4 Reconstruction of Retinal OCT Images with Sparse Representation
77
principle component analysis (PCA) algorithm is utilized to train a subdictionary
from each of the K clusters.
To effectively exploit the properties of different structures and textures on ocular
OCT images (e.g. each retinal layer has a different thickness and various kinds
of pathology), different scales information should be considered in the dictionary
training process. To achieve this, a multiscale strategy is incorporated into the above
structural learning process. Specifically, the training image is first zoomed in and out
via upsampling and downsampling processes. Then, the original and these magnified
images are divided into same sized patches. In this way, although the patch size is
fixed, patches from different magnified scales can be considered as variable sized
patches from a particular scale. Next, the structural learning process is applied on
the patches from the same scale (s) to create the multiscale structural dictionary,
which is the concatenation of the subdictionaries (
D
Mstr
k,s
, s 1, 2, . . . , S) from
all scales. A schematic representation of the multiscale structural dictionary learning
process is illustrated in Fig. 4.3. The upsampling and downsampling processes are
implemented by the bilinear interpolation.
4.3.1.2 Nonlocal Denoising Process
This subsection introduces how to utilize the learned dictionary for denoising OCT
images. For each patch x i , we find an appropriate subdictionary D
A
i from the learned
Fig. 4.1 Creation of a less
noisy (averaged) frame by an
average operation on the
multiple frames captured
form a unique position
77
principle component analysis (PCA) algorithm is utilized to train a subdictionary
from each of the K clusters.
To effectively exploit the properties of different structures and textures on ocular
OCT images (e.g. each retinal layer has a different thickness and various kinds
of pathology), different scales information should be considered in the dictionary
training process. To achieve this, a multiscale strategy is incorporated into the above
structural learning process. Specifically, the training image is first zoomed in and out
via upsampling and downsampling processes. Then, the original and these magnified
images are divided into same sized patches. In this way, although the patch size is
fixed, patches from different magnified scales can be considered as variable sized
patches from a particular scale. Next, the structural learning process is applied on
the patches from the same scale (s) to create the multiscale structural dictionary,
which is the concatenation of the subdictionaries (
D
Mstr
k,s
, s 1, 2, . . . , S) from
all scales. A schematic representation of the multiscale structural dictionary learning
process is illustrated in Fig. 4.3. The upsampling and downsampling processes are
implemented by the bilinear interpolation.
4.3.1.2 Nonlocal Denoising Process
This subsection introduces how to utilize the learned dictionary for denoising OCT
images. For each patch x i , we find an appropriate subdictionary D
A
i from the learned
Fig. 4.1 Creation of a less
noisy (averaged) frame by an
average operation on the
multiple frames captured
form a unique position