IEEE Consumer Electronics Magazine - March/April 2023 - 20

Special Section on Security, Trust and Privacy Solutions for Intelligent Internet of Vehicular Things
prove its universality in infrared target detection,
experiments are separately carried out on
FLIR infrared datasets (https://projects.asl.ethz.
ch/datasets), real infrared images captured statically
and real infrared images captured by a UAV
equipped with infrared camera.
EXPERIMENTS OF SR RECONSTRUCTION
1) Comparison Before and After Reconstruction
Experiment
1: FLIR thermal infrared image
sequence 1 is used in this experiment, with the
original resolution of 324 256 and the SR factor
of 4. The visual effects before and after SR
reconstruction are shown in Figure 7, the obvious
contrast area on both the LR input and the
SR result are scaled respectively to better compare
the visual saliency of the target before and
after reconstruction.
It is seen in Figure 7 that the texture structure of
FIGURE 7. Result of FLIR thermal infrared image
sequence 1.
where Ft1ðx; yÞ denotes the background area with
pixels of low gray level. Since the ultimate goal is
to enhance the texture structure of weak infrared
target, the original gray level distribution is maintained
in background part. Ft2ðx; yÞ can be seen as
the real texture part with pixels of relatively high
gray level, it is shown as the form of discrete texture
distribution on spatial domain. The histogram
of this part does not have statistical significance,
thus, Ft2ðx; yÞ is enhanced with linear gain. Then,
the final enhanced texture layer is expressed as
Fte x; yðÞ¼ Ft1 x; yðÞþ b Ft2 x; yðÞ:
(12)
It is found suitable to set the value of b as
1.05-1.1, through a large number of experiments.
More experimental results and analysis are also
shown in Section IV.
EXPERIMENTS AND ANALYSIS
In this section, extensive comparison experiments
are provided to verify the performance of
the proposed SR reconstruction method. To
20
the input LRimage is fuzzy with a lowcontrast. However,
in the SR result, the texture structure of the
image becomes clear, and the high-frequency information
lost in the original input images is restored
through the reconstruction process. To be specific,
the texture of the roof tiles in the original LR image
is difficult to identify, while it is clearly visible in the
SR result; and the steps at the entrance cannot be
distinguished from each other in LR images, though
there exists some thermal noise, the geometric
shape of these steps becomes clear in the SR result.
Experiment 2: FLIR thermal infrared image
sequence 2 is used in this experiment, with the
original resolution of 324 256 and the SR factor
of 4. The visual effects before and after SR
reconstruction are shown in Figure 8, the obvious
contrast area on both the LR input and the
SR result are scaled respectively to better compare
the visual saliency of the target before and
after reconstruction.
Due to the lack of enough resolution and contrast,
it is difficult to observe the contour of different
infrared targets inside input LR image, as
shown in Figure 8. Through reconstruction, the
real resolution of the image becomes higher,
while its texture structures of the lawn and the
buildings being clearer. In the red box marked
area, the weak target cannot be identified in the
LR image, however, it is clear to see a manhole
cover in the reconstruction result.
IEEE Consumer Electronics Magazine

https://projects.asl.ethz.ch/datasets https://projects.asl.ethz.ch/datasets

IEEE Consumer Electronics Magazine - March/April 2023

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - March/April 2023