ORIGINAL ARTICLE
Deep transfer learning for military object recognition under small
training set condition
Zhi Yang
1,2
•
Wei Yu
3
•
Pengwei Liang
3
•
Hanqi Guo
3
•
Likun Xia
4
•
Feng Zhang
5
•
Yong Ma
3
•
Jiayi Ma
3
Received: 29 November 2017 / Accepted: 27 March 2018
Ó The Natural Computing Applications Forum 2018
Abstract
Convolutional neural network is powerful for general object recognition. However, its excellent performance depends largely on
huge training set. Facing task like military object recognition in which image samples for training are scarce, its performance will
degrade sharply. To solve this problem, a deep transfer learning method is proposed in this paper. The main idea consists of two
parts: transfer learning for prior knowledge embedding and mixed layer for better feature extraction. It has been proved that the
ability of feature extraction learned in large dataset is helpful to related tasks and can be transferred to a new neural network. The
transfer learning process is achieved by fixing the weights of some layersandthenretrainingtheremained layers. The key problem
for deep transfer learning is which part should be transferred and which part should be retrained to adapt the network to the new task.
This problem is solved by extensive experiments, and it is found that retraining the last three layers and transferring prior to the other
layers can reach the best performance. Besides, we used mixed layer scheme to make use of the current information. In each mixed
layer, convolution filters in different scales are combined together, helping to adapt features in different scales. By employing these
two method s, the proposed method exhibits a large improvement in military object recognition under small training set. Experi-
ments demonstrate that our method can achieve a high recognition precision, superior to many other algorithms compared.
Keywords Object recognition Small training set Military Transfer learning Convolutional neural network
1 Introduction
Since convolutional neural network (CNN) was used for
general object recognition [14], deep learning methods
exhibit extraordinary performance in many vision
problems, such as detection [7, 8, 24], semantic segmen-
tation [19, 20], navigation [23], retrieval [18], human
posture estimation [30], artistic style transferring [12] and
video advertising [35]. The key to the success of CNN
methods is the powerful representation learned with hier-
archical architecture and large-scale training set that usu-
ally contains millions of images. However, in task like
military object recognition, it is hard to get enough training
data due to the expense or the scarcity of the target sam-
ples. As a result, CNN methods will suffer serious degra-
dation in these cases.
To fulfill the task of military object recognition under
small training set conditi on with neural network, two
subproblems should be solved. On the one hand, it is
necessary to take advantage of some prior knowledge in
order to compensate for the deficiency of training data. On
the other hand, network structure that enables more pow-
erful feature extract ion ability should be engineered. The
main idea of the scheme solves the problem from two
aspects: improving data diversity with prior from large
dataset and strengthening the representation to take the
most advantages of information from current data.
Zhi Yang and Wei Yu contributed equally to the work.
& Jiayi Ma
jiayima@whu.edu.cn
1
College of Computer Science and Technology, Wuhan
University of Science and Technology, Wuhan 430065,
China
2
Hubei Province Key Laboratory of Intelligent Information
Processing and Real-time Industrial System, Wuhan
University of Science and Technology, Wuhan 430065,
China
3
Electronic Information School, Wuhan University,
Wuhan 430072, China
4
College of Information Engineering, Capital Normal
University, Beijing 100048, China
5
China Academy of Electronics and Information Technology,
Beijing 100041, China
123
Neural Computing and Applications
https://doi.org/10.1007/s00521-018-3468-3
(0123456789().,-volV)(0123456789().,-volV)