Reconstructing site-resolved lattice occupation with high fidelity is crucial in quantum gas microscopy experiments for accurate physical observable extraction. Existing methods struggle with short interatomic separations and limited signal-to-noise ratios, especially when lattice spacing is below half the imaging resolution. This paper introduces a deep convolutional neural network (CNN)-based algorithm for high-fidelity reconstruction of site-resolved lattice occupation. The unsupervised algorithm trains directly on experimental fluorescence images, enabling fast reconstruction of large images (thousands of sites). Benchmarked using a cesium atom quantum gas microscope with short-spaced optical lattices (383.5 nm lattice constant, 850 nm Rayleigh resolution), the algorithm achieves promising reconstruction fidelities (≥96%) across various fillings. The algorithm's potential extends to experiments with shorter lattice spacing, enhancing readout fidelity and speed for lower-resolution imaging systems, and finding applications in trapped ion experiments.

Alexander Impertro, Julian F. Wienand, Sophie Häfele, Hendrik von Raven, Scott Hubele, Till Klostermann, Cesar R. Cabrera, Immanuel Bloch, Monika Aidelsburger