Multi-period Infrared Image Generation Based on MCGAN
235
to enhance the overall quality of the generated infrared images. In this paper, we define
L1 norm as follows:
L 1 (GA) = E
y i − GA(x, z i ) 1
(1)
L 1 (GB) = E
x − GB(y i , kz) 1
(2)
where GA and GB denote generator A and B respectively. The loss functions of
discriminator A and B are expressed as (3) (4):
L DA = −E
DA(y i )
+ E[DA(GA(x, z i ))] + L Di
(3)
L DB = −E[DB(x)] + E
DB(GB(y i , kz))
+ L Di
(4)
where L Di denotes gradient penalty [11] to the discriminator. We introduce L1 norm as
additional penalty term to construct the loss function of generator:
L GA = −E x,z [DA(GA(x, z i ))] + E
GA(GB(y i , kz), z i ) − y i 1
+ αL 1 (GA) + L Gi
(5)
L GB = −E x,z
DB(GB(y i , kz))
+ E
||GB(GA(x, z i ), kz) − x|| 1
+ βL 1 (GB) + L Gi
(6)
where α and β denote coefficients of L1 norm to generator A and B, respectively.
4 Experiments
In this section, we evaluate the performance of MCGAN. Since CGAN has a good
performance in image generation, we select it as a rival to MCGAN. Moreover, another
comparison network is a NW-MCGAN, which does not use W distance compared with
MCGAN.
4.1 The Datasets
There are two datasets, self-built dataset and LTIR [12] dataset, being used in the experiments. With using infrared imaging device ZENMUSE XT2, we captured 303 infrared
images, which were taken from 48 different visual angles of certain buildings in 5 different periods. These images are divided into training set containing 240 images and test
set containing the rest of them.
4.2 MCGAN Training Details
The generator GA and GB are the most important parts of MCGAN. They have the same
architecture and hyperparameters, as shown in the Fig. 2.