Advanced Precipitate and Dislocation Segmentation in STEM Images using U-net Architecture and Focused Region Training

Abstract

Recently, the U-net architecture, based on fully convolutional networks (FCNs), has gained significant attention for its exceptional performance in image segmentation. Particularly in the field of image analysis, this architecture has proven to be highly effective in generating accurate results. In this study, we employ the U-net architecture to segment precipitates and the matrix in Scanning Transmission Electron Microscopy (STEM) images and detect dislocations around the interface. To achieve this, we propose a novel system called the "Segregation and Merging system." In the segregation step, we convert the input image data into two distinct types of data. The first type is used for segmenting dislocations, while the background includes both precipitates and the matrix. To capture and learn localized features, such as dislocations, more effectively, we introduce a concept called "Focused Region Training." The second type of data is used to segment precipitates and the matrix, employing separate models for each type, resulting in unique outputs. In the merging step, we combine the outputs from both models to generate a multi-label segmentation map. This approach effectively addresses common challenges related to low accuracy and overfitting, commonly associated with limited computational power and image resolution. Notably, our method significantly improves upon existing techniques. Furthermore, we explore potential applications of our findings, such as measuring distances between dislocations. Leveraging the multi-label segmentation map, we extract valuable information that aids in the further analysis and characterization of materials. Overall, our study presents a novel method for image segmentation using the U-net architecture.