IEEE Signal Processing Magazine - January 2018 - 86

2) Since many methods have been proposed and breakthrough performances have been achieved in recent years,
it would be enlightening to review the recently proposed
object detection techniques, especially for those based on
deep-learning techniques.
3) It is of great interest to conduct in-depth discussions for a few
important questions. For example, why can the recent deeplearning-based frameworks significantly boost the performance of object detection? What is the most intrinsic
improvement of such frameworks when compared with previous ones? What are the problems that need to be solved in
the future for deep-learning-based approaches?
4) Comprehensive comparison and analysis for the experimental results on publicly available object detection benchmarks
would help readers to better understand the performance of
each object detection strategy as well as the corresponding
network architecture.

Preliminary knowledge
In recent years, there has been rapid development in the research
area of deep learning, including its popularization in computer
vision. In this section, we briefly introduce one of the advanced
deep-learning techniques that has been widely used in the object
detection task, i.e., the CNN.
The CNN is one of the most well-known and widely used
deep-learning architectures inspired by the natural visual perception mechanism of living creatures, which was first proposed in 1980 by Fukushima [21] and then improved by LeCun
[22]. CNNs are designed to process data that come in the form
of multiple arrays [23], for example, a color image composed
of three two-dimensional arrays that contain pixel intensities
in the three color channels. There are four key ideas behind
CNNs that take advantage of the properties of natural signals:

local connections, shared weights, pooling, and the use of
many layers [23].
As shown in Figure 2, the architecture of a typical CNN
model is structured as a series of layers as follows:
■ Convolutional layers: Convolutional layers are the most
important for feature extraction. The first several layers usually capture low-level features (like edges, lines, and corners)
while the deeper layers are able to learn high-level features
(like structures, objects, and shapes) by combining low-level
ones. Each unit in a convolutional layer is connected to a
local patch in the feature maps of the previous layer through
a set of kernels called a filter bank. The result of this local
weighted sum is then passed through a nonlinearity operation
such as a rectified linear unit (ReLU). All units in a feature
map share the same filter bank. Different feature maps in a
convolutional layer use different filter banks.
■ Pooling layers: Pooling layers aim to reduce the dimension
of the representation and create invariance to small shifts and
distortions. A pooling layer is usually placed between two
convolutional layers. Each feature map of a pooling layer is
connected to its corresponding feature map of the previous
convolutional layer. A typical pooling unit computes the
maximum of a local patch of units in one feature map.
■ Fully connected layers: Fully connected layers are typically
used as the last few layers of the network for better summarizing the information conveyed by lower-level layers in
view of the final decision. As a fully connected layer occupies
most of the parameters, overfitting can easily happen. To
prevent this, the dropout method is usually employed [24].
Starting with the breakthrough success of AlexNet [24] for
ImageNet classification in 2012, significant efforts have been
made in developing various CNN models, including VGGNet
[25], GoogLeNet [26], and ResNet [27].

Feature Maps

Input Image

Feature Maps
Convolutional Kernal

Convolutional Layer

Pooling Layer

Fully Connected Layer

FIGURE 2. The architecture of a typical CNN model.
86

IEEE SIGNAL PROCESSING MAGAZINE

January 2018

Table of Contents for the Digital Edition of IEEE Signal Processing Magazine - January 2018