Decemb er 10, 2010 / Vol. 8, No. 12 / CHINESE OPTICS LETTERS 1113
Ghosting reduction in scene-based nonuniformity correction
of infrared image sequences
Junqi Bai (白白白俊俊俊奇奇奇), Qian Chen (陈陈陈 钱钱钱), Weixian Qian (钱钱钱惟惟惟贤贤贤), and Xianya Wang (王王王娴娴娴雅雅雅)
∗
441 Lab, Scho ol of Electronic Engineering and Opto electronic Techniques, Nanjing University of Science and Technology,
Nanjing 210094, China
∗
E-mail: xiaotuo1212@yaho o.com.cn
Received May 24, 2010
Scene-based adaptive nonuniformity correction (NUC) is currently being applied to achieve higher perfor-
mance in infrared imaging systems. However, almost all scene-based NUC algorithms cause the production
of ghosting artifacts over output images. Based on constant-statistics theory, we propose a novel threshold
self-adaptive ghosting reduction algorithm to improve the space low-pass and temporal high-pass (SLP-
THP) NUC technique. The correction parameters of the previous frame are regarded as thresholds to
compute new correction parameters. Experimental results show that the proposed algorithm can obtain a
satisfactory performance in reducing unwanted ghosting artifacts.
OCIS codes: 040.3060, 100.2550, 100.2960, 100.2980.
doi: 10.3788/COL20100812.1113.
Infrared (IR) imaging systems have been widely used in
both military and civilian fields. However, fixed pattern
noise caused by nonuniform resp onse of detectors is an
intrinsic shortcoming of IR imaging. Nonuniformity cor-
rection (NUC) techniques were developed to perform the
necessary calibration. Reference-based corrections us-
ing calibrated images on startup cannot solve the drift
in the parameters of the detectors over time. As a re-
sult, scene-based corrections
[1−3]
have been studied to
continuously correct IR nonuniformity without using ref-
erences and interrupting detections. However, ghosting
artifacts are major problems of scene-based NUC. The
best known NUC technique based on temporal high-pass
(THP) filters is highly dependent on object motion. If an
object in the image moves slowly or remains stationary
for a large number of iterations, it will cause the produc-
tion of ghosting artifacts over the output images.
To reduce the ghost effect, Harris et al. pro-
posed a simple de-ghosting module in constant-statistics
NUC
[3]
, wherein the correction parameters will not be
updated until the changes of each pixel are greater
than the threshold. Aiming at the NUC based on neu-
ral network
[4]
, Esteban et al. presented two different
adaptive learning rate strategies to reduce ghosting
artifacts
[5]
. Recently, Qian et al. proposed an interesting
space low-pass and THP (SLP-THP) NUC to eliminate
ghosting artifacts
[6]
. In the abovementioned NUC tech-
niques, a proper threshold is important to reduce ghost-
ing artifacts effectively, but choosing the proper thresh-
old is a very difficult task. The SLP-THP NUC was
taken into consideration because using high threshold will
not remove enough of the ghost artifacts, while using low
threshold will cause few images for the estimation of cor-
rection parameters.
Aiming at the threshold problem of SLP-THP NUC,
this letter proposes a novel threshold self-adaptive ghost-
ing reduction NUC based on constant-statistics theory
[7]
.
We adjust the threshold in our algorithm by using the
correction parameters of the previous frame.
THP NUC is carried out by computing the difference
between the input image and the temporal low-pass filter
image, which can be expressed as
y(k) = x(k) − f(k), (1)
f(k) = x(k)/N + (1 − 1/N ) × f (k − 1), (2)
where x(k) is the input image, y(k) is the output image,
f(k) is the average of x(k) in the time domain, and N is
the number of accumulable frames.
In THP NUC, correction parameters are estimated us-
ing the whole space frequency of the input image. In
view of the space characteristic of nonuniformity (which
mainly exists in high-space frequency of image, as shown
in Fig. 1), SLP-THP NUC divides the space frequency
of image into two parts: low-space frequency and high-
space frequency.
Assuming x = z + nu, x is the input image, z is the
original image, and nu is the nonuniformity. Hence, x, z,
and nu can be expressed as
x(k) =
x
low
(k) +
x
high
(k), (3)
z(k) =
z
low
(k) +
z
high
(k), (4)
Fig. 1. Different space frequency images. (a) Input image;
(b) low-space frequency; (c) high-space frequency.
1671-7694/2010/121113-04
c
° 2010 Chinese Optics Letters