A boosted decision tree approach to shadow detection in scanning electron microscope (SEM) images for machine vision applications.

Scanning electron microscopy is important across a wide range of machine vision applications, and the ability to detect shadows in images could provide an important tool for evaluating attributes of the surfaces being imaged, such as the presence of defects or particulate impurities. One example where the presence of shadows can be important is in the reconstruction of elevation maps from stereo-pair scanning electron microscopy (SEM) images. Shadows can both interfere with determination of matching points for stereoscopic calculations, and confuse shape-from-shading algorithms which rely on pixel intensity gradients to calculate surface slope, leading to inaccurate reconstructions. This paper describes a machine learning method for identifying locations in SEM images impacted by shadows, based on a training set of photographic images. The method could be useful as a means of identifying parts of images likely to suffer from reconstruction artifacts in shape-from-shading-based reconstructions, or as a tool for automated defect identification. The method uses a boosted decision tree machine learning approach to identify shadows based on the features of images. The method is illustrated with four different natural surfaces exhibiting a range of different types of shadow features, and an example is used to illustrate how the method can identify regions likely to be impacted by shadows in reconstructions.

A hybrid 3D SEM reconstruction method optimized for complex geologic material surfaces.

Reconstruction methods are widely used to extract three-dimensional information from scanning electron microscope (SEM) images. This paper presents a new hybrid reconstruction method that combines stereoscopic reconstruction with shape-from-shading calculations to generate highly-detailed elevation maps from SEM image pairs. The method makes use of an imaged glass sphere to determine the quantitative relationship between observed intensity and angles between the beam and surface normal, and the detector and surface normal. Two specific equations are derived to make use of image intensity information in creating the final elevation map. The equations are used together, one making use of intensities in the two images, the other making use of intensities within a single image. The method is specifically designed for SEM images captured with a single secondary electron detector, and is optimized to capture maximum detail from complex natural surfaces. The method is illustrated with a complex structured abrasive material, and a rough natural sand grain. Results show that the method is capable of capturing details such as angular surface features, varying surface roughness, and surface striations.