Two Routes for Image-to-Image Translation: Rule-based vs. Learning-based
Dr. Minglun Gong of the Memorial University of Newfoundland
Friday, December 15, 2017 · 10:30AM · HEC 450
Abstract
Many image processing and computer vision tasks, such as image segmentation, stylization, and abstraction, can be posed as image-to-image translation problems. This talk presents two different image-to-image translation approaches, one is rule-base and the other is learning-based.
The rule-based algorithm is capable of stylizing an input face photo using a single exemplar image. Since the numbers and varieties of patch samples are highly limited, special cares are put into sample selection to best preserve the identity and content of the input face. A two-phase procedure is also designed, where colors are transferred first in a semantic-aware manner, followed by edge-preserving texture transfer.
The learning-based algorithm employs Conditional Generative Adversarial Networks (GANs) to perform general cross-domain image-to-image translation. It requires a large set of training images, but unlike existing approaches, the images do not need to be labeled. To train in an unsupervised manner, two GANs are constructed to translate images in opposite directions, forming a closed loop. As a result, images from either domain can be translated to the other and then reconstructed, enabling a reconstruction error term for training.
Biography
Dr. Minglun Gong is a Professor and Head at the Department of Computer Science, Memorial University of Newfoundland. He obtained his Ph.D. from the University of Alberta in 2003 and his M.Sc. from the Tsinghua University in 1997. Minglun’s research interests cover various topics in the broad area of visual computing (including computer graphics, computer vision, visualization, image processing, and pattern recognition). So far, he has published over 100 referred technical papers in journals and conference proceedings, including 18 articles in ACM/IEEE transactions. Currently an associate editor for Pattern Recognition, he has also served as program committee member for several top-tier conferences.
