Abstract We present a novel method for the analysis of quantum gas microscope images, which uses deep learning to improve the fidelity with which lattice sites can be classified as occupied or unoccupied. Our method is especially suited to addressing the case of imaging without continuous cooling, in which the accuracy of existing threshold-based reconstruction methods is limited by atom motion and low photon counts. We devise two neural network architectures which are both able to improve upon the fidelity of threshold-based methods, following training on large data sets of simulated images. We evaluate these methods on simulations of a free-space erbium quantum gas microscope, and a noncooled ytterbium microscope in which atoms are pinned in a deep lattice during imaging. In some conditions we see reductions of up to a factor of two in the reconstruction error rate, representing a significant step forward in our efforts to implement high fidelity noncooled site-resolved imaging.