txt/2204.14047.txt

A Deep Learning based No-reference Quality Assessment Model
for UGC Videos
Wei Sun
Shanghai Jiao Tong University
Shanghai, China
[email protected] Min
Shanghai Jiao Tong University
Shanghai, China
[email protected]
Wei Lu
Shanghai Jiao Tong University
Shanghai, China
[email protected] Zhai∗
Shanghai Jiao Tong University
Shanghai, China
[email protected]
ABSTRACT
Quality assessment for User Generated Content (UGC) videos plays
an important role in ensuring the viewing experience of end-users.
Previous UGC video quality assessment (VQA) studies either use
the image recognition model or the image quality assessment (IQA)
models to extract frame-level features of UGC videos for quality
regression, which are regarded as the sub-optimal solutions be-
cause of the domain shifts between these tasks and the UGC VQA
task. In this paper, we propose a very simple but effective UGC
VQA model, which tries to address this problem by training an
end-to-end spatial feature extraction network to directly learn the
quality-aware spatial feature representation from raw pixels of the
video frames. We also extract the motion features to measure the
temporal-related distortions that the spatial features cannot model.
The proposed model utilizes very sparse frames to extract spatial
features and dense frames (i.e. the video chunk) with a very low
spatial resolution to extract motion features, which thereby has
low computational complexity. With the better quality-aware fea-
tures, we only use the simple multilayer perception layer (MLP)
network to regress them into the chunk-level quality scores, and
then the temporal average pooling strategy is adopted to obtain
the video-level quality score. We further introduce a multi-scale
quality fusion strategy to solve the problem of VQA across differ-
ent spatial resolutions, where the multi-scale weights are obtained
from the contrast sensitivity function of the human visual system.
The experimental results show that the proposed model achieves
the best performance on five popular UGC VQA databases, which
demonstrates the effectiveness of the proposed model. The code is
available at https://github.com/sunwei925/SimpleVQA.
CCS CONCEPTS
•Computing methodologies →Modeling methodologies .
∗Corresponding author: Guangtao Zhai.
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from [email protected].
MM ’22, October 10–14, 2022, Lisboa, Portugal
©2022 Association for Computing Machinery.
ACM ISBN 978-1-4503-9203-7/22/10. . . $15.00
https://doi.org/10.1145/3503161.3548329KEYWORDS
video quality assessment, UGC videos, deep learning, feature fusion
ACM Reference Format:
Wei Sun, Xiongkuo Min, Wei Lu, and Guangtao Zhai∗. 2022. A Deep Learn-
ing based No-reference Quality Assessment Model for UGC Videos. In Pro-
ceedings of the 30th ACM International Conference on Multimedia (MM ’22),
October 10–14, 2022, Lisboa, Portugal. ACM, New York, NY, USA, 10 pages.
https://doi.org/10.1145/3503161.3548329
1 INTRODUCTION
With the proliferation of mobile devices and wireless networks in
recent years, User Generated Content (UGC) videos have exploded
over the Internet. It has become a popular daily activity for the gen-
eral public to create, view, and share UGC videos through various
social media applications such as YouTube, TikTok, etc. However,
UGC videos are captured by a wide variety of consumers, ranging
from professional photographers to amateur users, which makes
the visual quality of UGC videos vary greatly. In order to ensure
the Quality of Experience (QoE) of end-users, the service providers
need to monitor the quality of UGC videos in the entire streaming
media link, including but not limited to video uploading, compress-
ing, post-processing, transmitting, etc. Therefore, with billions of
video viewing and millions of newly uploaded UGC videos every
day, an effective and efficient video quality assessment (VQA) model
is needed to measure the perceptual quality of UGC videos.
Objective VQA can be divided into full-reference (FR), reduced-
reference (RR), and no-reference (NR) according to the amount of
pristine video information needed. Since there is no reference video
for in-the-wild UGC videos, only NR VQA models are qualified for
evaluating their quality. Although NR VQA algorithms [ 21,23,26]
have been studied for many years, most of them were developed
for Professionally Generated Content (PGC) videos with synthetic
distortions, where the pristine PGC videos are shot by photogra-
phers using professional devices and are normally of high quality,
and the distorted PGC videos are then degraded by specific video
processing algorithms such as video compression, transmission, etc.
So, previous VQA studies mainly focus on modeling several types
of distortions caused by specific algorithms, which makes them less
effective for UGC videos with in-the-wild distortions. To be more
specific, the emerging UGC videos pose the following challenges
to the existing VQA algorithms for PGC videos:arXiv:2204.14047v2  [cs.CV]  20 Oct 2022MM ’22, October 10–14, 2022, Lisboa, Portugal Wei Sun, Xiongkuo Min, Wei Lu, and Guangtao Zhai∗
First, the distortion types of UGC videos are diverse. A mass
of UGC videos are captured by amateur users, which may suffer
various distortion types such as under/over exposure, low visibility,
jitter, noise, color shift, etc. These authentic distortions are intro-
duced in the shooting processing and cannot be modeled by the
single distortion type, which thereby requires that the VQA models
have a more strong feature representation ability to qualify the
authentic distortions. Second, the content and forms of UGC videos
are extremely rich. UGC videos can be natural scenes, animation
[35], games [ 46,47], screen content, etc. Note that the statistics
characteristics of different video content vary greatly. For example,
the natural scenes statistics (NSS) features [ 22–24,26] are com-
monly used in the previous VQA studies to measure the distortions
of natural scene content, but they may be ineffective for computer-
generated content like animation or games. In addition, live videos,
videoconferencing, etc. are also ubiquitous for UGC videos nowa-
days, whose quality is severely affected by the network bandwidth.
Third, due to the advancement of shooting devices, more high res-
olution [ 18] and high frame rate [ 19,52,53] videos have emerged
on the Internet. The various kinds of resolutions and frame rates
are also important factors for video quality. What’s more, users
can view the UGC videos through mobile devices anywhere and at
any time, so the display [ 25] and the viewing environment such as
ambient luminance [ 29], etc. also affect the perceptual quality of
UGC videos to a certain extent. However, these factors are rarely
considered by previous studies.
The recently released large-scale UGC VQA databases such as
KoNViD-1k [ 8], YouTube UGC [ 36], LSVQ [ 44], etc. have greatly
promoted the development of UGC VQA. Several deep learning
based NR VQA models [ 14,15,37,40,44] have been proposed to
solve some challenges mentioned above and achieve pretty good
performance. However, there are still some problems that need to
be addressed. First, the previous studies either use the image recog-
nition model [ 15][44] or the pretrained image quality assessment
(IQA) models [ 37][40][14] to extract frame-level features, which
lacks an end-to-end learning method to learn the quality-aware
spatial feature representation from raw pixels of video frames. Sec-
ond, previous studies usually extract the features from all video
frames and have a very high computational complexity, making
them difficult to apply to real-world scenarios. Since there is much
redundancy spatial information between the adjacent frames, we
argue that there is not necessary to extract the features from all
frames. Third, the spatial resolution and frame rate of UGC videos
as well as other factors such as the display, viewing environment,
etc. are still rarely considered by these studies. However, these
factors are very important for the perceptual quality of UGC videos
since the contrast sensitivity of the human visual system (HVS) is
affected by them.
In this paper, to address the challenges mentioned above, we
propose a very simple but effective deep learning based VQA model
for UGC videos. The proposed framework is illustrated in Figure 1,
which consists of the feature extraction module, the quality regres-
sion module, and the quality pooling module. For the feature ex-
traction module, we extract quality-aware features from the spatial
domain and the spatial-temporal domain to respectively measure
the spatial distortions and motion distortions. Instead of using thepretrained model to extract the spatial features in the previous stud-
ies, we propose to train an end-to-end spatial feature extraction
network to learn quality-aware feature representation in the spatial
domain, which thereby makes full use of various video content
and distortion types in current UGC VQA databases. We then uti-
lize the action recognition network to extract the motion features,
which can make up the temporal-related distortions that the spatial
features cannot model. Considering that the spatial features are
sensitive to the resolution while the motion features are sensitive
to the frame rate, we first split the video into continuous chunks
and then extract the spatial features and motion features by using a
key frame of each chunk and all frames of each chunk but at a low
spatial resolution respectively. So, the computational complexity of
the proposed model can be greatly reduced.
For the quality regression module, we use the multilayer per-
ception (MLP) network to map the quality-aware features into
the chunk-level quality scores, and the temporal average pooling
strategy is adopted to obtain the final video quality. In order to
solve the problem of quality assessment across different resolu-
tions, we introduce a multi-scale quality fusion strategy to fuse
the quality scores of the videos with different resolutions, where
the multi-scale weights are obtained from the contrast sensitivity
function (CSF) of HVS by considering the viewing environment
information. The proposed models are validated on five popular
UGC VQA databases and the experimental results show that the
proposed model outperforms other state-of-the-art VQA models
by a large margin. What’s more, the proposed model trained on a
large-scale database such as LSVQ [ 44] achieves remarkable perfor-
mance when tested on the other databases without any fine-tuning,
which further demonstrates the effectiveness and generalizability
of the proposed model.
In summary, this paper makes the following contributions:
(1)We propose an effective and efficient deep learning based
model for UGC VQA, which includes the feature extraction
module, the quality regression module, and the quality pool-
ing module. The proposed model not only achieves remark-
able performance on the five popular UGC VQA databases
but also has a low computational complexity, which makes
it very suitable for practical applications.
(2)The feature extraction module extracts two kinds of quality-
aware features, the spatial features for spatial distortions and
the spatial-temporal features for motion distortions, where
the spatial features are learned from raw pixels of video
frames via an end-to-end manner and the spatial-temporal
features are extracted by a pretrained action recognition
network.
(3)We introduce a multi-scale quality fusion strategy to solve
the problem of quality assessment across different resolu-
tions, where the multi-scale weights are obtained from the
contrast sensitivity function of the human visual system by
considering the viewing environment information.
2 RELATED WORK
2.1 Handcrafted feature based NR VQA Models
A naive NR VQA method is to compute the quality of each frame
via popular NR IQA methods such as NIQE [ 24], BRISQUE [ 22],A Deep Learning based No-reference Quality Assessment Model for UGC Videos MM ’22, October 10–14, 2022, Lisboa, Portugal
Input video
Frame extraction
2D framesChunks
Global Average and STD Pooling
Spatial feature extractionStage 1Stage 2 Stage 3 Stage 4Motion feature extraction
Quality regressionQuality
ScoreMotion feature Spatial feature
Pooling3D CNN
Figure 1: The network architecture of the proposed model. The proposed model contains the feature extraction module, the
quality regression module, and the quality pooling module. The feature extraction module extracts two kinds of features, the
spatial features and the motion features.
CORNIA [ 42] etc., and then pool them into the video quality score.
A comparative study of various temporal pooling strategies on pop-
ular NR IQA methods can refer to [ 32]. The temporal information
is very important for VQA. V-BLIINDS [ 26] is a spatio-temporal
natural scene statistics (NSS) model for videos by quantifying the
NSS feature of frame-differences and motion coherency character-
istics. Mittal et al. [23] propose a training-free blind VQA model
named VIIDEO that exploits intrinsic statistics regularities of natu-
ral videos to quantify disturbances introduced due to distortions.
TLVQM [ 12] extracts abundant spatio-temporal features such as
motion, jerkiness, blurriness, noise, blockiness, color, etc. at two
levels of high and low complexity. VIDEVAL [ 33] further combines
the selected features from typical NR I/VQA methods to train a SVR
model to regress them into the video quality. Since video content
also affects its quality, especially for UGC videos, understanding the
video content is beneficial to NR VQA. Previous handcrafted feature
based methods are difficult to understand semantic information.
Hence, some studies [ 13,34] try to combine the handcrafted features
with the semantic-level features extracted by the pretrained CNN
model to improve the performance of NR VQA models. For example,
CNN-TLVQM [ 13] combines the handcrafted statistical temporal
features from TLVQM and spatial features extracted by 2D-CNN
model trained for IQA. RAPIQUE [ 34] utilizes the quality-aware
scene statistics features and semantics-aware deep CNN features
to achieve a rapid and accurate VQA model for UGC videos.
2.2 Deep learning based NR VQA Models
With the release of several large-scale VQA databases [ 8,36,44],
deep learning based NR VQA models [ 2,11,14,15,31,37,40,43,44]
attract many researchers’ attention. Liu et al. [17] propose a multi-
task BVQA model V-MEON by jointly optimizing the 3D-CNN
for quality assessment and compression distortion classification.
VSFA [ 15] first extracts the semantic features from a pre-trained
CNN model and then uses a gated recurrent unit (GRU) network
to model the temporal relationship between the semantic features
of video frames. The authors of VSFA further propose MDVSFA[16], which trains the VSFA model on the multiple VQA databases
to improve its performance and generalization. RIRNet [ 4] exploits
the effect of motion information extracted from the multi-scale
temporal frequencies for video quality assessment. Ying et al. [44]
propose a local-to-global region-based NR VQA model that com-
bines the spatial features extracted from a 2D-CNN model and the
spatial-temporal features from a 3D-CNN network. Wang et al. [37]
propose a feature-rich VQA model for UGC videos, which measures
the quality from three aspects, compression level, video content,
and distortion type and each aspect is evaluated by an individual
neural network. Xu et al. [40] first extract the spatial feature of
the video frame from a pre-trained IQA model and use the graph
convolution to extract and enhance these features, then extract
motion information from the optical flow domain, and finally inte-
grated the spatial feature and motion information via a bidirectional
long short-term memory network. Li et al. [14] also utilize the IQA
model pre-trianed on multiple databases to extract quality-aware
spatial features and the action recognition model to extract tem-
poral features, and then a GRU network is used to model spatial
and temporal features and regress them into the quality score. Wen
and Wang [ 39] propose a baseline I/VQA model for UGC videos,
which calculates the video quality by averaging the scores of each
frame and frame-level quality scores are obtained by a simple CNN
network.
3 PROPOSED MODEL
The framework of the proposed NR VQA model is shown in Fig-
ure 1, which consists of the feature extraction module, the quality
regression module, and the quality pooling module. First, we ex-
tract the quality-aware features from the spatial domain and the
spatial-temporal domain via the feature extraction module, which
are utilized to evaluate the spatial distortions and motion distor-
tions respectively. Then, the quality regression module is used to
map the quality-aware features into chunk-level quality scores. Fi-
nally, we perform the quality pooling module to obtain the video
quality score.MM ’22, October 10–14, 2022, Lisboa, Portugal Wei Sun, Xiongkuo Min, Wei Lu, and Guangtao Zhai∗
3.1 Feature Extraction Module
In this section, we expect to extract the quality-aware features that
can represent the impact of various distortion types and content
on visual quality. The types of video distortion can be roughly
divided into two categories: the spatial distortions and the motion
distortions. The spatial distortions refer to the artifacts introduced
in the video frames, such as noise, blur, compression, low visibility,
etc. The motion distortions refer to the jitter, lagging due, etc., which
are mainly caused by unstable shooting equipment, fast-moving
objects, the low network bandwidth, etc. Therefore, we need to
extract the quality-aware features from these two aspects.
Note that the characteristics of the spatial features and motion
features are quite different. The spatial features are sensitive to
the video resolution but insensitive to the video frame rate since
the adjacent frames of the video contain lots of redundancy spatial
information and higher resolution can represent more abundant
high-frequency information, while motion features are the opposite
because the motion distortions are reflected on the temporal dimen-
sion and these features are usually consistent for local regions of
the frames.
Therefore, considering these characteristics, given a video 𝑉,
whose number of frames and frame rate are 𝑙and𝑟respectively,
we first split the video 𝑉into𝑁𝑐continuous chunks 𝑐={𝑐𝑖}𝑁𝑐
𝑖=1at
an time interval 𝜏, where𝑁𝑐=𝑙/(𝑟∗𝜏), and there are 𝑁𝑓=𝑟∗𝜏
frames in each chunk 𝑐𝑖, which is denoted as 𝑐𝑖={𝑥𝑖,𝑗}𝑁𝑓
𝑗=1. Then
we only choose a key frame 𝑥𝑖,𝑘𝑒𝑦 in each chunk to extract the
spatial features and the motion features of each chunk are extracted
using all frames in 𝑐𝑖but at a very low spatial resolution. As a
result, we can greatly reduce the computation complexity of the
VQA model with little performance degradation.
3.1.1 Spatial Feature Extraction Module. Given a frame 𝑥, we de-
note𝑓𝑤(𝑥)as the output of the CNN model 𝑓with trainable pa-
rameters𝑤={𝑤𝑘}applied on the frame 𝑥. Assume that there are
𝑁𝑠stages in the CNN model, and 𝑓𝑘𝑤(𝑥)is the output feature maps
extracted from the 𝑘-th stage, where 𝑓𝑘𝑤(𝑥)∈R𝐻𝑘×𝑊𝑘×𝐶𝑘, and𝐻𝑘,
𝑊𝑘, and𝐶𝑘are the height, width, and the number of channels of
the feature maps 𝑓𝑘𝑤(𝑥)respectively. In the following, we use the
𝑓𝑘𝑤to replace the 𝑓𝑘𝑤(𝑥)for simplicity.
It is well known that the features extracted by the deep lay-
ers of the CNN model contain rich semantic information, and are
suitable for representing content-aware features for UGC VQA.
Moreover, previous studies indicate that the features extracted by
the shallow layers of the CNN models contain low-level informa-
tion [ 28,48], which responds to low-level features such as edges,
corners, textures, etc. The low-level information is easily affected
by the distortion and is therefore distortion-aware. Hence, we ex-
tract the quality-aware features via calculating the global mean
and stand deviation of feature maps extracted from all stages of the
CNN model. Then, we apply global average and stand deviation
pooling operations on the feature maps 𝑓𝑘𝑤:
𝜇𝑓𝑘𝑤=GPavg(𝑓𝑘
𝑤),
𝜎𝑓𝑘𝑤=GPstd(𝑓𝑘
𝑤),(1)where𝜇𝑓𝑘𝑤and𝜎𝑓𝑘𝑤are the global means and stand deviation of
feature maps 𝑓𝑘𝑤respectively. Finally, we concatenate the 𝜇𝑓𝑘𝑤and
𝜎𝑓𝑘𝑤to derive the spital feature representation of our NR VQA model:
𝐹𝑘
𝑠=cat([𝜇𝑓𝑘𝑤,𝜎𝑓𝑘𝑤]),
𝐹𝑠=cat({𝐹𝑘
𝑠}𝑁𝑠
𝑘=1).(2)
3.1.2 Motion Feature Extraction Module. We extract the motion
features as the complementary quality-aware features since UGC
videos are commonly degraded by the motion distortions caused
by the unstable shooting equipment or low bit rates in the living
streaming or videoconferencing. The spatial features are difficult
to handle these distortions because they are extracted by the intra-
frames while motion distortions occur in the interframes. Therefore,
the motion features are also necessary for evaluating the quality
of UGC videos. Here, we utilize the pretrained action recognition
model as the motion feature extractor to obtain the motion features
of each video chunk. The action recognition model is designed
to detect different kinds of action classes, so the feature represen-
tation of the action recognition network can reflect the motion
information of the video to a certain extent. Therefore, given the
video chunk 𝑐and the action recognition network MOTION , we
can obtain the motion features:
𝐹𝑚=MOTION(c) (3)
where𝐹𝑚represents the motion features extract by the action
recognition network.
Therefore, given the video chunk 𝑐, we first select a key frame
in the chunk to calculate the spatial features 𝐹𝑠. Then, we calculate
the motion features 𝐹𝑚using the whole frames but at a low spatial
resolution in the video chunk. Finally, we obtain the quality-aware
features for the video chunk 𝑐by concatenating the spatial features
and motion features:
𝐹=cat([𝐹𝑠,𝐹𝑚]), (4)
3.2 Quality Regression Module
After extracting quality-aware feature representation by the feature
extraction module, we need to map these features to the quality
scores via a regression model. In this paper, we use the multi-layer
perception (MLP) as the regression model to obtain the chunk-level
quality due to its simplicity and effectiveness. The MLP consists of
two fully connected layers and there are 128 and 1 neuron in each
layer respectively. Therefore, we can obtain the chunk-level quality
score via
𝑞=𝑓𝑤FC(𝐹), (5)
where𝑓𝑤FCdenotes the function of the two FC layers and 𝑞is the
quality of the video chunk.
3.3 Quality Pooling Module
As stated in Section 3.1, we split the video 𝑉into𝑁𝑐continuous
chunks{𝑐𝑖}𝑁𝑐
𝑖=1. For the chunk 𝑐𝑖, we can obtain its chunk-level
quality score 𝑞𝑖via the feature extraction module and the quality
regression module. Then, it is necessary to pool the chunk-level
scores into the video level. Though many temporal pooling methods
have been proposed in literature [ 32][15], we find that the temporalA Deep Learning based No-reference Quality Assessment Model for UGC Videos MM ’22, October 10–14, 2022, Lisboa, Portugal
Table 1: Summary of the benchmark UGC VQA databases. Time duration: Seconds.
Database Videos Scenes Resolution Time Duration Format Distortion Type DATA Environment
KoNViD-1k [8] 1,200 1,200 540p 8 MP4 Authentic MOS + 𝜎 Crowd
YouTube-UGC [36] 1500 1500 360p-4K 20 YUV, MP4 Authentic MOS + 𝜎 Crowd
LSVQ [44] 38,811 38,811 99p-4K 5-12 MP4 Authentic MOS + 𝜎 Crowd
LBVD [3] 1,013 1,013 240p-540p 10 MP4 Authentic, Transmission MOS + 𝜎 In-lab
LIVE-YT-Gaming [45] 600 600 360p-1080p 8-9 MP4 Authentic MOS Crowd
averaging pooling achieves the best performance from Section 4.3.2.
Therefore, the video-level quality is calculated as:
𝑄=1
𝑁𝑐𝑁𝑐∑︁
𝑖=1𝑞𝑖, (6)
where𝑞𝑖is the quality of the 𝑖-th chunk and 𝑄is the video quality
evaluated by the proposed model.
3.4 Loss Function
The loss function used to optimize the proposed models consists of
two parts: the mean absolute error (MAE) loss and rank loss [ 39].
The MAE loss is used to make the evaluated quality scores close to
the ground truth, which is defined as:
𝐿𝑀𝐴𝐸=1
𝑁𝑁∑︁
𝑖=1𝑄𝑖−ˆ𝑄𝑖, (7)
where the ˆ𝑄𝑖is the ground truth quality score of the 𝑖-th video in a
mini-batch and 𝑁is the number of videos in the mini-batch.
The rank loss is further introduced to make the model distinguish
the relative quality of videos better, which is very useful for the
model to evaluate the videos with similar quality. Since the rank
value between two video quality is non-differentiable, we use the
following formula to approximate the rank value:
𝐿𝑖𝑗
𝑟𝑎𝑛𝑘=max(0,ˆ𝑄𝑖−ˆ𝑄𝑗−𝑒(ˆ𝑄𝑖,ˆ𝑄𝑗)·(𝑄𝑖−𝑄𝑗)), (8)
where𝑖and𝑗are two video indexes in a mini-batch, and 𝑒(ˆ𝑄𝑖,ˆ𝑄𝑗)
is formulated as:
𝑒(ˆ𝑄𝑖,ˆ𝑄𝑗)=(
1,ˆ𝑄𝑖≥ˆ𝑄𝑗,
−1,ˆ𝑄𝑖<ˆ𝑄𝑗,(9)
Then,𝐿𝑟𝑎𝑛𝑘 is calculated via:
𝐿𝑟𝑎𝑛𝑘=1
𝑁2𝑁∑︁
𝑖=1𝑁∑︁
𝑗=1𝐿𝑖𝑗
𝑟𝑎𝑛𝑘(10)
Finally, the loss function can be obtained by:
𝐿=𝐿𝑀𝐴𝐸+𝜆·𝐿𝑟𝑎𝑛𝑘, (11)
where𝜆is a hyper-parameter to balance the MAE loss and the rank
loss.
3.5 Multi-scale Quality Fusion Strategy
Previous studies evaluate the video quality either using the original
spatial resolution or a fixed resized spatial resolution, which ig-
nore that videos are naturally multi-scale [ 53]. Some existing work
[38][25][20] shows that considering the multi-scale characteristics
can improve the performance of image quality assessment. So, wepropose a multi-scale quality fusion strategy to further improve
the evaluation accuracy of the VQA model and this strategy is
very useful to compare the quality of videos with different spatial
resolutions.
3.5.1 Multi-scale Video Quality Scores. We first resize the resolu-
tion of the video into three fixed spatial scales, which are 540p,
720p, and 1080p, respectively. We do not downscale the video from
the original scale to several lower resolution scales, which is a
more common practice in previous studies. That is because when
users watch videos in an application, the resolution of videos is
actually adapted to the resolution of the playback device, and the
modern display resolution is normally larger than 1080p. So, the
perceptual quality of the low-resolution videos is also affected by
the up-sampling artifacts, which also need to be considered by VQA
models. Therefore, given a VQA model, we can derive three qual-
ity of videos at three scales, which are denoted as 𝑄1,𝑄2, and𝑄3
respectively.
3.5.2 Adaptive Multi-scale Weights. The weight of each scale is
obtained by considering the human psychological behaviors and
the visual sensitivity characteristics. It is noted that the contrast
perception ability of the HVS depends on the spatial frequency
of the visual signal, which is modeled by the contrast sensitivity
function (CSF). Specifically, we first define a viewing resolution
factor𝜉as:
𝜉=𝜋·𝑑·𝑛
180·ℎ𝑠·2, (12)
where the unit of 𝜉is cycles per degree of visual angle (cpd), 𝑑is
the viewing distance (inch), ℎ𝑠is the height of the screen (inch),
and𝑛denotes the number of pixels in the vertical direction of the
screen. For the above three spatial scales of video, we can obtain the
corresponding 𝜉, which are denoted as 𝜉1,𝜉2, and𝜉3respectively.
We use𝜉to divide the spatial frequency range of the corresponding
scale, which covers one section of the CSF formulated by:
𝑆(𝑢)=5200𝑒(−0.0016𝑢2(1+100/𝐿)0.08)
√︃
(1+144
𝑋2
0+0.64𝑢2)(63
𝐿0.83+1
1−𝑒(−0.02𝑢2))(13)
where𝑢,𝐿, and𝑋2
0indicate spatial frequency (cpd), luminance
(cd/m2), and angular object area (squared degrees), respectively.
The weight of each scale is calculated as the area under the CSF
within the corresponding frequency covering range:
𝑤𝑖=1
𝑍∫𝜉𝑖
𝜉𝑖−1𝑆(𝑢)d𝑢,𝑖∈{1,2,3}, (14)
where𝑖from 1 to 3 corresponds the finest to coarsest scale respec-
tively, and𝜉0corresponds the viewing resolution factor of 0. 𝑍is a
normalization factor such thatÍ
𝑖𝑤𝑖=1.MM ’22, October 10–14, 2022, Lisboa, Portugal Wei Sun, Xiongkuo Min, Wei Lu, and Guangtao Zhai∗
Table 2: Performance of the SOTA models and the proposed model on the KoNViD-1k, YouTube-UGC, LBVD, and LIVE-YT-
Gaming databases. W.A. means the weight average results. The best performing model is highlighted in each column.
TypeDatabase KoNViD-1k YouTube-UGC LBVD LIVE-YT-Gaming W.A.
Criterion SRCC PLCC SRCC PLCC SRCC PLCC SRCC PLCC SRCC PLCC
IQANIQE 0.542 0.553 0.238 0.278 0.327 0.387 0.280 0.304 0.359 0.393
BRISQUE 0.657 0.658 0.382 0.395 0.435 0.446 0.604 0.638 0.513 0.525
GM-LOG 0.658 0.664 0.368 0.392 0.314 0.304 0.312 0.317 0.433 0.440
VGG19 0.774 0.785 0.703 0.700 0.676 0.673 0.678 0.658 0.714 0.712
ResNet50 0.802 0.810 0.718 0.710 0.715 0.717 0.729 0.768 0.744 0.751
KonCept512 0.735 0.749 0.587 0.594 0.626 0.636 0.643 0.649 0.650 0.660
VQAV-BLIINDS 0.710 0.704 0.559 0.555 0.527 0.558 0.357 0.403 0.566 0.578
TLVQM 0.773 0.769 0.669 0.659 0.614 0.590 0.748 0.756 0.699 0.689
VIDEVAL 0.783 0.780 0.779 0.773 0.707 0.697 0.807 0.812 0.766 0.762
RAPIQUE 0.803 0.818 0.759 0.768 0.712 0.725 0.803 0.825 0.767 0.781
VSFA 0.773 0.775 0.724 0.743 0.622 0.642 0.776 0.801 0.721 0.736
Liel al. 0.836 0.834 0.831 0.819 - - - - - -
Pro. 0.856 0.860 0.847 0.856 0.844 0.846 0.861 0.866 0.851 0.856
Table 3: Performance of the SOTA models and the proposed
models on the LSVQ database. Pro. M.S. refers to the pro-
posed model implemented by the multi-scale quality fusion
strategy. W.A. means the weighted average results. The best
performing model is highlighted in each column.
Database Test Test-1080p W.A.
Criterion SRCC PLCC SRCC PLCC SRCC PLCC
TLVQM 0.772 0.774 0.589 0.616 0.712 0.722
VIDEVAL 0.794 0.783 0.545 0.554 0.712 0.707
VSFA 0.801 0.796 0.675 0.704 0.759 0.766
PVQ 0.827 0.828 0.711 0.739 0.789 0.799
Liel al. 0.852 0.854 0.772 0.788 0.825 0.832
Pro. 0.864 0.861 0.756 0.801 0.829 0.841
Pro. M.S. 0.867 0.861 0.764 0.803 0.833 0.842
Therefore, the multi-scale fusion quality score 𝑄𝑚is calculated
as:
𝑄𝑚=3Ö
𝑖=1𝑄𝑤𝑖
𝑖, (15)
4 EXPERIMENTAL VALIDATION
4.1 Experimental Protocol
4.1.1 Test Databases. We test the proposed model on the five UGC
VQA database: KoNViD-1k [ 8], YouTube-UGC [ 36], LSVQ [ 44],
LBVD [ 3], and LIVE-YT-Gaming [ 45]. We summarize the main infor-
mation of the databases in Table 1. The LSVQ database is the largest
UGC VQA database so far, and there are 15 video categories such
as animation, gaming, HDR, live music, sports, etc. in the YouTube-
UGC database, which is more diverse than other databases. The
LBVD database focuses on the live broadcasting videos, of which
the videos are degraded by the authentic transmission distortions.
The LIVE-YT-Gaming database consists of streamed gaming videos,
where the video content is generated by computer graphics.4.1.2 Implementation Details. We use the ResNet50 [ 7] as the back-
bone of the spatial feature extraction module and the SlowFast R50
[6] as the motion feature extraction model for the whole experi-
ments. The weights of the ResNet50 are initialized by training on
the ImageNet dataset [ 5], the weights of the SlowFast R50 are fixed
by training on the Kinetics 400 dataset [ 10], and other weights are
randomly initialized. For the spatial feature extraction module, we
resize the resolution of the minimum dimension of key frames as
520 while maintaining their aspect ratios. In the training stage, the
input frames are randomly cropped with the resolution of 448 ×448.
If we do not use the multi-scale quality fusion strategy, we crop the
center patch with the same resolutions of 448 ×448 in the testing
stage. Note that we only validate the multi-scale quality fusion
strategy on the model trained by the LSVQ database since there
are enough videos with various spatial resolutions in it. For the
motion feature extraction module, the resolution of the videos is
resized to 224×224 for both the training and testing stages. We use
PyTorch to implement the proposed models. The Adam optimizer
with the initial learning rate 0.00001 and batch size 8 are used for
training the proposed model on a server with NVIDIA V100. The
hyper-parameter 𝜆is set as 1. For simplicity, we select the first
frame in each chunk as the key frame. For the multi-scale quality
fusion strategy, there are 𝑑=35,𝑛=1080 ,ℎ=11.3,𝐿=200, and
𝑋2
0=606, and the final multi-scale weights for UGC videos are
𝑤1=0.8317,𝑤2=0.0939, and𝑤3=0.0745.
4.1.3 Comparing Algorithms. We compare the proposed method
with the following no-reference models:
•IQA models: NIQE [ 24], BRISQUE [ 22], GM-LOG [ 41], VGG19
[27], ResNet50 [7], and KonCept512 [9].
•VQA models: V-BLIINDS [ 26], TLVQM [ 12], VIDEAL [ 33],
RAPIQUE [34], VSFA [15], PVQ [44], and Li et al. [14].
Since the number of videos in the LSVQ database is relatively
large, we only compare some representative VQA models on the
LSVQ database and omit the methods which perform poorly on the
other four UGC databases.A Deep Learning based No-reference Quality Assessment Model for UGC Videos MM ’22, October 10–14, 2022, Lisboa, Portugal
4.1.4 Evaluation Criteria. We adopt two criteria to evaluate the
performance of VQA models, which are Pearson linear correlation
coefficient (PLCC) and Spearman rank-order correlation coefficient
(SRCC). PLCC reflects the prediction linearity of the VQA algorithm
and SRCC indicates the prediction monotonicity. An excellent VQA
model should obtain the value of SRCC and PLCC close to 1. Before
calculating the PLCC, we follow the same procedure in [ 1] to map
the objective score to the subject score using a four-parameter
logistic function.
For KoNViD-1k, YouTube-UGC, LBVD, and LIVE-YT-Gaming
databases, we randomly split these databases into the training set
with 80% videos and the test set with 20% videos for 10 times, and
report the median values of SRCC and PLCC. For the LSVQ database,
we follow the same training and test split suggested by [ 44] and
report the performance on the test and test-1080p subsets.
4.2 Performance Comparison with the SOTA
Models
The performance results of the VQA models on the KoNViD-1k,
YouTube-UGC, LBVD, and LIVE-YT-Gaming databases are listed in
Table 2, and on the LSVQ database are listed in Table 3. From Table
2 and Table 3, we observe that the proposed model achieves the best
performance on all five UGC VQA databases and leads by a large
margin, which demonstrates that the proposed model does have a
strong ability to measure the perceptual quality of various kinds of
UGC videos. For the test-1080p subset of the LSVQ database, the
proposed model is inferior to Li et al., which may be because the spa-
tial resolution of most videos in the test-1080p subset is larger than
1080p while the proposed model resizes the spatial resolution of test
videos into 448×448, so the proposed model has a relatively poor
ability to represent the characteristics of high-resolution videos.
Through the multi-scale quality weighting fusion strategy, the pro-
posed model can significantly improve the performance on the
test-1080p subset.
Then, most handcrafted feature based IQA models perform poorly
on these UGC VQA databases especially for the LBVD and LIVE-
YT-Gaming databases since they are designed for natural scene
images with synthetic distortions and are difficult to handle the
complex in-the-wild distortions and other video types such gaming,
videoliving, etc. It is worth noting that through fine-tuning the deep
CNN baseline i.e. ResNet50 on the VQA databases, it can achieve a
pretty good performance, which also indicates that spatial features
are very important for VQA tasks. For the NR VQA methods, the
hand-crafted feature based NR VQA methods such as TLVQM and
VIDEVAL achieve pretty well performance by incorporating the
rich spatial and temporal quality features, such as NSS features,
motion features, etc., but they are inferior to the deep learning
based NR VQA methods due to the strong feature representation
ability of CNN. VSFA extracts the spatial features from the pre-
trained image recognition model, which are not quality-aware, and
achieves relatively poor performance when compared with other
deep learning based methods. PVQ and Li et al. methods both uti-
lize the pretrained IQA model and ptretrained action recognition
model to extract spatial and motion features respectively, and they
perform better than other compared NR I/VQA methods but are
inferior to the proposed model. Through training an end-to-endTable 4: The results of ablation studies on the LSVQ data-
base. S and M means the spatial features and motion features
respectively, and S∗means that the spatial features are ex-
tracted by the pretrained image classification network.
Database Test Test-1080p
Criterion SRCC PLCC SRCC PLCC
FeatureS∗+M 0.847 0.841 0.732 0.774
S 0.827 0.829 0.702 0.757
M 0.660 0.669 0.569 0.621
RegressionGRU 0.858 0.855 0.735 0.788
Transformer 0.860 0.861 0.753 0.799
PoolingMethod in [15] 0.860 0.858 0.733 0.786
1D CNN based 0.864 0.862 0.739 0.790
Table 5: The SRCC results of cross-database evaluation. The
model is trained on the LSVQ database.
Database KoNViD-1k YouTube-UGC LBVD LIVE-YT-Gaming
Pro. 0.860 0.789 0.689 0.642
Pro. M.S. 0.859 0.822 0.711 0.683
spatial feature extractor, the proposed model can take advantage of
various video content and distortion types in the UGC databases
and learn a better quality-aware feature representation. As a result,
the proposed model achieves the best performance on all five UGC
VQA databases.
4.3 Ablation Studies
In this section, we conduct several ablation studies to investigate
the effectiveness of each module in the proposed model, including
the feature extraction module, and the quality regression module.
All the experiments are tested on the LSVQ database since it is the
largest UGC VQA model and is more representative.
4.3.1 Feature Extraction Module. The proposed model consists
of the spatial feature extractor that learns the end-to-end spatial
quality-aware features and the motion feature extractor that utilizes
a pretrained action recognition model to represent motion informa-
tion. Therefore, we first do not train the spatial feature extractor
and directly use the weights trained on the ImageNet database to
study the effect of the end-to-end training strategy for the spatial
feature extractor. Then, we only use the end-to-end trained spatial
features or the pretrained motion features to evaluate the quality of
UGC videos to investigate the effect of these two kinds of features.
The results are listed in Table 4. First, it is observed that the model
using the motion features is inferior to the model using the spatial
features and both of them are inferior to the proposed model, which
indicates that both spatial and motion features are beneficial to the
UGC VQA task and the spatial features are more important. Then,
we find that end-to-end training for the spatial feature extractor can
significantly improve the evaluation performance, which demon-
strates that end-to-end trained spatial features represent better than
that extracted by the pretrained image classification model.MM ’22, October 10–14, 2022, Lisboa, Portugal Wei Sun, Xiongkuo Min, Wei Lu, and Guangtao Zhai∗
Table 6: Comparison of computational complexity for the six VQA models and two proposed models. Time: Second.
Methods V-BLIINDS TLVQM VIDEVAL VSFA RAPIQUE Li et al. Pro. Pro. M.S.
Time 61.982 219.992 561.408 56.424 38.126 61.971 6.929 8.448
4.3.2 Quality Regression Module. In this paper, we use the MLP
as the regression model to derive the chunk-level quality scores.
However, in previous studies, some sequential models such as GRU
[15], Transformer [ 14], etc. are also adopted to further consider the
influence of the features extracted from adjacent frames. Here, we
also adopt these methods as a comparison to investigate whether
sequential models can improve the performance of the proposed
models. Specifically, we replace the MLP module with the GRU and
Transformer and keep other experimental setups the same. The
results are listed in Table 4. We observe that models using GRU
and Transformer are both inferior to the proposed model, which
means that the MLP module is enough to regress the quality-aware
features to quality scores though it is very simple. This conclusion
is also consistent with [ 37]. The reason is that the proposed model
and the model in [ 37] calculate the chunk-level quality score and
the effect of adjacent frames are considered in the quality-aware
features (i.e. motion features), while other VQA models [ 15] [14]
calculate the frame-level quality scores, which may need to consider
the effect of adjacent frames in the quality regression module.
4.3.3 Quality Pooling Module. The proposed model uses the tem-
poral average pooling method to fuse the chunk-level quality scores
into the video level. It is noted that previous studies also propose
several temporal pooling methods for VQA. In this section, we test
two temporal pooling methods, which are the subjectively-inspired
method introduced in [ 15] and a learning based temporal pooling
method using the 1D CNN. The results are listed in Table 4. From
Table 4, we observe that the average pooling strategy achieves sim-
ilar performance to the learning based pooling method, and both of
them are superior to the subjectively-inspired methods. Since the
average pooling strategy is simpler and does not increase the extra
parameters, we use the temporal average pooling method in this
paper.
4.4 Cross-Database Evaluation
UGC videos may contain various kinds of distortions and content,
most of which may not exist in the training set. Hence, the gen-
eralization ability of the UGC VQA model is very important. In
this section, we use the cross-database evaluation to test the gener-
alization ability of the proposed model. Specifically, we train the
proposed model on the LSVQ database and test the trained model
on the other four UGC VQA databases. We list the results in Table
5. It is observed that the proposed model achieves excellent per-
formance in cross-database evaluation. The SRCC results on the
KoNViD-1k and YouTube-UGC databases both exceed 0.8, which
have surpassed most VQA models trained on the corresponding
database. We find that the multi-scale quality fusion strategy can
significantly improve the performance on the databases containing
videos with different spatial resolutions (YouTube-UGC, LBVD, and
LIVE-YT-Gaming), which further demonstrates its effectiveness.It is also observed that the performance on the LBVD and LIVE-
YT-Gaming databases is not good as the other two databases. The
reason is that the LBVD and LIVE-YT-Gaming databases contain
live broadcasting and gaming videos respectively, which may rarely
exist in the LSVQ database. Since the single database can not cover
all kinds of video types and distortions, we may further improve the
generalization ability of the proposed model via the multiple data-
base training strategy [ 30] [51] or the continual learning manner
[49] [50].
4.5 Computational Complexity
The computational complexity is a very important factor that needs
to be considered in practical applications. Hence, we test the com-
putational complexity in this section. All models are tested on a
computer with i7-6920HQ CPU, 16G RAM, and NVIDIA Quadro
P400. The deep learning based models and the handcrafted based
models are tested using the GPU and CPU respectively. We report
the running time for a video with the resolution of 1920 ×1080 and
time duration of eight seconds in Table 6. It is seen that the proposed
model has a considerably low running time compared with other
VQA models. The reason is that we use very sparse frames to calcu-
late the spatial features while other deep learning based methods
need dense frames. Moreover, we extract the motion features at a
very low resolution, which only adds little computational complex-
ity to the proposed model. The very low computational complexity
makes the proposed model suitable for practical applications.
5 CONCLUSION
In this paper, we propose an effective and efficient NR VQA model
for UGC videos. The proposed model extracts the quality-aware
features from the spatial domain and the spatial-temporal domain to
measure the spatial distortions and motion distortions respectively.
We train the spatial feature extractor in an end-to-end training
manner, so the proposed model can make full use of the various
spatial distortions and content in the current VQA database. Then,
the quality-aware features are regressed into the quality scores
by the MLP network, and the temporal average pooling is used
to obtain the video-level quality scores. We further introduce the
multi-scale quality fusion strategy to address the problem of quality
assessment across different spatial resolutions. The experimental
results show that the proposed model can effectively measure the
quality of UGC videos.
ACKNOWLEDGMENTS
This work was supported by the National Natural Science Foun-
dation of China (61831015, 61901260) and the National Key R&D
Program of China 2021YFE0206700.A Deep Learning based No-reference Quality Assessment Model for UGC Videos MM ’22, October 10–14, 2022, Lisboa, Portugal
REFERENCES
[1]Jochen Antkowiak, T Jamal Baina, France Vittorio Baroncini, Noel Chateau,
France FranceTelecom, Antonio Claudio França Pessoa, F Stephanie Colonnese,
Italy Laura Contin, Jorge Caviedes, and France Philips. 2000. Final report from
the video quality experts group on the validation of objective models of video
quality assessment march 2000. (2000).
[2]Yuqin Cao, Xiongkuo Min, Wei Sun, and Guangtao Zhai. 2021. Deep Neural Net-
works For Full-Reference And No-Reference Audio-Visual Quality Assessment.
In2021 IEEE International Conference on Image Processing (ICIP) . IEEE, 1429–1433.
[3]Pengfei Chen, Leida Li, Yipo Huang, Fengfeng Tan, and Wenjun Chen. 2019. QoE
evaluation for live broadcasting video. In 2019 IEEE International Conference on
Image Processing (ICIP) . IEEE, 454–458.
[4]Pengfei Chen, Leida Li, Lei Ma, Jinjian Wu, and Guangming Shi. 2020. RIRNet:
Recurrent-in-recurrent network for video quality assessment. In Proceedings of
the 28th ACM International Conference on Multimedia . 834–842.
[5]Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet:
A large-scale hierarchical image database. In 2009 IEEE conference on computer
vision and pattern recognition . Ieee, 248–255.
[6]Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. 2019. Slow-
fast networks for video recognition. In Proceedings of the IEEE/CVF international
conference on computer vision . 6202–6211.
[7]Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual
learning for image recognition. In Proceedings of the IEEE conference on computer
vision and pattern recognition . 770–778.
[8]Vlad Hosu, Franz Hahn, Mohsen Jenadeleh, Hanhe Lin, Hui Men, Tamás Szirányi,
Shujun Li, and Dietmar Saupe. 2017. The Konstanz natural video database
(KoNViD-1k). In 2017 Ninth international conference on quality of multimedia
experience (QoMEX) . IEEE, 1–6.
[9]Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. 2020. KonIQ-10k: An
ecologically valid database for deep learning of blind image quality assessment.
IEEE Transactions on Image Processing 29 (2020), 4041–4056.
[10] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra
Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al .2017.
The kinetics human action video dataset. arXiv preprint arXiv:1705.06950 (2017).
[11] Woojae Kim, Jongyoo Kim, Sewoong Ahn, Jinwoo Kim, and Sanghoon Lee. 2018.
Deep video quality assessor: From spatio-temporal visual sensitivity to a convo-
lutional neural aggregation network. In Proceedings of the European Conference
on Computer Vision (ECCV) . 219–234.
[12] Jari Korhonen. 2019. Two-level approach for no-reference consumer video quality
assessment. IEEE Transactions on Image Processing 28, 12 (2019), 5923–5938.
[13] Jari Korhonen, Yicheng Su, and Junyong You. 2020. Blind natural video quality
prediction via statistical temporal features and deep spatial features. In Proceed-
ings of the 28th ACM International Conference on Multimedia . 3311–3319.
[14] Bowen Li, Weixia Zhang, Meng Tian, Guangtao Zhai, and Xianpei Wang. 2021.
Blindly Assess Quality of In-the-Wild Videos via Quality-aware Pre-training and
Motion Perception. arXiv preprint arXiv:2108.08505 (2021).
[15] Dingquan Li, Tingting Jiang, and Ming Jiang. 2019. Quality assessment of in-the-
wild videos. In Proceedings of the 27th ACM International Conference on Multimedia .
2351–2359.
[16] Dingquan Li, Tingting Jiang, and Ming Jiang. 2021. Unified quality assessment
of in-the-wild videos with mixed datasets training. International Journal of
Computer Vision 129, 4 (2021), 1238–1257.
[17] Wentao Liu, Zhengfang Duanmu, and Zhou Wang. 2018. End-to-End Blind
Quality Assessment of Compressed Videos Using Deep Neural Networks.. In
ACM Multimedia . 546–554.
[18] Wei Lu, Wei Sun, Xiongkuo Min, Wenhan Zhu, Quan Zhou, Jun He, Qiyuan Wang,
Zicheng Zhang, Tao Wang, and Guangtao Zhai. 2022. Deep Neural Network for
Blind Visual Quality Assessment of 4K Content. arXiv preprint arXiv:2206.04363
(2022).
[19] Pavan C Madhusudana, Xiangxu Yu, Neil Birkbeck, Yilin Wang, Balu Adsumilli,
and Alan C Bovik. 2021. Subjective and objective quality assessment of high
frame rate videos. IEEE Access 9 (2021), 108069–108082.
[20] Xiongkuo Min, Kede Ma, Ke Gu, Guangtao Zhai, Zhou Wang, and Weisi Lin.
2017. Unified blind quality assessment of compressed natural, graphic, and screen
content images. IEEE Transactions on Image Processing 26, 11 (2017), 5462–5474.
[21] Xiongkuo Min, Guangtao Zhai, Jiantao Zhou, Mylene CQ Farias, and Alan Conrad
Bovik. 2020. Study of subjective and objective quality assessment of audio-visual
signals. IEEE Transactions on Image Processing 29 (2020), 6054–6068.
[22] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. 2012. No-
reference image quality assessment in the spatial domain. IEEE Transactions on
image processing 21, 12 (2012), 4695–4708.
[23] Anish Mittal, Michele A Saad, and Alan C Bovik. 2015. A completely blind video
integrity oracle. IEEE Transactions on Image Processing 25, 1 (2015), 289–300.
[24] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. 2012. Making a “completely
blind” image quality analyzer. IEEE Signal processing letters 20, 3 (2012), 209–212.
[25] Abdul Rehman, Kai Zeng, and Zhou Wang. 2015. Display device-adapted video
quality-of-experience assessment. In Human Vision and Electronic Imaging XX ,Vol. 9394. International Society for Optics and Photonics, 939406.
[26] Michele A Saad, Alan C Bovik, and Christophe Charrier. 2014. Blind prediction
of natural video quality. IEEE Transactions on Image Processing 23, 3 (2014),
1352–1365.
[27] Karen Simonyan and Andrew Zisserman. 2014. Very deep convolutional networks
for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014).
[28] Wei Sun, Xiongkuo Min, Guangtao Zhai, Ke Gu, Huiyu Duan, and Siwei Ma. 2019.
MC360IQA: a multi-channel CNN for blind 360-degree image quality assessment.
IEEE Journal of Selected Topics in Signal Processing 14, 1 (2019), 64–77.
[29] Wei Sun, Xiongkuo Min, Guangtao Zhai, Ke Gu, Siwei Ma, and Xiaokang Yang.
2020. Dynamic backlight scaling considering ambient luminance for mobile
videos on lcd displays. IEEE Transactions on Mobile Computing (2020).
[30] Wei Sun, Xiongkuo Min, Guangtao Zhai, and Siwei Ma. 2021. Blind quality
assessment for in-the-wild images via hierarchical feature fusion and iterative
mixed database training. arXiv preprint arXiv:2105.14550 (2021).
[31] Wei Sun, Tao Wang, Xiongkuo Min, Fuwang Yi, and Guangtao Zhai. 2021. Deep
learning based full-reference and no-reference quality assessment models for
compressed ugc videos. In 2021 IEEE International Conference on Multimedia &
Expo Workshops (ICMEW) . IEEE, 1–6.
[32] Zhengzhong Tu, Chia-Ju Chen, Li-Heng Chen, Neil Birkbeck, Balu Adsumilli,
and Alan C Bovik. 2020. A comparative evaluation of temporal pooling methods
for blind video quality assessment. In 2020 IEEE International Conference on Image
Processing (ICIP) . IEEE, 141–145.
[33] Zhengzhong Tu, Yilin Wang, Neil Birkbeck, Balu Adsumilli, and Alan C Bovik.
2021. UGC-VQA: Benchmarking blind video quality assessment for user generated
content. IEEE Transactions on Image Processing (2021).
[34] Zhengzhong Tu, Xiangxu Yu, Yilin Wang, Neil Birkbeck, Balu Adsumilli, and
Alan C Bovik. 2021. Rapique: Rapid and accurate video quality prediction of user
generated content. arXiv preprint arXiv:2101.10955 (2021).
[35] Tao Wang, Zicheng Zhang, Wei Sun, Xiongkuo Min, Wei Lu, and Guangtao
Zhai. 2022. Subjective Quality Assessment for Images Generated by Computer
Graphics. arXiv preprint arXiv:2206.05008 (2022).
[36] Yilin Wang, Sasi Inguva, and Balu Adsumilli. 2019. YouTube UGC dataset for video
compression research. In 2019 IEEE 21st International Workshop on Multimedia
Signal Processing (MMSP) . IEEE, 1–5.
[37] Yilin Wang, Junjie Ke, Hossein Talebi, Joong Gon Yim, Neil Birkbeck, Balu
Adsumilli, Peyman Milanfar, and Feng Yang. 2021. Rich features for percep-
tual quality assessment of UGC videos. In Proceedings of the IEEE/CVF Conference
on Computer Vision and Pattern Recognition . 13435–13444.
[38] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. 2003. Multiscale structural sim-
ilarity for image quality assessment. In The Thrity-Seventh Asilomar Conference
on Signals, Systems & Computers, 2003 , Vol. 2. Ieee, 1398–1402.
[39] Shaoguo Wen and Junle Wang. 2021. A strong baseline for image and video
quality assessment. arXiv preprint arXiv:2111.07104 (2021).
[40] Jiahua Xu, Jing Li, Xingguang Zhou, Wei Zhou, Baichao Wang, and Zhibo Chen.
2021. Perceptual Quality Assessment of Internet Videos. In Proceedings of the
29th ACM International Conference on Multimedia . 1248–1257.
[41] Wufeng Xue, Xuanqin Mou, Lei Zhang, Alan C Bovik, and Xiangchu Feng. 2014.
Blind image quality assessment using joint statistics of gradient magnitude and
Laplacian features. IEEE Transactions on Image Processing 23, 11 (2014), 4850–
4862.
[42] Peng Ye, Jayant Kumar, Le Kang, and David Doermann. 2012. Unsupervised
feature learning framework for no-reference image quality assessment. In 2012
IEEE conference on computer vision and pattern recognition . IEEE, 1098–1105.
[43] Fuwang Yi, Mianyi Chen, Wei Sun, Xiongkuo Min, Yuan Tian, and Guangtao
Zhai. 2021. Attention Based Network For No-Reference UGC Video Quality
Assessment. In 2021 IEEE International Conference on Image Processing (ICIP) .
IEEE, 1414–1418.
[44] Zhenqiang Ying, Maniratnam Mandal, Deepti Ghadiyaram, and Alan Bovik. 2021.
Patch-VQ:’Patching Up’the Video Quality Problem. In Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern Recognition . 14019–14029.
[45] Xiangxu Yu, Zhenqiang Ying, Neil Birkbeck, Yilin Wang, Balu Adsumilli, and
Alan C Bovik. 2022. Subjective and Objective Analysis of Streamed Gaming
Videos. arXiv preprint arXiv:2203.12824 (2022).
[46] Saman Zadtootaghaj, Nabajeet Barman, Steven Schmidt, Maria G Martini, and
Sebastian Möller. 2018. NR-GVQM: A no reference gaming video quality metric.
In2018 IEEE International Symposium on Multimedia (ISM) . IEEE, 131–134.
[47] Saman Zadtootaghaj, Steven Schmidt, Saeed Shafiee Sabet, Sebastian Möller, and
Carsten Griwodz. 2020. Quality estimation models for gaming video streaming
services using perceptual video quality dimensions. In Proceedings of the 11th
ACM Multimedia Systems Conference . 213–224.
[48] Matthew D Zeiler and Rob Fergus. 2014. Visualizing and understanding convolu-
tional networks. In European conference on computer vision . Springer, 818–833.
[49] Weixia Zhang, Dingquan Li, Chao Ma, Guangtao Zhai, Xiaokang Yang, and Kede
Ma. 2021. Continual learning for blind image quality assessment. arXiv preprint
arXiv:2102.09717 (2021).
[50] Weixia Zhang, Kede Ma, Guangtao Zhai, and Xiaokang Yang. 2021. Task-specific
normalization for continual learning of blind image quality models. arXiv preprintMM ’22, October 10–14, 2022, Lisboa, Portugal Wei Sun, Xiongkuo Min, Wei Lu, and Guangtao Zhai∗
arXiv:2107.13429 (2021).
[51] Weixia Zhang, Kede Ma, Guangtao Zhai, and Xiaokang Yang. 2021. Uncertainty-
aware blind image quality assessment in the laboratory and wild. IEEE Transac-
tions on Image Processing 30 (2021), 3474–3486.
[52] Qi Zheng, Zhengzhong Tu, Yibo Fan, Xiaoyang Zeng, and Alan C Bovik. 2022.
No-Reference Quality Assessment of Variable Frame-Rate Videos Using Tem-
poral Bandpass Statistics. In ICASSP 2022-2022 IEEE International Conference onAcoustics, Speech and Signal Processing (ICASSP) . IEEE, 1795–1799.
[53] Qi Zheng, Zhengzhong Tu, Pavan C Madhusudana, Xiaoyang Zeng, Alan C Bovik,
and Yibo Fan. 2022. FAVER: Blind Quality Prediction of Variable Frame Rate
Videos. arXiv preprint arXiv:2201.01492 (2022).