CAR-CNN: A Deep Residual Convolutional Neural Network for

Compression Artifact Removal in Video Surveillance Systems

Miloud Aqqa and Shishir K. Shah

Quantitative Imaging Laboratory, Department of Computer Science, University of Houston, U.S.A.

Keywords:

Compression Artifacts, Video Quality Enhancement, Deep Learning, Visual Surveillance.

Abstract:

Video compression algorithms are pervasively applied at the camera level prior to video transmission due to

bandwidth constraints, thereby reducing the quality of video available for video analytics. These artifacts

may lead to decreased performance of some core applications in video surveillance systems such as object

detection. To remove such distortions during video decoding, it is required to recover original video frames

from distorted ones. To this end, we present a fully convolutional residual network for compression artifact

removal (CAR-CNN) without prior knowledge on the noise distribution trained using a novel, differentiable

loss function. To provide a baseline, we also trained our model by optimizing the Structural Similarity (SSIM)

and Mean Squared Error (MSE). We test CAR-CNN on self-collected data, and we show that it can be applied

as a pre-processing step for the object detection task in practical, non-idealized applications where quality

distortions may be present.

1 INTRODUCTION

In real deployments of public safety video systems,

cameras are often backhauled via wireless media,

where packet loss and jitter impact video quality

available for video analytics. Furthermore, due to

bandwidth constraints, lossy compression algorithms

are often used to encode videos before transmission to

central storage and processing sites in order to avoid

network congestion and lower communication latency

at the expense of introducing undesired complex ar-

tifacts (e.g., blocking, blurring, ﬂoating), which re-

move textures and details in video frames as shown

in Figure 1. These artifacts are not only unpleasant to

the human eye but also adversely impact the perfor-

mance of various high-level vision algorithms such as

object detectors (Aqqa et al., 2019).

Typically, video compression algorithms come

with a parameter to tune the trade-off between ﬁle

size and quality of the video. The larger this pa-

rameter, the stronger are quality distortions stemming

from compression artifacts. However, merely opting

for low compression rates is not always a practical

solution. In video surveillance systems, video trans-

portation from the transmitting node to the video ana-

lytics compute engine is typically performed over an

IP network infrastructure, where transmission chan-

nels have limited bandwidth and are allowed a certain

quota per camera. This constraint is often ensured by

strong compression.

Most surveillance cameras adopt the H.264/AVC

standard for video encoding (Wiegand et al., 2003),

which is a lossy compression technique. A video

consists of images; an image is divided into slices

and blocks. A block is a square part (16×16, 8×8,

and 4×4) of the image. H.264 is a block-based

coder/decoder, meaning that a series of mathemat-

ical functions are applied on individual blocks to

achieve compression and decompression (Juurlink

et al., 2012). Also, it exploits spatial redundancy

within images and temporal redundancy in videos to

achieve appealing compression ratios with low la-

tency, making it a widely accepted standard for video

transmission for a myriad of applications.

Much work has been done on removing different

compression artifacts from JPEG images using differ-

ent techniques, from optimizing discrete cosine trans-

form (DCT) coefﬁcients (Zhang et al., 2013) to re-

cently leveraging the representational power of con-

volutional neural network (CNN) (Galteri et al., 2017;

Svoboda et al., 2016; Yu et al., 2015). CNNs have

been successfully used in many other image restora-

tion tasks including super-resolution (Kim et al.,

2016; Dong et al., 2014), denoising (Zhang et al.,

2017), structured noise removal (Eigen et al., 2013),

and blind deconvolution (Schuler et al., 2016) with

Aqqa, M. and Shah, S.

CAR-CNN: A Deep Residual Convolutional Neural Network for Compression Artifact Removal in Video Surveillance Systems.

DOI: 10.5220/0009184405690575

In Proceedings of the 15th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2020) - Volume 4: VISAPP, pages

569-575

ISBN: 978-989-758-402-2; ISSN: 2184-4321

569

Bitrate=2Mb/s, CRF=29 Bitrate=2Mb/s, CRF=35 Bitrate=2Mb/s, CRF=41 Bitrate=2Mb/s, CRF=47

Bitrate=1.5Mb/s, CRF=29 Bitrate=1.5Mb/s, CRF=35 Bitrate=1.5Mb/s, CRF=41 Bitrate=1.5Mb/s, CRF=47

Bitrate=1Mb/s, CRF=29 Bitrate=1Mb/s, CRF=35 Bitrate=1Mb/s, CRF=41 Bitrate=1Mb/s, CRF=47

Uncompressed frame

Uncompressed patch

Figure 1: An uncompressed patch taken from a video frame and its 12 compressed versions. The compression artifacts can

be visually perceived as CRF value increases and bitrate decreases. The combination of CRF=29 and maximum bitrate of

2Mb/s results in the lowest compressed version, thus better image quality. The combination of CRF=47 and maximum bitrate

of 1Mb/s results in highest compressed version, thus worst image quality.

state-of-the-art performance.

In this paper, we target the problem of compres-

sion artifact removal (CAR) in H.264/AVC encoded

videos. Compared to JPEG compression method

(Wallace, 1992), the degradation caused by H.264 is

not only originated from spatial artifacts (e.g. block-

ing, blurring, ringing) but also from temporal artifacts

(e.g., ﬂoating, ﬂickering). The aim is to devise an ap-

proach that can be applied as a post-processing step

on decompressed frames, and therefore it can be used

on many lossy video compression algorithms. In ad-

dition, it should be able to handle videos encoded at

different bit rates, and thus at different qualities. This

allows avoiding any changes to the existing compres-

sion pipelines, which are usually optimized (e.g., us-

ing dedicated hardware).

The contributions of our work are three-fold: (i)

We present a feed-forward fully convolutional resid-

ual generative network by optimizing a novel loss

function, and we show its superior performance com-

pared to de facto error metrics in image recon-

struction. (ii) We evaluate the performance of our

approach using self-collected data, which consists

of thirty uncompressed videos recorded in different

surveillance scenarios (indoor and outdoor). The

video frames from these videos are considered to be

of high quality. We augment this dataset by introduc-

ing complex artifacts under different levels of video

compression using H.264/AVC standard. (iii) We

show that our method can be used as a pre-processing

step for object detectors in video surveillance sys-

tems.

In section 2, we review some of the related work.

In section 3, we detail the proposed method and dif-

ferent loss functions used in this work. We describe

in section 4 the dataset, performance metrics, and im-

plementation details. Section 5 reports the results ob-

tained from our experiments. In section 6, we con-

clude our work.

2 RELATED WORK

There is a vast literature of image restoration, address-

ing image compression artifacts. These methods are

ranging from hand-designed ﬁlters relying on infor-

mation in the DCT domain to recently using deep con-

volutional neural networks (DCNN) following their

success in other machine vision tasks.

Simple artifact removal ﬁlters are included in

many software for handling images, videos, and other

multimedia ﬁles. For example, the FFmpeg frame-

work includes the simple postprocessing (ssp) ﬁlter

(Nosratinia, 1999), which applies JPEG compression

to the shifted versions of the already-compressed im-

age, and averages the results. Foi et al. devel-

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

570

oped Pointwise Shape-Adaptive DCT (SA-DCT) (Foi

et al., 2006), in which the thresholded transform co-

efﬁcients are used to reconstruct a local estimate of

the image signal within the adaptive-shape support.

Yang et al. presented in (Yang et al., 2000) a differ-

ent approach, which consists of applying DCT-based

lapped transform on the signal already in the DCT do-

main. The authors in (Li et al., 2014) decompose im-

ages into texture and structure components, then elim-

inate artifacts that are part of the texture component

due to contrast enhancement. Chang et al. (Chang

et al., 2014) propose to remove blocking artifacts of

JPEG compression images by ﬁnding a sparse rep-

resentation over a learned dictionary from a training

set of images. The main disadvantage of these algo-

rithms is that they explicitly attempt to reverse the ef-

fect of DCT-domain quantization optimally, and thus

they are very speciﬁc to the applied compressor. Fur-

thermore, they tend to overly smooth texture regions

without reproducing sharp edges.

The work presented in this paper was inspired

by DCNN based approaches. The main idea is to

learn an image transformation function that can pro-

duce a restored version of the given input image.

Dong et al. (Dong et al., 2015) propose artifact re-

duction CNN (AR-CNN), which extends their super-

resolution CNN (SRCNN) architecture with feature

enhancement layers. They trained AR-CNN in two

stages - a shallow network is trained ﬁrst, then it is

used as an initialization for a ﬁnal 4 layer CNN due

to training difﬁculties encountered when training the

later from scratch. Differently from AR-CNN, Svo-

boda et al. (Svoboda et al., 2016) report better re-

sults by training a feed-forward CNN that combines

residual learning and skip architecture to get a better

reconstruction quality.

Convolutional networks have successfully shown

their ability in different image transformation prob-

lems, such as image denoising (Zhang et al., 2017),

super-resolution (Kim et al., 2016; Dong et al., 2014),

and style-transfer (Gatys et al., 2016). Zhang et al.

(Zhang et al., 2017) propose a denoising convolu-

tional neural network (DnCNN) to eliminate Gaus-

sian noise, showing that residual learning and batch

normalization are beneﬁcial for this task. Kim et al.

(Kim et al., 2016) addressed the problem of image

super-resolution, using a deep architecture trained on

residual images. Ledig et al. (Ledig et al., 2017) pro-

pose a deep residual convolutional network, trained in

an adversarial fashion by optimizing a perceptual loss

that combines an adversarial loss and a content loss.

The authors state that their model can recover photo-

realistic textures from heavily downsampled images.

A style-transfer method of Gatys et al. (Gatys et al.,

2016) uses image representations from convolutional

neural network optimized for object recognition while

optimizing a loss that accounts for both image content

and style to keep the content of an arbitrary photo-

graph with the appearance of numerous well-known

artworks.

To the best of our knowledge, we are the ﬁrst to

adapt the idea of residual learning (Kim et al., 2016)

for H.264/AVC compression artifact removal based

on CNN in video surveillance systems. We follow

the assumption ”deeper is better”, and we propose 34-

layer fully convolutional residual neural network and

is, therefore, able to restore images of any resolution.

Moreover, we investigate the use of two loss functions

(MSE and SSIM), and deﬁne a new loss function that

combines the advantages of MSE and SSIM.

3 METHODOLOGY

In H.264/AVC compression artifact removal task, the

goal is to restore a video frame I

from a compressed

one I

. Most video compression algorithms (e.g.,

H.264/AVC, H.265/HEVC) encode images by imple-

menting less resolution for chroma information than

for luma information, taking advantage of the human

visual system’s lower acuity for color differences than

for luminance. Therefore, we transform the color im-

ages to YCbCr space, and only the luminance channel

Y is used further.

In [0, 255]

W ×H×C

, we deﬁne I

, I

, and I

as im-

age tensors with width W , height H and number of

image channels C. During H.264/AVC video encod-

ing process, an uncompressed image I

is encoded

by:

= E(I

, QP) (1)

using H.264 encoder E with some quantization pa-

rameter QP. The goal is to learn an inverse function

φ ≈ E

−1

to remove compression artifacts introduced

by E, thus restoring I

from I

≈ I

= φ(I

) (2)

To this end, we deﬁne φ(.) as a fully convolutional

residual network φ(I

;θ) with parameters θ that are

learned by optimizing a loss function l

CAR

. Given N

training video frames, we solve:

θ = argmin

∑

i=1

CAR

, φ(I

;θ)) (3)

In the following, we describe the architecture of

CAR-CNN and different loss functions used to remove

H.264/AVC compression artifacts.

CAR-CNN: A Deep Residual Convolutional Neural Network for Compression Artifact Removal in Video Surveillance Systems

571

Conv 3x3 n64 s1

Residual Block

…

Residual Block

Conv 3x3 n64 s1

Conv 3x3 n1 s1

Tanh

Conv 3x3 n64 s1

Figure 2: An Overview of CAR-CNN architecture. The network contains 15 residual blocks indicating with n the number of

ﬁlters and s the stride of each convolutional layer.

3.1 CAR-CNN Architecture

An overview of our proposed network is shown in

Figure 2. Inspired by (Kim et al., 2016), we propose

a 34-layers residual generative network that contains

only blocks of convolutional layers and LeakyReLU

non-linearities. In particular, we use convolutional

layers with 3 × 3 kernels and 64 feature maps. All

convolutional layers are followed by a LeakyReLU

activation with a slope of 0.2 for negative inputs. Af-

ter every convolution, we apply a padding of 1 pixel

to keep the same image size across all convolution

layers. Since the input images were transformed to

gray-scale considering only the luminance channel Y ,

the last layer is a simple convolutional layer with one

feature map followed by a tanh activation function to

keep the output values between the [−1, 1] range.

3.2 Loss Functions

In this section, we describe different loss functions

used to train CAR-CNN.

3.2.1 Mean Squared Error

As ﬁrst attempt to remove compression artifact, we

use Mean Squared Error loss (MSE):

MSE

W H

∑

i=1

∑

j=1

i, j

− I

i, j

)

(4)

Although MSE has shown improved performance

in JPEG artifact removal task (Svoboda et al., 2016),

its results are still sub-optimal as it doesn’t recover

most of the high-frequency details from a distorted

video frame.

3.2.2 Stuctural Similarity

A popular index in the image restoration task is the

structural similarity index (SSIM) (Wang et al., 2004).

It evaluates images accounting for the fact that the

human visual system is sensitive to changes in local

structure. If the goal is to produce visually pleasing

images, the loss function should be perceptually mo-

tivated, as is the case of SSIM.

SSIM for pixel p is deﬁned as:

SSIM(p) =

2µ

+ µ

2σ

i j

+ σ

(5)

The loss function for SSIM can be then written as:

SSIM

= 1 − SSIM(p) (6)

The network is trained to optimize the structural

similarity between the reference images and the re-

stored ones.

3.2.3 The Best of Both Worlds: SSIM + MSE

To capture the best characteristics of both SSIM and

MSE, we propose to combine them:

both

= α · l

SSIM

+ (1 − α) · l

MSE

(7)

where we empirically set α = 0.79. This is motivated

by the fact that MSE can excel in recovering low fre-

quencies while SSIM can better preserve contrast in

high-frequency regions.

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

572

Figure 3: Samples of video frames from uncompressed videos recorded in indoor and outdoor surveillance scenarios.

4 EXPERIMENTAL SETUP

4.1 Dataset

Previous work for JPEG compression artifact removal

were tested on BSDS500 (Martin et al., 2001) and

LIVE1 (Sheikh et al., 2014) datasets. These datasets

contain still images that have distinctly different char-

acteristics as compared to video frames encountered

in video surveillance systems. For this reason, we

have collected thirty uncompressed videos that rep-

resent common scenarios where video surveillance

cameras are deployed. The videos are 5 minutes long

movie clips and were acquired using AXIS P3227-

LVE network camera and recorded in 1080p high def-

inition (1920 × 1080) at 30fps. Samples of video

frames are shown in Figure 3.

H.264/AVC encoding uses Constant Rate Factor

(CRF) as the default quality (and rate control) setting.

CRF achieves constant quality by compressing dif-

ferent frames by different amounts, thus varying the

Quantization Parameter (QP) as necessary to main-

tain a certain level of perceived quality. It does this

by taking motion into account, similar to the encoder

on a surveillance camera. CRF ranges between 0 and

51, where lower values would result in better qual-

ity, and higher values lead to more compression. To

simulate the trade-of between quality and bitrate, we

have used CRF in conjunction with Video Buffer Ver-

iﬁer (VBV) mode to ensure that the bitrate is con-

strained to a certain maximum as in real-world set-

tings. An exhaustive combination of CRF values (29,

35, 41, and 47) and maximum bitrate values (2Mb/s,

1.5Mb/s, and 1Mb/s) are selected to create a total of

12 data variants.

4.2 Performance Metrics

The most wide-spread evaluation metrics for qual-

ity assessment in compression artifact removal task

are MSE, and the peak signal-to-noise ratio (PSNR),

which is the MSE normalized to the maximum pos-

sible signal values expressed in decibel (dB). An-

other alternative is to use the structural similarity in-

dex (SSIM), which is the mean of the product of three

terms assessing similarity in luminance, contrast, and

structure over multiple localized windows. We report

the mean PSNR and SSIM across the 12 data variants

created.

4.3 Implementation Details

For training our networks, we use 90k video frames as

the training set and 12k for the validation set. Testing

is performed on 17k video frames for each of the 12

data variants.

We use the PyTorch framework (Paszke et al.,

2017) for our evaluations. The training process was

distributed over two Nvidia Tesla v100 GPUs with a

mini-batch of 128 images and have been carried on

for 360 epochs. For each image, we ﬁrst rescale it

to (910 × 512), and then we randomly crop a 64 ×

64 patch with horizontal ﬂipping. We optimize the

network’s parameters with Adam (Kingma and Ba,

2015), starting with a learning rate of 10

−4

and a mo-

mentum of 0.9.

5 RESULTS

5.1 Artifact Removal

The evaluation results of the mean PSNR and SSIM

across different data variants are shown in Table 1

and Table 2, respectively. We compare our results

to the simple postprocessing ﬁlter (ssp) in the FFm-

peg framework. As can be seen from the results, our

approach outperforms ssp in all data variants. More

speciﬁcally, in the case of the highest compression

level (i.e., Bitrate=1Mb/s and CRF=47), we can see

an improvement in PSNR of 0.30 dB over ssp while

SSIM is improved from 0.604 to 0.693. At the low-

est compression (i.e., Bitrate=2Mb/s and CRF=29),

CAR-CNN: A Deep Residual Convolutional Neural Network for Compression Artifact Removal in Video Surveillance Systems

573

Table 1: Restoration Quality Comparison. Results reported for average PSNR (dB) using luminance.

Method

Bitrate = 2 Mb/s Bitrate = 1.5 Mb/s Bitrate = 1 Mb/s

CRF-29 CRF-35 CRF-41 CRF-47 CRF-29 CRF-35 CRF-41 CRF-47 CRF-29 CRF-35 CRF-41 CRF-47

ssp 29.65 28.09 25.93 24.53 29.25 28.07 25.90 24.53 28.58 27.70 25.85 24.53

Our MSE 30.07 28.36 26.65 25.27 29.61 28.36 26.64 25.17 28.88 27.98 26.62 25.05

Our SSIM 29.69 28.12 26.02 24.78 29.32 28.12 25.91 24.78 28.54 27.72 25.88 24.78

CAR-CNN 29.76 28.19 26.07 24.81 29.38 28.18 25.96 24.81 28.60 27.78 25.93 24.81

Table 2: Restoration Quality Comparison. Results reported for average SSIM using luminance.

Method

Bitrate = 2 Mb/s Bitrate = 1.5 Mb/s Bitrate = 1 Mb/s

CRF-29 CRF-35 CRF-41 CRF-47 CRF-29 CRF-35 CRF-41 CRF-47 CRF-29 CRF-35 CRF-41 CRF-47

ssp 0.817 0.789 0.725 0.604 0.807 0.788 0.723 0.604 0.787 0.774 0.720 0.604

Our MSE 0.827 0.805 0.758 0.681 0.822 0.801 0.757 0.681 0.794 0.784 0.756 0.681

Our SSIM 0.845 0.829 0.777 0.689 0.837 0.819 0.763 0.689 0.817 0.794 0.761 0.687

CAR-CNN 0.873 0.867 0.786 0.695 0.858 0.826 0.762 0.693 0.835 0.796 0.765 0.693

Table 3: Detection performance of YOLO measured as mean average precision (mAP) at IoU=0.50 on the 12 data variants

for different reconstruction methods.

Method

Bitrate = 2 Mb/s Bitrate = 1.5 Mb/s Bitrate = 1 Mb/s

CRF-29 CRF-35 CRF-41 CRF-47 CRF-29 CRF-35 CRF-41 CRF-47 CRF-29 CRF-35 CRF-41 CRF-47

ssp 0.766 0.736 0.661 0.522 0.756 0.735 0.661 0.519 0.745 0.733 0.557 0.519

Our MSE 0.774 0.759 0.698 0.577 0.768 0.761 0.697 0.577 0.758 0.753 0.669 0.577

Our SSIM 0.783 0.772 0.703 0.587 0.779 0.773 0.706 0.587 0.764 0.759 0.675 0.587

CAR-CNN 0.789 0.781 0.723 0.589 0.783 0.776 0.727 0.589 0.769 0.763 0.676 0.588

we see a gain of 0.42 dB in PSNR over ssp, and

we improve SSIM from 0.817 to 0.873. From a

quality index point of view, MSE based model out-

performs SSIM and (SSIM+MSE) based models in

PSNR, which can be explained by the fact that MSE

tends to evaluate better more blurry and smooth re-

gions than realistic textures.

5.2 Object Detection

In video surveillance systems, we are more inter-

ested in understanding how video quality affects ma-

chine vision algorithms. During video compression,

quality distortions stemmed from spatial and tempo-

ral artifacts are introduced to the video frames lead-

ing to decreased performance of object detectors,

as shown in (Aqqa et al., 2019). This degradation

in performance can be explained because compres-

sion artifacts remove textures and details in these

video frames. These high-frequency features repre-

sent edges and shapes of objects that the detector may

be looking for to classify an object.

In this experiment, we select YOLO (Redmon

et al., 2016) as the object detector. We consider its

detections on uncompressed videos as ground-truth

bounding boxes, and we compare them against its

detections on reconstructed versions of the 12 com-

pressed variants. As a lower bound, we report per-

formance on images restored using ssp; results are

reported in Table 3. As can be seen, the perfor-

mance drop of YOLO with respect to its perfor-

mance on uncompressed video frames is signiﬁcant.

Even for moderate compression levels (i.e., CRF=29),

the performance decreases by at least 21.1%. Our

(SSIM+MSE) based network outperforms MSE and

SSIM based networks across all data variants. In

particular, at the lowest compression level (i.e., Bi-

trate=2Mb/s and CRF=29) the improvement in mAP

using CAR-CNN is the following: ssp(+2.3%),

MSE(+1.5%) and SSIM(+0.6%) highlighting the

beneﬁts of combining both MSE and SSIM for the

compression artifacts removal task. As can be ob-

served in Table 3, at the highest compression level

(i.e., Bitrate=1Mb/s and CRF=47), our approach has

an improvement in mAP of 6.9% over ssp, 1.1% over

MSE, and 0.1% over SSIM.

6 CONCLUSION

We have presented a 34-layer fully convolutional

residual neural network for H.264/AVC compres-

sion artifact removal in video surveillance systems.

Our network was trained using a novel loss func-

tion and outperforms MSE and SSIM based networks.

Moreover, we have shown that it is possible to im-

prove video quality, both for human viewers and ma-

chines and thus avoiding changes to the compression

pipelines through applying a pre-processing step that

removes compression artifacts.

VISAPP 2020 - 15th International Conference on Computer Vision Theory and Applications

574

ACKNOWLEDGEMENT

This work was performed in part through the ﬁnan-

cial assistance award, Multi-tiered Video Analytics

for Abnormality Detection and Alerting to Improve

Response Time for First Responder Communications

and Operations (Grant No. 60NANB17D178), from

U.S. Department of Commerce, National Institute of

Standards and Technology.

REFERENCES

Aqqa, M., Mantini, P., and Shah, S. K. (2019). Understand-

ing how video quality affects object detection algo-

rithms. In 14th International Conference on Computer

Vision Theory and Application.

Chang, H., Ng, M. K., and Zeng, T. (2014). Reducing ar-

tifacts in jpeg decompression via a learned dictionary.

IEEE Transactions on Image Processing, 62:718–728.

Dong, C., Deng, Y., Loy, C. C., and Tang, X. (2015). Com-

pression artifacts reduction by a deep convolutional

network. In IEEE International Conference on Com-

puter Vision (ICCV).

Dong, C., Loy, C. C., He, K., and Tang, X. (2014). Learn-

ing a deep convolutional network for image super-

resolution. In European Conference on Computer Vi-

sion (ECCV).

Eigen, D., Krishnan, D., and Fergus, R. (2013). Restoring

an image taken through a window covered with dirt or

rain. In IEEE International Conference on Computer

Vision (ICCV).

Foi, A., Katkovnik, V., and Egiazarian, K. (2006). Point-

wise shape-adaptive dct for high-quality deblocking

of compressed color images. In 14th European Signal

Processing Conference.

Galteri, L., Seidenari, L., Bertini, M., and Bimbo, A. D.

(2017). Deep generative adversarial compression ar-

tifact removal. In IEEE International Conference on

Computer Vision (ICCV).

Gatys, L. A., Ecker, A. S., and Bethge, M. (2016). Im-

age style transfer using convolutional neural networks.

In IEEE Conference on Computer Vision and Pattern

Recognition (CVPR).

Juurlink, B., Alvarez-Mesa, M., Chi, C. C., Azevedo, A.,

Meenderinck, C., and Ramirez, A. (2012). Under-

standing the application: An overview of the h.264

standard. Scalable Parallel Programming Applied to

H.264/AVC Decoding, pages 5–15.

Kim, J., Lee, J. K., and Lee, K. M. (2016). Accurate image

super-resolution using very deep convolutional net-

works. In IEEE Conference on Computer Vision and

Pattern Recognition (CVPR).

Kingma, D. P. and Ba, J. (2015). Adam: A method for

stochastic optimization. In the 3rd International Con-

ference for Learning Representations.

Ledig, C., Theis, L., Husz

ar, F., Caballero, J., Cunning-

ham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J.,

Wang, Z., and Shi, W. (2017). Photo-realistic single

image super-resolution using a generative adversarial

network. In IEEE Conference on Computer Vision and

Pattern Recognition (CVPR).

Li, Y., Guo, F., Tan, R. T., and Brown, M. S. (2014). A

contrast enhancement framework with jpeg artifacts

suppression. In European Conference on Computer

Vision (ECCV).

Martin, D. R., Fowlkes, C., Tal, D., and Malik, J. (2001). A

database of human segmented natural images and its

application to evaluating segmentation algorithms and

measuring ecological statistics. In IEEE International

Conference on Computer Vision (ICCV).

Nosratinia, A. (1999). Embedded post-processing for en-

hancement of compressed images. In Proceedings

DCC’99 Data Compression Conference.

Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E.,

DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., and

Lerer, A. (2017). Automatic differentiation in Py-

Torch. In NeurIPS Autodiff Workshop.

Redmon, J., Divvala, S., Girshick, R., and Farhadi, A.

(2016). You only look once: Uniﬁed, real-time object

detection. In IEEE Conference on Computer Vision

and Pattern Recognition (CVPR).

Schuler, C. J., Hirsch, M., Harmeling, S., and Sch

olkopf,

B. (2016). Learning to deblur. IEEE Transactions on

Pattern Analysis and Machine Intelligence, 38:1439–

1451.

Sheikh, H. R., Wang, Z., Cormack, L., , and Bovik, A. C.

(2014). Live image quality assessment database re-

lease 2.

Svoboda, P., Hradis, M., Ba

rina, D., and Zemc

ık, P. (2016).

Compression artifacts removal using convolutional

neural networks. Journal of WSCG, 24:63–72.

Wallace, G. K. (1992). The jpeg still picture compression

standard. IEEE Transactions on Consumer Electron-

ics, 38.

Wang, Z., Bovik, A., Sheikh, H., and Simoncelli, E. (2004).

Image quality assessment: from error visibility to

structural similarity. IEEE Transactions on Image

Processing, 13:600–612.

Wiegand, T., Sullivan, G. J., Bjontegaard, G., and Luthra,

A. (2003). Overview of the h.264/avc video coding

standard. IEEE Transactions on Circuits and Systems

for Video Technology, 13:560–576.

Yang, S., Kittitornkun, S., Hu, Y.-H., Nguyen, T., and Tull,

D. (2000). Blocking artifact free inverse discrete co-

sine transform. In Proceedings 2000 International

Conference on Image Processing.

Yu, K., Dong, C., Deng, Y., Loy, C. C., and Tang, X. (2015).

Compression artifacts reduction by a deep convolu-

tional network. In IEEE International Conference on

Computer Vision (ICCV).

Zhang, K., Zuo, W., Chen, Y., Meng, D., and Zhang, L.

(2017). Beyond a gaussian denoiser: Residual learn-

ing of deep cnn for image denoising. IEEE Transac-

tions on Image Processing, 26:3142–3155.

Zhang, X., Xiong, R., Fan, X., and Gao, W. (2013).

Compression artifact reduction by overlapped-block

transform coefﬁcient estimation with block similarity.

IEEE Transactions on Image Processing, 22:4613–

4626.

CAR-CNN: A Deep Residual Convolutional Neural Network for Compression Artifact Removal in Video Surveillance Systems

575