Spatially Regularized Multiscale Graph Clustering for Electron Microscopy

We propose an unsupervised, multiscale learning method for the segmentation of electron microscopy (EM) imagery. Large EM images are first coarsely clustered using spectral graph analysis, thereby non-locally and non-linearly denoising the data. The resulting coarse-scale clusters are then considered as vertices of a new graph, which is analyzed to derive a clustering of the original image. The two-stage approach is multiscale and enjoys robustness to noise and outlier pixels. A quasilinear and parallelizable implementation is presented, allowing the proposed method to scale to images with billions of pixels. Strong empirical performance is observed compared to conventional unsupervised techniques.