Abstract:

Scanning transmission electron microscopy (STEM) can realize the characterization and analysis of the microstructure at nano and atomic scales. However, the imaging process will be affected by unpredictable non-uniform noise, which will affect the atomic structure analysis. In the past few years, predictive models based on deep learning networks can reduce the impact of noise through de-noising or data fitting, but these models still meet an overfitting problem. In this paper, a high-noise STEM atomic structure segmentation method based on structural condition priors of materials is designed by introducing structural condition priors of materials into deep neural network models. The structure condition prior of materials is modeled into segmented network attention using the contrast learning method, which includes the self-attention mechanism and cross-attention mechanism. By calculating these two kinds of attention, the segmentation network can not only focus on the key regions in the image but also focus on the control information from the structure coordinate vector mode. The experimental results show that the proposed segmentation network can segment the atomic structure of high-noise STEM images more accurately than other traditional segmentation networks, and accurately predict the secondary structure information that is blocked by noise or the top layer.

Key words: atomic structure segmentation, high noise STEM, prior information model