Learning Spatially Adaptive Blur Representations for Stable Deconvolution in Electron Microscopy

Spatially heterogeneous blur is frequently observed in practical electron microscopy (EM) acquisitions and can lead to nonuniform resolution across the field of view. Herein, we propose a patch-wise anisotropic blur-kernel estimation framework that performs uncertainty-aware regularized shift-variant deconvolution. The proposed method is not intended to replace standard scanning electron microscopy (SEM) acquisition optimization, such as focus correction, aperture selection, working-distance control, beam-current adjustment, or astigmatism correction. Rather, it is designed as a complementary post-acquisition restoration and quality-control tool for cases in which spatially heterogeneous blur remains after acquisition or when reacquisition is impractical. The method is expected to be most useful in high-magnification imaging, large-area SEM mapping, serial-section or volume EM workflows, and archived datasets, where local degradation may vary across the field of view. In controlled simulations with known references, the method improved restoration consistency relative to the evaluated global-point spread function (PSF) conditions. Application to experimental SEM images demonstrated exploratory feasibility, but the absence of matched blur-free references precluded conclusions regarding true spatial-resolution recovery or faithful morphology.