Denoising of Transmission Electron Microscopy Images for Atomic Defect Identification

Two-dimensional (2D) materials possess remarkable electrical, mechanical, thermal, and optical properties, making them applicable in the fields of electronics and optoelectronics. Atomic defects could significantly affect device performance and reliability. Clearly identify and analysis of atomic defects in 2D materials is crucial. Transmission electron microscopy (TEM) with atomic resolution serves as an ideal method for the defect characterization of 2D materials. However, TEM images often suffer from various types of noise, which severely hampers the analysis of 2D materials. In this study, we propose a neural network based on the optimized CycleGAN architecture for TEM image denoising even when obtaining strictly paired datasets for TEM is challenging. The optimized CycleGAN model with attention gate could improve the signal-to-noise ratio and structural similarity of the denoised images. The effectiveness of this method was validated on the scanning transmission electron microscopy image dataset. Enhancing the visibility of atomic contrast and improving the accuracy of identifying individual atomic structures are achieved by removing the surrounding noise from the original image.