88
L. Fang and S. Li
Fig. 4.10 Outline of the SBSDI framework
where T is the maximum number of nonzero coefficients in α
i+w
L ,L , and the position
of the nonzero coefficients in
ˆ
α
i+w
L ,L
W
w−W
are the same while coefficient values
become different. The SOMP algorithm [50] can be employed to efficiently solve
the above problem. Then, the joint operation can reconstruct the current HH patch
as: ˆ
x
i
H,H
W
w−W
b
w
i ˆ
x
i+w
H,H , where ˆ
x
i+w
H,H D
A
H,H M
A
i ˆ
α
i+w
L ,L is the estimated patch and
b
w
i is the weight, computed by
b
w
i exp
−
x
i+w
L ,L − x
i
L ,L
2
2
/ h
/N orm.
(4.16)
N orm is a normalization factor and h is a predetermined scalar. Finally, we return
the estimated patches to their original positions to reconstruct the OCT image. The
outline of the SBSDI framework is illustrated in Fig. 4.10.
4.3.2.3 Experimental Results
The proposed SBSDI method was tested on two kinds of retinal OCT images: (1)
synthetic images created from high resolution images that were then subsampled and
(2) real test images consisting of images captured at a low sampling rate. For test
synthetic images, we subsampled the previously acquired high-resolution images
with both random and regular patterns, thus reducing the number of A-scans in
each B-scan. For real test datasets, we directly acquired low-resolution images with
a regularly sampled pattern. Both the human and mouse retinal images were used
in these experiments. All these studies followed the tenants of the Declaration of
Helsinki.