11-08-2012, 05:02 PM
Robust Web Image/Video Super-Resolution
Robust Web Image.pdf (Size: 1.67 MB / Downloads: 42)
INTRODUCTION
WITH the Internet flourishing and the rapid progress in
hand-held photographic devices, image and video are
becoming more and more popular on the web, due to their rich
content and easy perception. Consequently, image search engines
and online video websites have experienced an explosion
of visits during the past few years. However, limited by the
network bandwidth and server storage, most web image/video
exists in a low quality version degraded from the source. The
most common degradations are downsampling and compression.
Downsampling exploits the correlation in the spatial domain
while compression further exploits the correlation in the
frequency and temporal (for video) domains. Quality degradation
greatly lowers the required bandwidth and storage, making
the access to web image/video practical and convenient.
PDE Regularization
Among various available regularization techniques,
anisotropic PDE’s [27]–[30] are considered to be one of
the best, due to their ability to smooth data while preserving
visually salient features in images. A brief restatement of PDE
regularization is given below. Suppose is a 2D scalar image,
the PDE regularization can be formulated as the juxtaposition
of two oriented 1D heat flows along the gradient direction
Energy Change During Regularization
To obtain appropriate regularization strength that well balances
artifacts removal and primitive preservation, we propose
to investigate the image energy change characteristics during
the iterative regularization process. For this purpose, an image
is first divided into primitive field and nonprimitive field. This
partition can be determined by the orientation energy edge detection
COMPRESSED VIDEO SUPER-RESOLUTION
Since the above introduced single-image SR method does
not require frame registration or motion estimation, it can be
directly applied into the compressed video SR in a frame-byframe
style. By integrating certain interframe interactions on the
regularization strength and simple spatio-temporal coherency
constraints, our scheme is competent for the SR task of web
videos with dynamic content and different degradation levels.
The framework of our solution is shown in Fig. 9. Similar to
that of image SR, it consists of three steps. First, a th frame
from an LR video is divided into PF and NPF and iterative
PDE regularization is performed on , during which the energy
change velocities in both PF and NPF are recorded. When
the ratio of these two velocities converges (judged by a parameter
, which is also influenced by that of the previous frame
), regularization stops and the accumulated noise image
is subtracted from , resulting in an artifacts-relieved
frame . Then, is upsampled to the desired resolution
through bicubic interpolation.
Video Results
Our solution for compressed video SR is tested on a variety of
web videos downloaded from YouTube [35]. They are generally
in a 320 240 resolution but with different degradation levels.
We perform a uniform on them, still using the above
database. In the pair matching stage, the candidate number of
the enhancing patch , and the SAD threshold .
Fig. 15 shows three frames extracted from a super-resolved
web cartoon video. This result demonstrates the effectiveness
of our solution in three aspects. First, the total iteration number
of regularization, as enclosed in the caption, is appropriately
dependent on the degradation level of each frame.