Multi-stage generation for semantic image segmentation

Computer vision is one of the most active fields of research in artificial intelligence. Recently, state–of–the–art results have been achieved in a variety of tasks related to computer vision adopting deep learning models. However, most of these approaches rely on supervised end–to–end training of complex neural networks that contain a huge number of parameters. Therefore, training these models requires large sets of supervised examples to obtain good generalization capabilities, which prevents their use in scenarios where there is scarcity of data. The main goal of this thesis is to provide new methods to generate synthetic images that can be used to enlarge small sets of real samples. A domain–shift is generally present between generated and real data, since synthetic images rarely reproduce the exact appearance of real ones. In this thesis different approaches are proposed to improve the realism of generated data, thus implicitly reducing the domain–shift. Specifically, machine learning based methods are used to better simulate the spatial position of the objects in the scene and to increase the realism of the simulation. The proposed approaches have been successfully employed in a specific domain, the automatic analysis of agar plate images, in which we have used synthetic data to train a deep semantic segmentation network. Our experiments show that the generated images can be a viable alternative to real data when a large set of supervised samples is not available. Moreover, a completely data–driven approach is proposed to generate both synthetic images and the corresponding supervision. Based on few real examples, the presented framework is able to learn to reproduce images of high–resolution and quality, that can be used to integrate existing datasets of real images. The method has been evaluated for the segmentation of blood vessels in retinal images and allowed to obtain state–of–the–art results.