- Ponente: Dr. Mehran Ebrahimi, profesor de la Ontario Tech University, Canada
- Fecha: 24 de febrero
- Hora: 12:00 horas
- Lugar: sala de reuniones del CITIC (Centro de Investigación Tecnologías de Información y las Comunicaciones, Calle Periodista Rafael Gómez Montero, 2, 18014 Granada
In many practical problems in the field of applied sciences, the features of most interest cannot be observed directly, but have to be inferred from other, observable quantities. The problem of solving an unknown object from the observed quantities is called an inverse problem. Many classical problems in imaging can be modelled as inverse problems. Many real-world inverse problems are ill-posed, mainly due to the lack of existence of a unique solution. The procedure of providing an acceptable unique solution to such problems is known as regularization. Indeed, much of the progress in image processing in the past few decades has been due to advances in the formulation and practice of regularization. This, coupled with the progress in the areas of optimization and numerical analysis, has yielded much improvement in computational methods of solving inverse imaging problems.
In this talk, we will revisit a number of inverse problems including image registration (alignment), image inpainting (completion), super-resolution (resolution enhancement), and present some recent research ideas mainly aimed at medical imaging applications. Furthermore, we present an approach based on deep convolutional neural networks to address two image restoration problems, image inpainting and super-resolution. The method applies our so-called “Edge-Connect”, a two-stage adversarial model that contains an edge generator followed by an image completion network. We evaluate the model and observe that it outperforms current state-of-the-art techniques quantitatively and qualitatively.