DATE:

Friday, November 8, 2019

TIME:

3:30 PM

LOCATION:

GMCS-314

SPEAKER:

Dr. Jerome Gilles, Associate Professor, Department of Mathematics and Statistics, SDSU.

ABSTRACT:

Texture analysis in image processing remains a challenging problem mainly due
to the fact that textures can vary a lot which makes difficult to find a mathematical
model for textures. However, the community usually agree that some characteristics
like orientation, periodicity and scale are very important to characterize textures.
These observations have naturally led to the use of wavelet to extract features
suitable for textures. Such features can then be used to feed either an unsupervised
or supervised classifier to perform texture segmentation/classification. In this work,
we show that empirical wavelets are particularly relevant to characterize textures
since they are well adapted to capture “optimized” features. We then use those feature
with standard classifier like k-means or the Nystrom algorithm to perform unsupervised
texture segmentation. Finally, we introduce the use of these features to train neural
networks to classify textures. We show that such easy change largely boost existing
neural networks and permits reaching performances close to the perfection.

HOST:

Jose Castillo

DOWNLOAD:

PDF File