Remote Sensing Image Object Recognition Based on
Convolutional Neural Network
Yumei Zhen, Huanyu Liu, Junbao Li
Department of Automatic Test and
Control, Harbin Institute of Technology,
Harbin 150080, China
zhenyumei1995@163.com
Cong Hu
School of Electronic Engineering and
Automation, Guilin University of
Electronic Technology,
Guilin 541004, China Email:
hucong@guet.edu.cn
Jeng-Shyang Pan
*
Fujian Provincial Key Lab of Big Data
Mining and Applications,
Fujian University of Technology,
Fuzhou 350108, China
jspan@cc.kuas.edu.tw
*
corresponding author
Abstract—The development of space remote sensing
technology brings a lot of remote sensing image data. The
traditional target detection method is difficult to adapt to the
large amount of high-resolution remote sensing image data. It is
necessary to find a way to automatically learn the most effective
features from the image data, and to fully recover the correlation
between the data. Based on the recognition of the typical targets
in remote sensing image data, this paper proposes a method of
remote sensing target recognition based on deep learning. In this
paper, the pre-train method is introduced to improve the
simulation of the model. The experiment of the test set proves the
validity of the method.
Keywords—deep learning;object recognition; pre-train
I. INTRODUCTION
With the development of space remote sensing technology
and high resolution satellite technology, the acquisition of
spatial remote sensing data becomes more convenient. The data
from remote sensing images has been greatly increased. It
provides the possibility to develop a more intelligent target
interpretation system for remote sensing images
[6]
. Therefore, it
is urgent to study the algorithm of detection recognition based
on large data. Extracting the target's valid characteristics from a
huge amount of high score data is a key technology.
Traditional target detection methods are difficult to adapt to
large amounts of data. Its dependence is characterized by
artificial design. This is time-consuming and relies heavily on
the expertise and the characteristics of the data itself. It is
difficult to produce an effective classifier from a large amount
of data to fully exploit the correlation between the data. So we
need to find a way to learn the features automatically. By
studying a large amount of data itself, the most effective
feature representation is obtained. By establishing a relatively
complex network structure, the correlation between the data is
fully mined to establish a strong feature extraction and
classifier. In recent years, more active depth learning has
provided an effective possibility for automatic extraction of
target characteristics. It can also get better results in remote
sensing image processing.
II. OBJECTION RECOGNITION PROCESS
A. Introduction of netual networks
Artificial neural network is a kind of imitate biological
neural network (animal central nervous system, especially the
brain structure and function of the mathematical model or
computational model. Its simple structure is shown in Fig 1.,
including the input layer, the implicit layer, and the output
layer, where the implied layer may have multiple layers. Each
neuron in the neural network is connected to all the nodes in
the previous layer, which is called a full connection, where
each connection is connected to a certain weight.
[4]
.The
process of network training is the process of gaining weight.
Fig 1 Simple structure of the artificial neural network
Structurally, the Convolutional Neural Network (CNN)is a
multilayer Neural Network. It consists mainly of the
convolution layer, subsampling layer, and full connection
layer
[5]
. Each layer consists of multiple two-dimensional
planes, each composed of multiple independent neurons. In
1998, LeNet5
[1]
marked the launch of CNN.
B. Design of the model
This article selects the winner of the ILSVRC14 contest
[2]
,
the GoogLeNet model, as the base model. Generally speaking,
the most straightforward way to improve network performance
is to increase the depth and width of the network, which means
a huge amount of parameters. However, the large quantity
parameters are easily produced and the calculation is greatly
increased. The ability to maintain the sparse network structure
and the high computational performance of the dense matrix is
an important problem. For this reason, GoogLeNet comes up