xi
Preface
is book comes during a deep learning revolution in computer vision, when performance of,
e.g., object classification on ImageNet [Russakovsky et al., 2014] has improved vastly from top-
5 error of 26% in 2011 to 16% in 2012. e major paradigm shift has been to move from
engineered image features (“pixel f***ing” according to Koenderink and van Doorn [2002]) to
learned deep features. So why write this text now? Many recent publications making use of deep
learning show a lack of rigor and their way to throw data at the problem is unsatisfying from
a theoretical perspective. Attempts to put deep networks into established frameworks as done
by Mallat [2016] are essential contributions to the field. Deep learning is very important from
a practical perspective, but having a well-founded understanding of the underlying features and
how they relate to common approaches can only help whether you are using deep learning or
engineered features. Indeed, there might be a twist to use channel representations inside of deep
networks, but there are further motivations to write this text. One reason is to honor the work
by Gösta Granlund, the father of channel representations, who recently finished his academic
career. A second motivation is to summarize one of my own branches of research, as I have
been working on feature representations since my master’s thesis 20 years ago (Felsberg [1998]).
Last but not least, this text addresses many mathematical and algorithmic concepts that are
useful to know and thus I want to share with students, colleagues, and practitioners. None of
those groups of people is addressed exclusively and presumably none will see this as a primary
source of information, but I hope that all will find new aspects and try to formulate new research
questions as a consequence.
Michael Felsberg
April 2018
..................Content has been hidden....................
You can't read the all page of ebook, please click here login for view all page.