With technological advancements in recent years, single particle cryogenic electron microscopy (cryo-EM) has become a major methodology for structural biology. Structure determination by single particle cryo-EM is premised on randomly orientated particles embedded in thin layer of vitreous ice to resolve high-resolution structural information in all directions. Otherwise, preferentially distributed particle orientations will lead to anisotropic resolution of the structure. Here we established a deconvolution approach, named AR-Decon, to computationally improve the quality of three-dimensional maps with anisotropic resolutions reconstructed from datasets with preferred orientations. We have tested and validated the procedure with both synthetic and experimental datasets and compared its performance with alternative machine-learning based methods.