Using Spatial Transform Networks and Transforming Auto-encoders to Build Advanced Deep Learning Methods for Tomographic Imaging Models

Deep Learning methods have revolutionised machine learning in the last few years. Recently Spatial Transform networks and Transforming Auto-encoders have been developed that specifically encode spatial transformation parameters into the network architecture. In this project, these ideas will be applied to the recovery of volumetric images from noisy and limited projection data. In fast CT imaging applications, the problem is the efficient reconstruction of volumetric x-ray attenuation images from noisy and limited data. Deep learning approaches have the advantage that they can learn prior information about objects that are to be imaged, including shape and material distributions. In this project, Spatial Transform Networks and Transforming Auto-encoders will be used to build these models.