Infrared Image Derivation Method
for Generative Adversarial Network with Object
Constraint
Ying Zhu (B) , Min Li, Le Ma, Xiao Xiang Qi, and Yu Song
Xi’an Institute of High-Tech, Xi’an, China
18270826122@163.com
Abstract. The generative adversarial network provides a method to derive
infrared images from visible images, but the contour and details of the object
of the generated infrared images are inaccurate. to solve this problem, we investigate how to improve the framework of the generative adversarial network, four
different constraints are designed in this paper. Namely, the sequential cascade
generation model of the subject in the segmentation map(GA), the sequential cascade generation model of the visible subject(GB), the superimposed generation
model(GC), and the previous convolution feature cascade generation model(GD).
In GB, the first three channels is the main body, and their inputs are the target
segmentation map of visible images, which directly cascade Object segmentation
graph in the last three channels. The experimental results show that the infrared
image generated by GB has more intact target outlines, clearer target details, and
is better on the whole.
Keywords: Infrared image · GAN · Deep learning · Image generation
1 Introduction
The infrared image can solve the shortcomings of the visible image. Due to the high cost
of infrared equipment, difficulty in maintenance and inconvenient use, it is difficult and
costly to directly capture high-resolution infrared images. Investigating how to generate
infrared images from visible images has an urgent application background.
In this paper, four infrared image derivation methods (TarGAN) [1] based on target
constraints are designed. The experiment was carried out on the Thermal-World [2]
dataset.
2 Network Design
2.1 Framework Design
In order to obtain the target segmentation graph for constraint, this paper uses the CRFasRNN [3] network as the preprocessing network. The whole network framework is based
© 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. 239–244, 2021.
https://doi.org/10.1007/978-3-030-64690-5_23
for Generative Adversarial Network with Object
Constraint
Ying Zhu (B) , Min Li, Le Ma, Xiao Xiang Qi, and Yu Song
Xi’an Institute of High-Tech, Xi’an, China
18270826122@163.com
Abstract. The generative adversarial network provides a method to derive
infrared images from visible images, but the contour and details of the object
of the generated infrared images are inaccurate. to solve this problem, we investigate how to improve the framework of the generative adversarial network, four
different constraints are designed in this paper. Namely, the sequential cascade
generation model of the subject in the segmentation map(GA), the sequential cascade generation model of the visible subject(GB), the superimposed generation
model(GC), and the previous convolution feature cascade generation model(GD).
In GB, the first three channels is the main body, and their inputs are the target
segmentation map of visible images, which directly cascade Object segmentation
graph in the last three channels. The experimental results show that the infrared
image generated by GB has more intact target outlines, clearer target details, and
is better on the whole.
Keywords: Infrared image · GAN · Deep learning · Image generation
1 Introduction
The infrared image can solve the shortcomings of the visible image. Due to the high cost
of infrared equipment, difficulty in maintenance and inconvenient use, it is difficult and
costly to directly capture high-resolution infrared images. Investigating how to generate
infrared images from visible images has an urgent application background.
In this paper, four infrared image derivation methods (TarGAN) [1] based on target
constraints are designed. The experiment was carried out on the Thermal-World [2]
dataset.
2 Network Design
2.1 Framework Design
In order to obtain the target segmentation graph for constraint, this paper uses the CRFasRNN [3] network as the preprocessing network. The whole network framework is based
© 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. 239–244, 2021.
https://doi.org/10.1007/978-3-030-64690-5_23