Enhancement of noisy EDX HRSTEM spectrum-images by combination of filtering and PCA

Abstract

STEM spectrum-imaging with collecting EDX signal is considered in view of the extraction of maximum information from very noisy data. It is emphasized that spectrum-images with weak EDX signal often suffer from information loss in the course of PCA treatment. The loss occurs when the level of random noise exceeds a certain threshold. Weighted PCA, though potentially helpful in isolation of meaningful variations from noise, might provoke the complete loss of information in the situation of weak EDX signal. Filtering datasets prior PCA can improve the situation and recover the lost information. In particular, Gaussian kernel filters are found to be efficient. A new filter useful in the case of sparse atomic-resolution EDX spectrum-images is suggested.

Keywords: Filtering; PCA; Spectrum-imaging.