performance as, being independent of local image structure, the method is forced to
operate in the same way near an image edge or inside a uniform image patch. In this
paper, an evolution of our previous distance-based algorithm is presented, which
makes explicit use of local image representations. As in our previous approach, the
proposed system learns from examples how to perform superresolution. In this way,
the computational load is mostly displaced to the off-line learning process, enabling a
fast, non-iterative response in the network deployment phase.
The proposed method is based on the sequential application of two processing
steps. In the first step, sequence pixels are projected onto the high-resolution frame,
and local image representations are built for each site of an embedded high-resolution
grid. In the second step, the image representation coefficients in the neighborhood of
each grid site are processed by a neural network to estimate the high-resolution image
values. The dimensionality of the network input data is previously reduced by appli-
cation of a PCA (Principal Component Analysis) technique.
Out method has shown to provide excellent results over a wide range of input noise
levels. In the following sections, we detail the processing steps involved, and present
experimental results that include a quantitative comparison of several methods. In the
last section, a brief discussion of the main results obtained is presented.
Local image representation
The first step of our superresolution method computes a local image representation
for each site of the high-resolution (HR) grid to be estimated. These local representa-
tions are built using the sequence pixels values projected onto the HR grid. The pro-
jection operation requires the previous knowledge of the geometrical transformations
that relate input sequence frames. To estimate this data, an adaptation of a classical
sub-pixel registration procedure [8] has been used.
Table 1. RMS errors for different interpolation schemes.
|
Method |
σ=0 |
σ=5 |
σ=10 |
σ =20 |
|
NN SEQ |
9.28 |
10.40 |
13.48^ |
21.97 ' |
|
__________________________Distance-based interpolation_______________________ | ||||
|
Inverse distance weight_______ |
6.82 |
7.48 |
________9.23 |
14.16 |
|
MLP-PNN_____________ |
5.61 |
6.00 |
________6.85 |
________8.43 |
|
_____________________________Polynomial Models__________________________ | ||||
|
Order 1___________________ |
7.24 |
7.37 |
________7.51 |
________8.48 |
|
Order 2__________________ |
2.90 |
3.56 |
________5.04 |
________8.75 |
|
Order 3__________________ |
2.34 |
3.29 |
________5.25 |
________9.72 |
Polynomial models have been used to describe the local image structure at sub-
pixel level. Polynomial coefficients are determined by a general linear least squares
technique, with matrix inversion performed using singular value decomposition, SVD
[9]. The use of SVD improves robustness when the problem is close to singular, due,
for instance, to an inadequate model order selection in an image patch, or induced by
a large image noise level.