Fig. 2: Example reconstruction of an experimental image.
From: An unsupervised deep learning algorithm for single-site reconstruction in quantum gas microscopes
a Raw fluorescence image showing a Mott-insulating state (~2500 atoms) in a box potential (arrows mark the lattice orientation). Crops containing 16 × 16 sites are extracted from the full image and fed into the encoder part. b The deconvolved site counts are extracted from the bottleneck layer before binarization and reassembled according to their position in the full image. c A histogram of these values reveals a bimodal distribution with a clear separation between empty and occupied lattices sites, allowing to set a threshold for the site occupation. d Applying this threshold finally gives the reconstructed lattice occupation corresponding to the input image. Light gray circles denote empty and dark purple occupied sites.
