Precisely identify the geometry of catalyst particles from S/TEM images via a boundary attention deep learning network

Abstract

The size and shape of the nanoparticles have a great impact on their performances as catalysts. Traditional method of manually measuring particles from scanning/transmission electron microscopy (S/TEM) images has limited counting ability of only up to hundreds of particles, not enough to support an accurate statistic and besides, the one dimensional (1D) manual measurement by drawing a line across nanoparticles would induce large bias given the practical supported catalysts whose particles always present irregular geometry. To develop an automatic and precise recognition method based on TEM data has been a long-standing demand to facilitate a highly efficient catalytic dispersion evaluation. In this paper, we propose an end-to-end boundary attention deep learning network to recognize nanoparticles in S/TEM images. The proposed model is an extension of the U-Net embedding a residual boundary attention module to extract the edge information of catalyst particles and a fusion module to exploit the edge features, gradient features and the results out from decoder to enhance recognition performance. Results demonstrate that the proposed boundary attention structure is superior to well-accepted U-Net framework in dealing with the precise particle recognition. Using our new method, a particle size distribution of PdCu catalyst at statistic scale of 14,000 + is achieved, through which more reliable and accurate results of S/TEM characterization can be obtained to facilitate efficient and reliable evaluation of catalysts dispersion.