Aerospace and Electronic Systems - March 2019 - 21

Duan et al.

Figure 3.
Two halves of the split CSF of eagle along the perceptual orientation. (a) CSF of eagle. (b) CSFON . (c) CSFOFF .

DENOISING LAYER
After the OSNL, the energy value Es;u ðx; yÞ is obtained.
The DL is applied to Es;u ðx; yÞ to achieve the non-noisy
estimation E~s . In DL, the PoEdges which is a probabilistic
method based on wavelet statistics is utilized for image
denoising [38]. In this method, the input image is firstly
decomposed into several different subbands. Then the
wavelet coefficients are transformed into the decorrelated
domain. For each wavelet coefficients, an iterated Wiener
filter is used:
z

tþ1

À1

¼ ðI þ diagðh wÞÞ z~

ht ¼ 1

t

sﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ
ﬃ
X
2
wj ðztj Þ

(6)

(7)

j

CONTRAST SENSITIVITY BASED SURROUND
INHIBITION LAYER
In CSSIL, the surround inhibition is implemented based
on the spatial properties of eagle's CSF. The CSF is shown
in Figure 3(a). Then along the perceptual orientation u,
we used the truncated half-rings detection [39] to split the
CSF in two halves CSFON and CSFOFF , as shown in
Figure 3(b) and (c).
The specific mathematical description of this procedure is as follows:
CSFuON ðx; yÞ ¼ CSFðx; yÞ Á jx cos u þ y sin u À ajþ
CSFuOFF ðx; yÞ ¼ CSFðx; yÞ Á ja À x cos u À y sin ujþ
(8)

MARCH 2019

The convolutions of the orientation of each pixel are calculated and the minimum value is defined as the inhibition
term. The specific expression of this procedure is as follows:
SIðx; yÞ ¼ minfE~s Ã CSFuON ; E~s Ã CSFuOFF g

(10)

Typically, when there are consistent characteristics in
center and surround, the inhibition will be very strong,
otherwise, the inhibition will get weak even disappear.
Therefore, we use the gray-level co-occurrence matrix
(GLCM) method [40] to construct the texture weighting.
For the pixel ðx; yÞ, the texture weight v is determined by:
vðx; yÞ ¼ expðÀtjjx À yjj2 Þ

where z and z~ represent the clean and noisy wavelet coefficients respectively, and w denotes the scale of wavelet
coefficients which is estimated by moment matching [38].
By inverting the wavelet transforms we can reconstruct
the denoised image.

(

where the excluded band area is determined by the parameter a and the operator j Á jþ is defined as:

d;
d < 0
jdjþ ¼
(9)
0;
d 0

(11)

where the parameter t controls the strength of texture
suppression. x ¼ ½contrastc ; energyc ; homogeneityc T and
y ¼ ½contrasts ; energys ; homogeneitys T are the feature
vectors of surround and center, respectively. The introduced
texture weight can suppress the texture edges effectively
under the condition of retaining salient contours of target.
Through the contrast sensitivity based surround inhibition
layer the results can be calculated by:
cs ðx; yÞ ¼ jE~s ðx; yÞ À k Á vðx; yÞ Á SIðx; yÞjþ

(12)

where the strength of the surround inhibition can be controlled by adjusting the inhibition factor k.

MULTIRESOLUTION FUSION LAYER
Through the handle of the three layers before, we acquire
the corresponding contours of four different scales s.
Then the MFL will be employed to fuse these obtained
contours together to form the salient contour of image. In
this process, we first utilize the non-maxima suppression
followed by hysteresis thresholding to construct the binary
contour maps bs ðx; yÞ [41]. Then the logic AND operator

IEEE A&E SYSTEMS MAGAZINE

21
Aerospace and Electronic Systems - March 2019

Table of Contents for the Digital Edition of Aerospace and Electronic Systems - March 2019
