Goals
Generating huge amounts of visual data, be it images or videos, has
never been easier than today. This creates a growing demand for lossy
codecs (coders and decoders) that produce visually convincing results
also for very high compression rates. Popular transform-based codecs
such as JPEG and JPEG 2000 have reached a state where one cannot expect
significant improvements anymore. To go beyond their limitations,
fundamentally different ideas are needed.
Inpainting-based codecs can change this situation. They store only a
small, carefully optimised part of the data. In the decoding step, the
missing information is filled in with a suitable inpainting mechanism.
A successful realisation of inpainting-based codecs can offer decisive
advantages over transform-based codecs: The stored information is more
intuitive and closer to the mechanisms of human perception. Moreover,
the concept is very flexible: It allows to integrate a number of
different features and can be tailored towards dedicated applications.
Most importantly, the higher the compression rate, the larger are the
qualitative advantages over transform-based codecs. However, the
potential of these codecs was widely unexplored, since difficult
fundamental problems had to be solved first.
The INCOVID project has addressed these challenges in an integrated
approach that covers a wide spectrum of aspects. We have established
important theoretical foundations, introduced better data selection
strategies, developed dedicated codecs for specific applications,
implemented highly efficient numerical algorithms, and created a
demonstrator that offers real-time performance in 4K resolution. These
achievements show that inpainting methods are far more than a visually
pleasant image editing tool: They have become the key component of a
well-founded and powerful alternative paradigm in coding. It can play
an important role in future codecs that we will use in our daily life.
This project was funded by the EU through the ERC Advanced Grant
INCOVID for Joachim Weickert (Grant Agreement No. 741215). Its
budget for research was 2.46 million Euro, to be spent over a
period of five years (from October 1, 2017 until September 30,
2022).
Main Achievements
The INCOVID project has accomplished its envisioned goals in all four
key areas: It has led to fundamental theoretical insights, novel
approaches with substantial quality improvements, extensions to new
application areas, and highly efficient algorithms that allow a real-time
demonstrator. Here is a list of the highlights:
I. Establishing the Foundations
-
We have introduced a general framework for pseudodifferential inpainting.
This new inpainting paradigm connects inpainting using linear shift
invariant differential operators with radial basis function interpolation
(Augustin et al. 2019).
-
Among the nonlinear inpainting operators, edge-enhancing anisotropic
diffusion (EED) is known to performs particularly well. We have
established a rigorous existence theory for this nonlinear
integrodifferential equation (Bildhauer et al. 2021).
-
We have laid the theoretical foundations of inpainting-based codecs
in terms of sparsification and quantisation scale-spaces
(Cárdenas et al. 2019, Peter 2021).
This has strengthened ties between two hitherto fairly unconnected
scientific communities: the coding community and the scale-space
community
-
We have extended the Mumford-Shah model to image compression. The
resulting energy functional serves as an important step towards a
full optimisation model that also incorporates the coding cost
(Jost et al. 2020).
II. Pushing the Quality to the Limit
-
The aforementioned Mumford-Shah based codec of Jost et al. (2020)
is particularly well-suited for piecewise smooth
data, e.g. depth maps. Here it significantly outperformed
state-of-the-art codecs such as HEVC by up to 3 dB.
-
We have pioneered data optimisation for exemplar-based inpainting.
It introduces densification by dithering of error maps, and it
allows an inpainting-based sparse representaton of highly textured
data in unprecented quality (Karos et al. 2018).
-
Deep learning methods for selecting inpainting data have been
(Alt et al. 2022, Peter et al. 2022, Peter et al. 2023,
Schrader et al. 2023).
They offer a similar quality as model-based approaches at
a substantially reduced runtime.
-
We have achieved a breakthrough in the simultaneous incorporation
of different feature types for inpainting-based image representations
(Jost et al. 2023). This approach is simple, fairly general,
and provides both spatial and tonal optimisation.
Moreover, it allowed us to find a novel promising feature type
based on local averages.
III. Extending the Spectrum
-
Various alternative data feature types have been evaluated in
student theses, e.g. isophotes (Nebel, 2019), local binary patterns
(Sirazitdinov, 2019), and junctions or corners (Gusenburger, 2020).
They have merits for specific imagery.
-
Inpainting-based compression has been extended to audio signals where
it offers competitive performance in terms of PSNR (Peter et al. 2019).
Moreover, we have established novel algorithms for
piecewise constant signal approximations with superior quality
(Bergerhoff et al. 2019).
-
As a component for inpainting-based video compression, we have
proposed a novel edge-aware inpainting technique that allows to
represent motion fields in high quality (Jost et al. 2020).
Moreover, we have introduced a video codec that outperforms previous
inpainting-based ones in both quality and speed (Andris et al., SSVM 2021).
IV. Algorithms, Prototypes, and Demonstrators
-
We have performed a systematic evaluation of coding strategies
for sparse inpainting masks (Mohideen et al. 2021), in order
to find and adapt the most powerful ones.
-
We have developed an neural surrogate algorithm for inpainting with
Euler's elastica that is simpler than classical ones, but offers
competitive quality (Schrader et al. 2022). Moreover, we
have also introduced a novel hybrid codec that combines the
advantages of inpainting- and transform-based compression methods
(Andris et al., PCS 2021).
-
Exemplar-based inpainting offers good quality for highly textured
images, but suffers from slow nonlocal algorithms. As a remedy, we
have introduced space-filling curves as acceleration structures
(Dahmen et al. 2020).
-
Highly efficient codecs have been developed that combine Shepard
interpolation with joint inpainting and prediction (Peter 2019).
-
For homogeneous diffusion inpainting, we have developed fast finite
element algorithms that exploit adaptive triangulations (Chizhov/Weickert
2021). Moreover, we have used domain decomposition methods on GPUs
(Kämper/Weickert 2022). They allow us to inpaint almost 40 colour
images in 4K resolution (3840 * 2160 pixels) in one second on a
contemporary GPU. This serves as the envisioned 4K real-time
demonstrator which concludes the INCOVID project.
Key References
-
T. Alt, P. Peter, J. Weickert:
Learning sparse masks for diffusion-based image inpainting.
In A. J. Pinho, P. Georgieva, L. F. Teixeira,
J. A. Sánchez (Eds.): Pattern Recognition and Image Analysis.
Lecture Notes in Computer Science, Vol. 13256, Springer, Cham, 528-539,
2022.
Also available as
arXiv:2110.02636 [eess.IV], revised March 2022.
-
S. Andris, P. Peter, R. M. K. Mohideen, J. Weickert, S.
Hoffmann:
Inpainting-based video compression in FullHD.
In A. Elmoataz, J. Fadili, Y. Quéau, J. Rabin, L. Simon (Eds.):
Scale Space and Variational Methods in Computer Vision. Lecture Notes
in Computer Science, Vol. 12679, Springer, Cham, 425-436, 2021.
Also available as
arXiv:2008.10273 [eess.IV], revised May 2021.
-
S. Andris, J. Weickert, T. Alt, P. Peter:
JPEG meets PDE-based image compression.
Proc. 35th Picture Coding Symposium (PCS 2021,
Bristol, UK, June 2021), IEEE, 2021.
Also available as
arXiv:2102.01138 [eess.IV], revised May 2021.
-
M. Augustin, J. Weickert, S. Andris:
Pseudodifferential inpainting: The missing link between PDE- and
RBF-based interpolation.
In J. Lellmann, M. Burger, J. Modersitzki (Eds.):
Scale Space and Variational Methods in Computer Vision.
Lecture Notes in Computer Science, Vol. 11603, 67-78, Springer, Cham, 2019.
-
L. Bergerhoff, J. Weickert, Y. Dar:
Algorithms for piecewise constant signal approximations.
Proc. 27th European Signal Processing Conference
(EUSIPCO 2019, A Coruña, Spain, Sept. 2-6, 2019), IEEE, 2019.
Also available as
arXiv:1903.01320v3 [eess.SP], June 2019.
-
M. Bildhauer, M. Cárdenas, M. Fuchs, J. Weickert:
Existence theory for the EED inpainting problem.
St. Petersburg Mathematical Journal, Vol. 32, No. 3, 481-497, May 2021.
Invited Paper.
Also available as
arXiv:1906.04628v2 [math.AP], September 2019.
-
M. Cárdenas, P. Peter, J. Weickert:
Sparsification scale-spaces.
In J. Lellmann, M. Burger, J. Modersitzki (Eds.):
Scale Space and Variational Methods in Computer Vision.
Lecture Notes in Computer Science, Vol. 11603, 303-314, Springer, Cham, 2019.
-
V. Chizhov, J. Weickert:
Efficient data optimisation for harmonic inpainting with finite
elements.
In N. Tsapatsoulis, A. Panayides, T. Theocharides, A. Lanitis,
C.S. Pattichis, M. Vento (Eds.): Computer Analysis of Images and Patterns.
Part 2. Lecture Notes in Computer Science, Vol. 13053, Springer, Cham,
432-441, 2021.
Also available as
arXiv:2105.01586 [eess.IV], revised July 2021.
-
T. Dahmen, P. Trampert, P. Peter, P. Bheed, J. Weickert, P. Slusallek:
Space-Filling Curve Indices as Acceleration Structure for
Exemplar-Based Inpainting.
arXiv:1712.06326 [cs.CV], January 2020.
-
F. Jost, V. Chizhov, J. Weickert:
Optimising different feature types for inpainting-based image
representations.
Proc. 2023 IEEE International Conference on Acoustics,
Speech, and Signal Processing (ICASSP 2023, Rhodes, Greece,
June 2023), 2023.
Also available as
arXiv:2210.14949 [eess.IV], October 2022.
-
F. Jost, P. Peter, J. Weickert:
Compressing piecewise smooth images with the Mumford-Shah cartoon
model.
Proc. 28th European Signal Processing Conference
(EUSIPCO 2020, Amsterdam, Netherlands, January 2021), 511-515, 2021.
Also available as
arXiv:2003.05206 [eess.IV], March 2020.
-
N. Kämper, J. Weickert:
Domain decomposition algorithms for real-time homogeneous diffusion
inpainting in 4K.
Proc. 2022 IEEE International Conference on Acoustics,
Speech and Signal Processing (ICASSP 2022, Singapore, May 2022),
1680-1684, 2022.
Also available as
arXiv:2110.03946 [eess.IV], revised February 2022.
-
L. Karos, P. Bheed, P. Peter, J. Weickert:
Optimising data for exemplar-based inpainting.
In J. Blanc-Talon, D. Helbert, W. Philips, D. Popescu, P. Scheunders
(Eds.): Advanced Concepts for Intelligent Vision Systems.
Lecture Notes in Computer Science, Vol. 11182, 547-558, Springer,
Cham, 2018.
-
R. M. K. Mohideen, P. Peter, J. Weickert:
A systematic evaluation of coding strategies for sparse binary
images.
Signal Processing: Image Communication, Vol. 99, Article 116424,
November 2021.
Also available as
arXiv:2010.13634 [eess.IV], revised July 2021.
-
P. Peter:
Fast inpainting-based compression: Combining Shepard interpolation
with joint inpainting and prediction.
Proc. 2019 IEEE International Conference on Image Processing
(ICIP 2019, Taipei, Taiwan, Sept. 2019), 3557-3561, 2019.
-
P. Peter:
Quantisation scale-spaces.
In A. Elmoataz, J. Fadili, Y. Quéau, J. Rabin, L. Simon (Eds.):
Scale Space and Variational Methods in Computer Vision. Lecture Notes
in Computer Science, Vol. 12679, Springer, Cham, 15-26, 2021.
Also available as
arXiv:2103.10491 [eess.IV], March 2021.
-
P. Peter:
A Wasserstein GAN for joint learning of inpainting and its spatial
optimisation.
In H. Wang, W. Lin, P. Manoranjan, G. Xiao, K.L. Chan, X. Wang,
G. Ping, H. Jiang: Image and Video Technology.
Lecture Notes in Computer Science, Vol. 13763,
Springer, Cham, 132-145, 2023.
Also available as
arXiv:2202.05623 [eess.IV], February 2022.
-
P. Peter, J. Contelly, J. Weickert:
Compressing audio signals with inpainting-based sparsification.
In J. Lellmann, M. Burger, J. Modersitzki (Eds.):
Scale Space and Variational Methods in Computer Vision.
Lecture Notes in Computer Science, Vol. 11603, 92-103, Springer, Cham, 2019.
-
P. Peter, K. Schrader, T. Alt, J. Weickert:
Deep spatial and tonal optimisation for homogeneous diffusion
inpainting.
Pattern Analysis and Applications, 2023.
Invited Paper.
Also available as
arXiv:2208.14371 [eess.IV], revised March 2023.
-
K. Schaefer, J. Weickert:
Diffusion-shock inpainting.
In L. Calatroni, M. Donatelli, S. Morigi, M. Prato, M. Santavesaria (Eds.):
Scale Space and Variational Methods in Computer Vision. Lecture Notes
in Computer Science, Vol. 14009, Springer, Cham, 588-600, 2023.
Also available as
arXiv:2303.09450 [eess.IV], March 2023.
-
K. Schrader, T. Alt, J. Weickert, M. Ertel:
CNN-based Euler's elastica inpainting with deep energy and
deep image prior.
Proc. 10th European Workshop on Visual Information
Processing (EUVIP 2022, Lisbon, Portugal, Sept. 2022), IEEE, 2022.
Also available as
arXiv:2207.07921 [cs.CV], July 2022.
-
K. Schrader, P. Peter, N. Kämper, J. Weickert:
Efficient neural generation of 4K masks for homogeneous diffusion
inpainting.
In L. Calatroni, M. Donatelli, S. Morigi, M. Prato, M. Santavesaria (Eds.):
Scale Space and Variational Methods in Computer Vision. Lecture Notes
in Computer Science, Vol. 14009, Springer, Cham, 16-28, 2023.
Also available as
arXiv:2303.10096 [eess.IV], March 2023.
|