This image-processing algorithm produces robust segmented images from low-resolution scanning electron microscope images of integrated circuits without a reference ground-truth image. Arriving at images of integrated circuits that are segmented, meaning that each pixel is labeled as part of the integrated circuit or not, is a vital step when checking that an integrated circuit was built as intended (hardware assurance) or determining the logical function of an integrated circuit from its physical structure (reverse engineering). Scanning electron microscopes are powerful tools for imaging integrated circuits, but achieving images of today’s intricate integrated circuits high-quality enough for segmentation requires the microscope to dwell for a long time over each cell of the integrated circuit, quickly adding up to weeks of imaging times. Therefore, a procedure for segmenting low-quality integrated circuit images produced with short dwell times is necessary to avoid the time costs of integrated circuit reverse engineering and hardware assurance spiraling out of control.
Researchers at the University of Florida have developed an image-processing algorithm that is more effective in segmenting low dwell-time images, resulting in impressive time savings during scanning electron microscope imaging without sacrificing segmentation reliability. This parallelizable algorithm also leverages machine learning so that it can segment features of various length scales without fine-tuning, making it widely applicable.
Produce reliable segmented images of integrated circuits using only easily available, low-quality scanning electron microscope data as input
Scanning electron microscopes work by firing a beam of electrons at a small region of the sample. These electrons subsequently scatter off and are collected to form an image of the region. Dwelling over the same region in a firing/collecting state for a longer time correlates with higher-quality images and shorter dwell times can significantly increase image noise as well as reduce image quality. For reverse-engineering purposes, each pixel of the final scanning electron microscope image must be classified according to whether it is part of the integrated circuit or not. The many different length scales of integrated circuit structures render this task challenging. However, machine learning algorithms are well suited to learn how to identify whatever features may be present. This algorithm deploys Gaussian mixture models in combination with different low-pass filters to train itself to identify the various frequencies present in the image. Rather than being programmed to pick out certain structures, this algorithm benefits from an unsupervised workflow that allows it to choose the best filter to remove noise but recover sharp edges. It achieves all this with significant time savings due to accepting low dwell time scanning electron microscope images and being parallelizable.
