Computer vision for advanced electron microscopy analysis

Abstract

Transmission electron microscopy (TEM) is one of the most powerful techniques used to characterize materials. However, characterization of a set of heterogenous sample images still remains as a persistent challenge. This can be a painstaking process in which researches may spend a tremendous amount of time to classify particle’s crystal structures manually. Moreover, this may also lead to sampling bias by reporting the favourable image in the report that is not representative of the whole sample. Therefore, this study aims to provide solution to this problem by utilizing machine learning. The project focuses especially in building an optimized model to recognize synthetic images with convolutional neural network (CNN). Synthetic images were generated with MATLAB due to relevant dataset is not present at the moment. Afterwards, a CNN model was built and optimized by adjusting several parameters including choice of optimizer, loss function and regularization in Python. The model presented was able to attain 98.44% accuracy and 0.034 loss. This preliminary model is hopefully able to accelerate future works in recognizing synthetic images and real TEM images by transfer learning.