Label-free deeply subwavelength optical microscopy

We report the experimental demonstration of deeply subwavelength far-field optical microscopy of unlabeled samples. We beat the similar to lambda/2 diffraction limit of conventional optical microscopy several times over by recording the intensity pattern of coherent light scattered from the object into the far-field. We retrieve information about the object with a deep learning neural network trained on scattering events from a large set of known objects. The microscopy retrieves dimensions of the imaged object probabilistically. Widths of the subwavelength components of the dimer are measured with a precision of lambda/10 with the probability higher than 95% and with a precision of lambda/20 with the probability better than 77%. We argue that the reported microscopy can be extended to objects of random shape and shall be particularly efficient on object of known shapes, such as found in routine tasks of machine vision, smart manufacturing, and particle counting for life sciences applications.