中国科学院自动化研究所机构知识库

Recently, flow-based approaches have shown considerable success in interpolating video images. However, in contrast to video images, electron microscope (EM) images are further complex due to noise and severe deformation between consecutive sections. Consequently, conventional flow-based interpola-tion algorithms, which assume a single offset per position, are not able to robustly model the movement of such complicated data. To address the aforementioned problems, this study propose a novel EM image interpolation framework that accommodates a range of offsets per location and further distills the inter-mediate features. First, a spatio-temporal ensemble (STE) interpolation module for capturing the missing middle features is presented. The STE is subdivided into two modules: temporal interpolation and resid-ual spatial-correlated block (RSCB). The former predicts the intermediate features in two directions with several offsets at each location. Moreover, the RSCB uses the correlation coefficients for aggregated sam-pling. Thus, even if intermediate features are severely deformed, the STE effectively improves their accu-racy. Second, a stackable feedback distillation block (SFDB) is introduced, which enhances the quality of intermediate features by distilling them from the input, and interpolated images, using a feedback mech-anism. Extensive experiments demonstrate that the proposed method presents a superior performance compared with previous studies, both quantitatively and qualitatively.(c) 2022 Elsevier B.V. All rights reserved.