Dealing with damage: Region-based image segmentation despite artifacts and bias
Event details
| Date | 18.06.2015 |
| Hour | 10:30 |
| Speaker |
Dr Dominique Zosso, University of California, Los Angeles Bio: Dominique Zosso received the M.Sc. degree in electrical and electronics engineering and the Ph.D. degree from EPFL in 2006 and 2011, respectively. He was Research and Teaching Assistant at the Signal Processing Laboratory LTS5, EPFL, from 2007 to 2012. He then became SNSF postdoctoral fellow, and is currently a CAM Assistant Adjunct Professor with the Department of Mathematics, University of California, Los Angeles, CA, USA, working with Luminita A. Vese, Andrea L. Bertozzi and Stanley J. Osher. His research interests include variational and PDE methods, and efficient algorithms to solve inverse problems in imaging and computer vision, and related problems in machine learning and data science. He develops mathematical models and cutting edge algorithms, to help understand images and solve problems in medicine, biology, robotics and consumer electronics. |
| Location |
DIA003
|
| Category | Conferences - Seminars |
Region-based image segmentation has essentially been solved by the Chan-Vese (CV) model. However, this model fails when images are affected by artifacts (outliers) and illumination bias that outweigh the actual image contrast. Here, we introduce a model for segmenting such images. In a single energy functional, we introduce 1) a dynamic artifact class preventing intensity outliers from skewing the segmentation, and 2), in Retinex-fashion, we decompose the image into a piecewise-constant structural part and a smooth bias part. The CV-segmentation terms then only act on the structure, and only in regions not identified as artifacts. The segmentation is parameterized using a phase-field, and efficiently minimized using threshold dynamics.
We demonstrate the proposed model on a series of sample images from diverse modalities exhibiting artifacts and/or bias. Our algorithm typically converges within 10-50 iterations and takes fractions of a second on standard equipment to produce meaningful results. We expect our method to be useful where image damage prevents classical CV-segmentation from working, and anticipate use in applications where artifacts and bias are actual features of interest, such as lesion detection and bias field correction in medical imaging, e.g. in magnetic resonance imaging (MRI).
We demonstrate the proposed model on a series of sample images from diverse modalities exhibiting artifacts and/or bias. Our algorithm typically converges within 10-50 iterations and takes fractions of a second on standard equipment to produce meaningful results. We expect our method to be useful where image damage prevents classical CV-segmentation from working, and anticipate use in applications where artifacts and bias are actual features of interest, such as lesion detection and bias field correction in medical imaging, e.g. in magnetic resonance imaging (MRI).
Practical information
- General public
- Free
Organizer
- Prof. Jean-Philippe Thiran
Contact
- Prof. Jean-Philippe Thiran