An effective image segmentation technique for the SEM image

Abstract

There have been lots of efforts to replace the human eye inspection of the SEM image by automatic inspection based on the reference comparison method. The two kinds of inspection methods exist: the direct comparison method and the indirect comparison method. The more widely used indirect comparison method requires segmentation step which is to split the original image into two regions, foreground region and background region. Especially the segmentation of SEM image is not easy due to high noise level, variation of the image offset, and the diversity of patterns. In previous work, the ridge detector had been used to overcome such characteristics of the SEM image. In this paper, we present an effective segmentation method developed on the watershed segmentation algorithm, global-local threshold method, Laplacian of Gaussian filter, and non-maximum suppression. Applied for segmentation of various SEM images, the presented method showed the accuracy of 94% for ID image type and 98% for 2D image type.