Title: Primal-Dual Network in Image segmentation

Abstract: Many approaches in traditional image analysis such as image segmentation, denoising, motion blur etc can be framed as an energy functional using total variational models (TVM). Framing the TVM as a deep neural network would allow end to end joint training of energy functionals with convolutional neural networks (CNNs) providing better solutions. The recent emergence for proximal operator show a lot of promise in this direction.

The seminar is dedicated to presenting an introduction to the Proximal operator and its possible use in solving the generic TVM and energy minimization techniques. Some background on the previous attempts in depth and multi-class labeling and their

limitations will be discussed. This will be followed by a presenting an architecture that combines the primal-dual unrolled scheme with CNN architecture. The current approach overcomes few of the issues discussed and has been successfully applied in historical document binarization.