XClose

UCL Centre for Medical Image Computing

Home
Menu

Tammy Riklin Raviv - CMIC/WEISS joint seminar series

29 January 2020, 1:00 pm–2:00 pm

Tammy Riklin Raviv - Ben Gurion University - a talk as part of CMIC/WEISS joint seminar series

Event Information

Open to

All

Organiser

cmic-seminars-request@cs.ucl.ac.uk

Location

90 HH function room
90 High Holborn
90 High Holborn
LONDON
WC1V 6LJ

Tammy Riklin Raviv - Ben Gurion University

 

Title - Differentiable Histogram-based Loss Functions for Image-to-Image translation with Deep Learning

 

Abstract

In the first part of my talk, I will give a brief overview of some of our recent deep-learning methods for Biomedical Image computing. Specifically, I will focus on image reconstruction, 3D modeling, image segmentation and quantitative estimation of instance segmentation quality.

In the second part, I will present a novel technique for a differentiable construction of cyclic and joint (2D) intensity histograms and show how these can be utilized for defining differentiable loss functions as part of a deep learning framework for image-to-image translation. While the proposed method can be applied to a variety of problems in which the output is an image, we choose to demonstrate its strength on color transfer problems, where the aim is to paint a source image with the colors of a different target image. Note that for this type of problems, the desired output-image does not exist and, therefore standard pixel-to-pixel loss functions, such as L_2 or cross-entropy cannot be used. Instead, we define an intensity-based loss function that is built on the Earth Movers’ Distance (EMD) between the histograms of the output and the target images. In addition, having a differentiable approximation of the joint histogram between the source and the output images, we define a statistical pixel-to-pixel similarity loss by the Mutual Information of these images. The incorporation of histogram loss functions in addition to an adversarial loss enables the construction of  meaningful and realistic images.