Multi-period Infrared Image Generation Based
on Multi-conditional Cycle Generative
Adversarial Networks
Xiaoxiang Qi 1(B) , Min Li 1(B) , Le Ma 1 , Ying Zhu 1 , and Yu Song 1,2
1 Xi’an Research Institute of Hi-Tech, Xi’an 710025, China
qixiaoxiang08@163.com, clwn@163.com
2 College of Information and Communication, National University of Defense Technology,
Xi’an 710106, China
Abstract. Since the number of most infrared image samples is too small, many
deep learning networks don’t work well on these infrared datasets. In this paper,
we propose a Multi-conditional Cycle Generative Adversarial Network, which can
derive multi-time infrared images from some fixed-time infrared images, and can
be used to augment the small infrared image dataset. By taking several different
time periods as multi-condition constraints, the model we proposed, which is based
on generative adversarial network (GAN), can derive infrared images in different
time periods. In order to improve the quality of derived infrared images, this paper
introduces L1 norm and Wasserstein distance to construct the objective function,
which make it easier to stabilize the training of the network, and the model can
learn more useful features. The small infrared images dataset in our experiment
is built by our team own. The experiments show that the method proposed in this
paper can derive multi-period infrared images. Compared with the CGAN, our
model can learn the feature mapping between the original images and the derived
images more accurately, and the generated infrared images are richer in details
and more realistic by vision.
Keywords: Infrared images · GAN · Deep learning · Image generation
1 Introduction
In recent years, deep learning is outstanding in target recognition and image segmentation, but the high cost of infrared equipment and few infrared image datasets become the
bottleneck of the application of deep learning in infrared target detection and segmentation. Therefore, it is necessary to study the augmentation method to infrared image
datasets.
The aim of this paper is to derive multi-period infrared images from one fixed-time
infrared image, so as to expand the infrared image dataset. Recently, researchers have
make it a reality to generate images and transfer its style [1]. However, the existing
network models, such as CGAN and CycleGAN, could not derive multi-period infrared
images well (see Subsect. 4.3). Therefore we propose a novel network, multi-conditional
cycle generative adversarial networks (MCGAN), to solve this problem.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
S. N. Atluri and I. Vušanovi´ c (Eds.): ICCES 2020, MMS 97, pp. 233–238, 2021.
https://doi.org/10.1007/978-3-030-64690-5_22
on Multi-conditional Cycle Generative
Adversarial Networks
Xiaoxiang Qi 1(B) , Min Li 1(B) , Le Ma 1 , Ying Zhu 1 , and Yu Song 1,2
1 Xi’an Research Institute of Hi-Tech, Xi’an 710025, China
qixiaoxiang08@163.com, clwn@163.com
2 College of Information and Communication, National University of Defense Technology,
Xi’an 710106, China
Abstract. Since the number of most infrared image samples is too small, many
deep learning networks don’t work well on these infrared datasets. In this paper,
we propose a Multi-conditional Cycle Generative Adversarial Network, which can
derive multi-time infrared images from some fixed-time infrared images, and can
be used to augment the small infrared image dataset. By taking several different
time periods as multi-condition constraints, the model we proposed, which is based
on generative adversarial network (GAN), can derive infrared images in different
time periods. In order to improve the quality of derived infrared images, this paper
introduces L1 norm and Wasserstein distance to construct the objective function,
which make it easier to stabilize the training of the network, and the model can
learn more useful features. The small infrared images dataset in our experiment
is built by our team own. The experiments show that the method proposed in this
paper can derive multi-period infrared images. Compared with the CGAN, our
model can learn the feature mapping between the original images and the derived
images more accurately, and the generated infrared images are richer in details
and more realistic by vision.
Keywords: Infrared images · GAN · Deep learning · Image generation
1 Introduction
In recent years, deep learning is outstanding in target recognition and image segmentation, but the high cost of infrared equipment and few infrared image datasets become the
bottleneck of the application of deep learning in infrared target detection and segmentation. Therefore, it is necessary to study the augmentation method to infrared image
datasets.
The aim of this paper is to derive multi-period infrared images from one fixed-time
infrared image, so as to expand the infrared image dataset. Recently, researchers have
make it a reality to generate images and transfer its style [1]. However, the existing
network models, such as CGAN and CycleGAN, could not derive multi-period infrared
images well (see Subsect. 4.3). Therefore we propose a novel network, multi-conditional
cycle generative adversarial networks (MCGAN), to solve this problem.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021
S. N. Atluri and I. Vušanovi´ c (Eds.): ICCES 2020, MMS 97, pp. 233–238, 2021.
https://doi.org/10.1007/978-3-030-64690-5_22