78
L. Fang and S. Li
Fig. 4.2 Dictionaries trained on a the original low-SNR OCT image by the K-SVD algorithm,
b averaged high-SNR image by the K-SVD algorithm, and c averaged high-SNR image by the
proposed MSBTD training algorithm (due to the limited space, only the first atom of each learned
subdictionary is shown)
Upsampled and
downsampled
Clustering
Structural
dictionary
learning
Training image
Patches extracted from
images at different scales
Upsampled Image
(Finer scale)
Multiscale
structural
dictionary
Structural clusters
at different scales
Downsamped image
(Coarser scale)
1,1
Mstr
D
K,S
Mstr
D
K,1
Mstr
D
1,
Mstr
s
D
K,
Mstr
s
D
1,S
Mstr
D
Cluster 1, 1 Cluster K,1
s
Cluster 1,
s
Cluster K,
Cluster 1, S Cluster K, S
Fig. 4.3 Algorithmic flowchart of multiscale structural dictionary learning process
multiscale structural dictionary to sparsely represent the patch, which achieves
denoising of that patch. The denoising steps are detailed as follows.
To seek the best subdictionary among the learned subdictionaries for each noisy
patch, the representative features of the patch x i is compared with each subdictionary.
To represent each subdictionary, we use the centroid atom (c k,s ∈ R
(w·z)×1 ) of the
corresponding k-means cluster (noted in the above Section) [36]. For the representative feature of each patch, its high-frequency component (denoted by x
H f
i ) is used.
We find the best fitted subdictionary D
A
i for the patch x i based on the normalized
correlation [21] between c k,s and x
H f
i :
A (k i , s i ) arg max
k,s
c k,s , x
H f
i
.
(4.4)