Signal Processing - November 2017 - 24

FROM THE GUEST EDITORS
Fatih Porikli, Shiguang Shan, Cees Snoek,
Rahul Sukthankar, and Xiaogang Wang

Deep Learning for Visual Understanding

n the past decade, there has been a transformative and permanent revolution in
computer vision cultivated by the reinvigorated adoption of deep learning for
visual understanding tasks. Driven by the
increasing availability of large annotated
data sets, efficient training techniques, and
faster computational platforms, deeplearning-based solutions have been progressively employed in a broader
spectrum of applications from image classification to activity recognition.
Deep learning, in general, refers to
a range of artificial neural networks that
consist of multiple layers, mimicking the
structure and cognitive process of the
human brain. Instead of relying on handcrafted features, they allow the acquisition
of knowledge directly from data. They
regress intricate objective functions in a
nested hierarchy, where more sophisticated representations with larger receptive
fields computed in terms of less abstract
ones with localized supports. Deep learning also makes it possible to incorporate
explicit domain knowledge and replace
a large variety of conventional algorithmic blocks with trainable differentiable
modules. These all give deep learning an
exceptional power and flexibility in modeling the relationship between the input
data and target output.
Efforts are now shifting toward
the remaining challenges. For instance,
the majority of current methods have

Digital Object Identifier 10.1109/MSP.2017.2744538
Date of publication: 13 November 2017

been designed to solve supervised learning problems where data comes with its
labeled attributes and how to reliably
apply deep learning to unsupervised settings in a similar degree of success is an
active area of research. Similarly, recent
efforts aim at working with small data,
focusing on how to take advantage of large
quantities of unlabeled examples as well as
with a few labeled samples.
Another area where deep agents may
play a significant role is to integrate positive and negative rewards into deep learning to choose the actions that yield the
best cumulative reward by interacting
with the environment. Also, the fusion of
multimodal and structured data into existing deep-learning models would open up
more extended application domains.
This special issue of IEEE Signal
Processing Magazine (SPM) is therefore
devoted to providing survey articles on
the latest advances in deep learning for
visual understanding. Its objective is to
encourage a diverse audience of researchers and enthusiasts toward an effective
participation in the solution of analogous
problems in other signal processing fields
by inseminating similar ideas.
The range of articles in this two-part
special issue indicates the breadth of the
computer vision discipline. (Part two
will be published in January 2018.) Many
fundamental areas are surveyed from the
computer vision perspective, including
■■ reinforcement learning
■■ learning with limited and no supervision (unsupervised learning)
IEEE SIGNAL PROCESSING MAGAZINE

November 2017

weakly supervised learning
zero- and few-shot learning
■■ domain adaptation
■■ multimodal learning
■■ metric learning
■■ generative adversarial networks
■■ recurrent networks
■■ regression with Bayesian networks
■■ model compression and robustness.
In addition, in-depth overviews of several
deep-learning-based computer vision
applications are provided, including
■■ inverse problems such as superresolution and image enhancement
■■ picture quality prediction
■■ saliency detection
■■ image and video segmentation with
conditional random fields
■■ image-to-text generation
■■ visual question answering
■■ face image analytics.
We would like to wholeheartedly thank
all of the contributing authors and reviewers of this special issue. We also sincerely
appreciate SPM's editor-in-chief, Prof.
Min Wu, Managing Editor Jessica Welsh,
and the entire magazine's editorial staff
for their extremely valuable support.
■■
■■

Meet the guest editors
Fatih Porikli (fatih
.porikli@anu.edu.au)
received his B.Sc.
degree in electrical
e n g i n e e r i n g from
Bilkent University,
Turkey, in 1992 and his Ph.D. degree in
electrical and computer engineering from

http://www.B.Sc

Table of Contents for the Digital Edition of Signal Processing - November 2017