16-05-2012, 05:05 PM
Robust super resolution of compressed video
[attachment=22190]
Introduction
Super-resolution involves the problem of inferring a highresolution
(HR) image (or a sequence) from observed multiple
low-resolution (LR) images or a segment of a video,
without changing the resolution of the image sense (see
[20] and [26] for reviews). There are mainly two categories
of SR algorithms, i.e., reconstruction-based algorithms
and learning-based algorithms. Recently, learningbased
methods [4, 8, 10, 12–14, 17, 34, 35] have been successfully
applied to SR algorithms for images and videos.
Xiong et al. [35] combine adaptive partial differential equations
(PDE) regularization with learning-based pair matching
to eliminate the compression artifacts and meanwhile
to preserve and enhance the high-frequency details. Before
starting learning-based SR, decompressing the compressed
videos is a necessary step, during which, some useful information
in the compressed bitstream may be lost. So for
images with rich texture regions, neither PDE regularization
nor learning-based SR works well [35]. Furthermore,
the example-based HR results largely depend on the similarity
between the input image and the samples in the database,
and may introduce new noise from the training set.
Modified acquisition and compression model
For compressed LR videos to be decompressed, we can collect
the following information: the LR images, motion compensated
predictions, quantized transform coefficients, motion
vectors, and quantized steps. The information of the
neighboring frames, from frame k − TB to frame k + TF,
serve as the input to reconstruct the HR solution fk .
Robust super resolution
Our algorithm can be summarized as follows: we first isolate
the frames individually and get their corresponding initial
super-resolution estimates within the Bayesian framework
by exploiting the information available in the compressed
bitstream. Then we use a robust optical flow algorithm [5]
to estimate the motion vectors between the estimated initial
HR frames. Finally, a final SR image can be obtained with
the neighboring initial estimates.
[attachment=22190]
Introduction
Super-resolution involves the problem of inferring a highresolution
(HR) image (or a sequence) from observed multiple
low-resolution (LR) images or a segment of a video,
without changing the resolution of the image sense (see
[20] and [26] for reviews). There are mainly two categories
of SR algorithms, i.e., reconstruction-based algorithms
and learning-based algorithms. Recently, learningbased
methods [4, 8, 10, 12–14, 17, 34, 35] have been successfully
applied to SR algorithms for images and videos.
Xiong et al. [35] combine adaptive partial differential equations
(PDE) regularization with learning-based pair matching
to eliminate the compression artifacts and meanwhile
to preserve and enhance the high-frequency details. Before
starting learning-based SR, decompressing the compressed
videos is a necessary step, during which, some useful information
in the compressed bitstream may be lost. So for
images with rich texture regions, neither PDE regularization
nor learning-based SR works well [35]. Furthermore,
the example-based HR results largely depend on the similarity
between the input image and the samples in the database,
and may introduce new noise from the training set.
Modified acquisition and compression model
For compressed LR videos to be decompressed, we can collect
the following information: the LR images, motion compensated
predictions, quantized transform coefficients, motion
vectors, and quantized steps. The information of the
neighboring frames, from frame k − TB to frame k + TF,
serve as the input to reconstruct the HR solution fk .
Robust super resolution
Our algorithm can be summarized as follows: we first isolate
the frames individually and get their corresponding initial
super-resolution estimates within the Bayesian framework
by exploiting the information available in the compressed
bitstream. Then we use a robust optical flow algorithm [5]
to estimate the motion vectors between the estimated initial
HR frames. Finally, a final SR image can be obtained with
the neighboring initial estimates.