Challenges and Opportunities in Computational Imaging and Sensing – Prof. Pier Luigi Dragotti, Imperial College London

May 14, 2020 @ 3:00 pm – 4:00 pm
JDB Seminar Room
Dr Ramji Venkataramanan

In many areas of science and engineering new signal acquisition methods allow unprecedented access to physical measurements and are challenging the way in which we do signal and image processing. Within this broad theme related to the interplay between sensing and processing, the main focus of this talk is on new sampling methodologies inspired by the advent of event-based video cameras and on solving selected inverse imaging problems in particular when multi-modal images are acquired.

In the first part of the talk, we investigate biologically-inspired time-encoding sensing systems as an alternative method to classical sampling, and address the problem of reconstructing classes of sparse signals from time-based samples. Inspired by a new generation of event-based audio-visual sensing architectures, we consider a sampling mechanism based on first filtering the input, before obtaining the timing information using leaky integrate-and-fire architectures. We show that, in this context, sampling by timing is equivalent to non-uniform sampling, where the reconstruction of the input depends on the characteristics of the filter and on the density of the non-uniform samples. Leveraging specific properties of the proposed filters, we derive sufficient conditions and propose novel algorithms for perfect reconstruction from time-based samples of classes of sparse signals. We then highlight further avenues for research in the emerging area of event-based sensing and processing.

We then move on to discuss the single-image super-resolution problem which refers to the problem of obtaining a high-resolution (HR) version
of a single low-resolution (LR) image. We consider the multi-modal case where a scene is observed using different imaging modalities and when these modalities have different resolutions. In this context, we use dictionary learning and sparse representation framework as a tool to model dependency across modalities in order to dictate the architecture of deep neural networks and to initialize the parameters of these networks. Numerical results show that this approach leads to state-of-the-art results in multi-modal image super-resolution applications. If time permits will also present applications in the area of art investigation.

Leave a Reply

Your email address will not be published. Required fields are marked *