Recently, solutions based on machine learning for image restoration have been proposed in the literature. In general, image distortions are modeled by linear and spatial invariant point spread functions (PSFs). However, these distortions are nonlinear in practice, which justifies the use of nonlinear solutions to mitigate their effect. In this work, we intend to use neural networks to improve the quality of images. In particular, we will consider: (I) multilayer perceptron (MLP), (II) convolutional neural network (CNN), and (III) generative adversarial network (GAN). The goal is to study the feasibility of training these networks in order to maximize the mean structural similarity, a measure that takes into account characteristics of the human visual system and has been widely used in the literature to compare the similarity between two images. Finally, to obtain good results independently of distortion, we will also combine such networks in a mixture of experts.
News published in Agência FAPESP Newsletter about the scholarship: