209:4 M. Wang et al.
new standards to better support diverse application scenarios such as city security, online learning,
and cloud gaming.
Screen content coding tools are specically designed to facilitate the coding of computer-
generated content, orienting to the application of screen sharing, animation, gaming, and a mix-
ture of content. The new generation of video coding standards species the screen content coding
technologies as low-level coding tools, catering to the market regarding the increased capacity
of the non-camera captured content. In VVC, ve screen content coding tools are designed [30],
including the Transform Skip with Residual Coding (TSRC) [29], Block-based Dierential
Pulse-Coded Modulation (BDPCM) [2], Intra Block Copy (IBC) [50], Adaptive Color Trans-
form (ACT) [60] and the palette mode [36]. The IBC, palette mode, and ACT are inherited from
the predecessor HEVC screen content coding extensions. The VVC is anticipated to become the
main solution for screen content and mixture content coding, owing to the considerable coding
performance improvement over the previous standard.
2.2 Image Super Resolution and Restoration
Numerous eorts have been dedicated to the restoration of the high-quality images where the low-
quality to high-quality mapping relationship could be eectively represented through deep neural
networks. To be more specic, with deep neural networks, low-level features are extracted and
formulated from the low-quality input LR image, approaching the HR by feature accumulation and
reorganization. Learning-based SR schemes have successfully surpassed the traditional SR schemes
from the perspective of the quantitative and qualitative evaluations, as the intrinsic connections
between the LR and HR could be well understood by the neural network.
Beyond the classical SRCNN, Kim et al. [16] proposed to progressively increase the network
depth, with the goal of exploring the restoration capability of the large-scale model. The net-
work design exhibits to be eective in enhancing the performance of the learning-based SR, such
that bundles of works have been proposed, including the residual networks, densely connected
networks, attention-based mechanism, recursive learning, and transformer networks. Enhanced
Deep Super Resolution (EDSR) network was investigated by Lim et al. [25], wherein the struc-
ture of the residual net [15] was modied to better adapt to the low-level recovery task. In [61], the
attention mechanism is involved in the SR network wherein the feature channels are adaptively re-
scaled according to the attention allocation. Residual dense network [63] was proposed for image
SR which exploited the hierarchical features from the LR with densely connected convolutional
layers. Yang et al. [57] investigated a deep edge guided recurrent residual network to progres-
sively compensate the high-frequency information, which could well handle JPEG artifacts during
high resolution recovery. Anwar et al. [3]proposedtheDensely Residual Laplacian Network
(DRLN), which exploits features at multiple scales with cascading residual structure, densely con-
nected structure, and Laplacian attention model. In [20], symmetrical residual connection struc-
tures are explored, which employ symmetrical nested residual connections with multiple paths,
leading to the enhancement of the restoration performance and the increase of computing speed.
Kernel attention network [58] is investigated for single image super resolution wherein the net-
work could adjust the receptive eld size by changing the input scales and kernel selection. In this
way, the network can learn the distinguished features through multi-scale perceiving. Recently,
Ma et al. [27] propose a meta-learning-based fusion network, which is capable of generating an
HR image by fusing deterministic and stochastic images, resulting the improvement of the per-
ceptual quality. A sequential hierarchical SR network is studied which considers the correlations
among features with dierent scales [26]. Meanwhile, image restoration, especially in terms of
resolving the compression distorted images, has attracted much attention. The restoration could
be implemented as out-loop lters for the reconstructed images, aiming at eliminating the coding
ACM Trans. Multimedia Comput. Commun. Appl., Vol. 19, No. 6, Article 209. Publication date: July 2023.