SEM-DCNet: A Self-Supervised Network for SEM Images Zonary Non-Linear Distortion Correction

Large-scale scanning electron microscope (SEM) has been used to reconstruct the gigantic and complex neural network in the brain, where the natural intelligence is embedded. Due to the limited field of view and the stability of the scanning coils, the ultra-large tiles imaged by single-beam SEM will inevitably exist some zonary non-linear distortions, which could have an unexpected impact on the down-stream stitching, alignment and segmentation. In this paper, we first analyze the parametric expression of the zonary distortion on SEM and then propose a novel self-supervised convolutional neural network named SEM-DCNet to automatically detect and correct the randomly distributed zonary distortion on the SEM images. Experiment on the real-world dataset shows that the proposed method can efficiently correct the distortion to a considerable degree, hence greatly improve the stitching performance under rigid transformation model.