Defocus blur synthesis and deblurring through interpolation in the latent space

High quality images play an important role in medical diagnosis and analysis. To ensure that microscopic imaging results are qualitative, most microscope systems nowadays are equipped with autofocusing hardware or software-based components. Nonetheless, there are cases when the optimal focal distance is not correctly identified and images present out-of-focus areas or are completely affected by defocus blur. In this paper, we investigate a generative model with twofold applicability. It can be used for recovering from defocus blur as well as for synthesizing defocus blur for data augmentation purposes. Both these tasks are achieved through interpolation in the latent space of an autoencoder. We apply two forms of linear enforcement to the latent space of an autoencoder trained to synthesize defocus blur in microscopy images. We evaluate the models and find that the regularized autoencoders outperform the baseline model in terms of synthesizing blur and deblurring images.