in multidimensional vector spaces from multiple images. The proposed algorithm projects subtraction vectors calculated from test and reference images to control the noise by reducing the dimensionality of vector spaces. The linear projection vectors are optimized using a physical defect model, and the noise distribution is calculated from the images. Because the noise distribution varies with the intensity or texture of the pixels, the target image is divided into small regions and the noise distribution of the subtraction images are calculated for each divided region. The bidirectional local perturbation pattern matching (BD-LPPM) which is an enhanced version of the LPPM, is proposed to increase the sensitivity when calculating the subtraction vectors, especially when the reference image contains more high-frequency components than the test image. The proposed algorithm is evaluated using defect samples for three different scanning electron microscopy images. The results reveal that the proposed algorithm increases the signal-to-noise ratio by a factor of 1.32 relative to that obtained using the Mahalanobis distance algorithm.