Image 2 PDF
Image 2 PDF
Image 2 PDF
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806281, IEEE
Transactions on Image Processing
Abstract—Transform domain downward conversion (TDDC) block is first transformed by the discrete cosine transform
for image coding is usually implemented by discarding some (DCT). Then, the transform coefficients are quantized by
high-frequency components from each transformed block. As a uniform quantization and further compressed through an
a result, a block of fewer coefficients is formed and a lower
compression cost is achieved due to the coding of only a entropy coding. All these operations aim at removing the
few low-frequency coefficients. In this paper, we focus on the spatial redundancy within the raw image data, and a number
design of a new TDDC-based coding method by using our of modifications on them have been developed to improve the
proposed interpolation-compression directed filtering (ICDF) and coding performance over the past few years [3-8]. Meanwhile,
error-compensated scalar quantization (ECSQ), leading to the some advanced techniques, such as image inpainting [9, 10],
compression-dependent TDDC (CDTDDC) based coding. More
specifically, ICDF is first used to convert each 16 × 16 macro- graph-based transform [11, 12], object extraction [13, 14],
block into an 8 × 8 coefficient block. Then, this coefficient block just-noticeable difference (JND) based perceptual optimiza-
is compressed with ECSQ, resulting in a smaller compression tion [15], and compressive sensing [16], have been applied
distortion for those pixels that locate at some specific positions to image compression, and significant improvements for the
of a macro-block. We select these positions according to the 4:1 coding performance have been achieved.
uniform sub-sampling lattice and use the pixels locating at them
to reconstruct the whole macro-block through an interpolation. In image coding, removing the spatial redundancy from the
The proposed CDTDDC-based coding can be applied to compress raw image data is always one of the top-priority considerations
both grayscale and color images. More importantly, when it to reach a high compression. One straightforward way to
is used in the color image compression, it offers not only a achieve this goal is to perform the spatial sub-sampling on
new solution to reduce the data-size of chrominance components the source image to reduce its resolution, which accordingly
but also a higher compression efficiency. Experimental results
demonstrate that applying our proposed CDTDDC-based coding produces a low-resolution (LR) image and forms the pixel-
to compress still images can achieve a significant quality gain domain downward conversion (PDDC). Then, all compression
over the existing compression methods. operations are performed on the LR image to conduct the
Index Terms—Image coding, sub-sampling, interpolation, dis- encoding. After the decoding is completed, an image inter-
tortion, quantization polation is normally performed on the decoded LR image
to reconstruct a full-resolution image. The decimation of the
source image makes a lower bit-cost, which may offer a po-
I. I NTRODUCTION tential improvement on the compression efficiency. The coding
efficiency of such a PDDC-based compression for grayscale
T HE popular JPEG coding scheme for still images [1]
is built upon the block-based framework, which leads
to a rather low complexity. Such a low complexity also
images has been verified in [17], where the LR image is
generated by down-sampling the source image according to
makes JPEG used more frequently than JPEG2000 [2] for a uniform 4:1 sub-sampling lattice and the LR image is com-
the compression of image signals, although the latter one pressed by JPEG. This method has been effectively improved
achieves a higher compression efficiency. In JPEG, the source through introducing more sub-sampling modes to implement
image is partitioned into non-overlapped blocks and each the spatial down-sampling for the compression [18].
Although the PDDC-based coding offers a lower bit-cost, its
This work was supported by the National Natural Science Foundation compression efficiency is often limited by image interpolation.
of China under Grant 61672134, Grant 61720106004, by the National Key
Basic Research Program of China under Grant 2015CB351804, by the To solve this problem, the adaptive interpolation-directed sub-
Fundamental Research Funds for Central Universities of China under Grant sampling is proposed to get a better interpolation [19]. It has
ZYGX2016J038, by the “111” Projects under Grant B17008, by the Macau been noticed that a so-called collaborative adaptive down-
Science and Technology Development Fund under Grant FDCT/022/2017/A1
and by the Research Committee at the University of Macau under Grant sampling and up-conversion (CADU) [20] is developed to pre-
MYRG2016-00137-FST. (Corresponding authors: Jiantao Zhou; Bing Zeng.) serve the edge information in the down-sized image and it has
S. Y. Zhu, Z. Y. He and B. Zeng are with School of Information and Com- been used in the PDDC-based coding. Also, the interpolation-
munication Engineering, University of Electronic Science and Technology of
China, Chengdu, China. dependent image down-sampling (IDID) [21] is designed to
X. D. Meng is with Department of Electronic and Computer Engineering, build up a block-based image compression. Meanwhile, the
The Hong Kong University of Science and Technology, Kowloon, Hong Kong, super-resolution directed down-sampling (SRDDS) [22] that
China.
J. T. Zhou is with Faculty of Science and Technology, University of Macau, optimizes the sub-sampling via minimizing the mean square
Macau, China. error (MSE) between the source image and the up-converted
1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806281, IEEE
Transactions on Image Processing
Encoding
zeros. As a result, this 8 × 8 small-sized coefficient block is
reserved for the further compression and the other coefficients
0
ICDF Truncation ECSQ are all removed, achieving the downward conversion in the
transform domain. Then, the small-sized coefficient block is
0 0
compressed with our proposed error-compensated scalar quan-
tization (ECSQ) to produce a high-quality compression for
...
Dequantization those pixels locating at some specific positions of the macro-
0
Interpolation IDCT
Zero Padding block. In this work, these positions are defined according to
the 4:1 uniform sub-sampling lattice, and the pixels locating at
0 0
these positions are used to interpolate the whole macro-block
Decoding
after compression.
Coupling ICDF and ECSQ together, we build up the
Fig. 1. Framework of CDTDDC-based coding. compression-dependent TDDC (CDTDDC) for the compres-
sion of image signals and the framework of this coding scheme
image has also been used successfully to implement a block- is shown in Fig. 1. Moreover, when the proposed CDTDDC-
based image compression. To make the sub-sampling based based coding is adopted in the compression of grayscale
coding more efficient, a rate distortion directed decimation images, it will work competitively with the JPEG baseline
filter is proposed in [23], where the filtering parameters coding as two coding modes for each macro-block. On the
are designed by minimizing both the interpolation distortion other hand, when it is used to compress color images, it is only
and the quantization distortion. Recently, another spatial sub- performed on two chrominance components after the RGB-
sampling based compression is proposed in [24], where the to-YCbCr conversion. In this way, it offers not only a new
sub-sampling of image blocks is determined adaptively by solution to reduce the data-size of color images but also a
a JND-based threshold. Although the spatial down-sampling high compression efficiency.
based coding offers considerable benefits to image compres- It is important to point out that our proposed method is quite
sion [25, 26], how to effectively reduce both the interpolation different from the traditional PDDC-based method. Firstly, we
distortion and the quantization distortion is still a challenge reduce the data-size of the source image in the transform
problem in this coding scheme. domain rather than in the pixel domain. Secondly, we design
So far, most of the advanced PDDC-based methods have the ICDF algorithm to preserve the high-frequency information
been developed for the compression of grayscale images. for the source image in the transform domain, making it
When these methods are adopted in color image compression more efficient than the adaptive spatial down-sampling [20-22]
after the RGB-to-YCbCr conversion, they are only applied to adopted in the PDDC-based method. Thirdly, in our proposed
the chrominance components for an overall bit-cost saving. CDTDDC-based coding, the ICDF algorithm is designed
The most popular PDDC-based color image coding is the 4:2:0 not only to make a high-quality reconstruction but also to
chroma format coding in which the chrominance components guarantee a low bit-cost for the compression. Meanwhile,
are directly sub-sampled along both the horizonal and vertical the ECSQ algorithm is proposed to reduce the compression
directions by a factor of 2. Based on such a sub-sampling, distortion and to improve the final reconstruction quality. In
the spatial redundancy of the chrominance component is re- the traditional PDDC-based compression, the adaptive down-
moved effectively and a higher overall compression efficiency sampling is only designed for a a high-quality reconstruction,
is achieved accordingly. This coding format has been used where the optimization of the compression-cost is ignored.
most frequently in the compression of color images and also The rest of this paper is organized as follows. We first
improves the compression efficiency very significantly. present a brief review on the TDDC-based coding and the
The image downward conversion may also be implemented interpolation-directed image down-sampling in Section II,
in the transform domain by discarding some high-frequency where both techniques are related to our proposed method.
components from the transformed block [27], which pro- Then, we describe how to implement our CDTDDC-based
duces a small-sized coefficient block and yields the transform coding in Section III in which both ICDF and ECSQ are intro-
domain downward conversion (TDDC). However, when the duced and the effectiveness of them are verified. Experimental
reconstruction is carried out, a serious quality degradation results are presented in Section IV and conclusions are finally
often happens due to the lack of necessary high-frequency drawn in Section V.
information. This also limits the application of TDDC in
practical image coding. II. BACKGROUND
In this work, we focus on the design of a new TDDC-based
coding to compress still images, including both grayscale A. TDDC-based coding
and color images. To achieve this goal, we first propose an Let us use b to denote an N × N image block, extracted
interpolation-compression directed filtering (ICDF) to convert from either a grayscale image or the chrominance component
each 16×16 macro-block into an 8×8 coefficient block. After of an RGB-to-YCbCr converted image. After the 2-D DCT
ICDF is carried out, only the 8 × 8 sub-block locating at the is performed on it, we get the transformed block B, where
top-left corner of the transformed macro-block receives mean- B = CbCT and C is the DCT matrix. Moreover, we can
ingful transform coefficients, while all other positions receive represent B as
1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806281, IEEE
Transactions on Image Processing
1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806281, IEEE
Transactions on Image Processing
III. P ROPOSED M ETHOD Based on Eqs. (4) and (8), the interpolation by using xΩ
may be represented by the coefficient vector as
In this work, we propose a new TDDC-based coding to
compress each 16 × 16 macro-block. In general, the transform
xI = HDX. (10)
domain downward conversion is implemented by using an
advanced pre-filtering and the reconstruction of a compressed According to the IDID-based solution, an interpolation-
macro-block is achieved based on the image interpolation. optimized X may be determined as
The pre-filtering used in our work not only facilitates an
efficient interpolation at the decoder side but also produces a X̃ = arg min kx − HDXk22 . (11)
X
specific transformed macro-block (as Eq. (2) represents) at the
encoder side, which leads to the transform domain downward Therefore, the optimal X is calculated as
conversion. To this end, an interpolation-compression directed −1
X̃ = (HD)T (HD) (HD)T x .
filtering (ICDF) is proposed and performed on each macro- (12)
block before compression. Then, an error-compensated scalar In this work, we aim at an optimal X which not only
quantization (ECSQ) is designed to reduce the distortion produces an efficient interpolation for a better reconstruction,
occurring on some specific pixels of the compressed macro- but also effectively controls the compression cost, i.e., the bit-
block, where these pixels will be collected to implement the rate, for practical coding. However, the IDID-based solution
interpolation for the reconstruction of a macro-block. only targets an optimal interpolation, but ignores the cost for
the compression.
A. Interpolation-compression directed filtering It has been noticed that a lower compression cost is highly-
related to the sparsity of the transformed image signal in
We present our analysis with the 1-D representation for the image compression. In other words, a very sparse transformed
corresponding 2-D image block in this subsection. Firstly, we image signal often leads to a much easier compression. DCT
compose a transform matrix Ψ by performing the Kronecker has shown an excellent energy-compaction capability [35, 36],
product, denoted as ⊗, on the DCT matrix C as Ψ = C ⊗ C. which makes the image signal sparse enough in the transform
Then, we concatenate all columns of B to form a coefficient domain, as proved by the Laplacian distribution of transform
vector X. Based on Ψ and x, X may be obtained by X = coefficients [37-39]. Therefore, it has been adopted in all of
Ψx, which implies that the 2-D transform has been achieved the block-based image and video codecs [1, 40, 41] to help
via the 1-D operation. Finally, the inverse transform will be compress image and video signals. On the other hand, the
implemented as sparsity of the transformed image signal is usually evaluated
x = Ψ−1 X. (7) by the l0 -norm.
To achieve a better interpolation as well as a lower compres-
In this work, the compressed pixels locating at some pre- sion cost, the l0 -norm of the transformed block is involved in
determined positions (denoted as Ω) of an N ×N macro-block the interpolation-directed pre-filtering, leading to our proposed
will be used to reconstruct a decompressed macro-block via ICDF algorithm. Moreover, after such a filtering, the optimal
image interpolation. Specifically, these positions are defined interpolation-compression directed transform coefficients are
according to the 4:1 uniform sub-sampling lattice. Regardless obtained and they are determined as
of the compression, let xΩ be composed by K = N 2 /4 pixels
X̃ = arg min kx − HDXk22 + λkXk0 .
locating at Ω in b. Typically, these pixels are picked out from (13)
b along the vertical direction. If we compose a K ×N 2 matrix X
D by using K rows of Ψ−1 corresponding to K selected The minimization of l0 -norm is an NP-hard problem and
positions belonging to Ω, we may get thus the l1 -norm is often used as an alternative solution
in practical applications. As a result, the optimal transform
xΩ = DX. (8) coefficients may be obtained as
Moreover, if the positions of X’s elements are changed,
X̃ = arg min kx − HDXk22 + λkXk1 .
(14)
some necessary column-swaps must be performed on D to X
guarantee the correct output for xΩ . More specifically, let The solution to Eq. (14) may be effectively achieved
us use X< to represent a column-vector composed by K by using the iterative shrinkage-thresholding algorithm
coefficients locating at < in B and use X<¯ to denote another (ISTA) [42, 43]. In our work, we aim at a TDDC-based coding
column-vector composed by the other (N 2 − K) coefficients and want to make the coefficients locating at < ¯ be zeros
¯ Then, we combine X< and X ¯ together to
locating at <. < after the ICDF is carried out. To achieve this goal, we design
construct X as X = [XT< XT<¯ ]T . Meanwhile, we use K a compression-dependent solution based on the fast iterative
corresponding columns of D to compose a matrix D< and use shrinkage-thresholding algorithm (FISTA) [44] and make the
the other (N 2 − K) columns of it to compose another matrix ¯ change to zeros in the process of
coefficients locating at <
D<¯ . Coupling D< and D<¯ together, we get D = [D< D<¯ ] solving Eq. (14). The details about how to implement the
which makes ICDF algorithm have been summarized in Algorithm 1, where
X< all parameters and pre-defined thresholds are initialized by the
xΩ = [D< D<¯ ] . (9)
X<¯ corresponding default values in [44].
1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806281, IEEE
Transactions on Image Processing
35
B. Error-compensated scalar quantization
30
25
The application of ICDF in the TDDC-based coding aims at
a better interpolation and a lower compression cost. However,
PSNR(dB)
20
when the compression happens, the interpolation efficiency as
15
well as the coding efficiency will be limited by the distortion
10
Lena occurring on those filtered pixels (denoted as x̃Ω ) that will
Boat
5
0 10 20 30 40 50 60
Iteration number
70 80 90 100 be used for interpolation. To solve this problem, we purpose
to reduce the sum of square error (SSE) distortion of x̃Ω
Fig. 2. Verification of the convergence for ICDF. as much as possible via controlling the quantization error of
the transformed macro-block based on an error-compensated
scalar quantization (ECSQ).
Similarly to the IDID-based coding, applying ICDF to
image compression also needs to avoid producing the mis- Due to the use of a unitary transform in the compression, the
matched interpolation matrices at the encoder side and at the SSE distortion of all pixels and the SSE distortion of all trans-
decoder side. In this work, we choose to use an interpolation form coefficients are exactly identical. In this work, we only
with fixed coefficients to implement ICDF as well as the final focus on the SSE distortion occurring on x̃Ω rather than the
reconstruction. To this end, we verify the effectiveness of whole filtered block. First of all, the coefficient vector obtained
different interpolation methods by separately integrating them by using ICDF may be represented as X̃ = [X̃T< X̃T<¯ ]T , where
in our proposed method to compress images. More details and X̃< is a column-vector composed by K coefficients locating
results about such a verification will be presented in Section at < in a coefficient block after our proposed filtering and
IV, where the bicubic interpolation performs best among all X̃<¯ is composed by the other (N 2 − K) coefficients. Since
compared methods. Therefore, the bicubic interpolation is X̃<¯ = 0, according to Eq. (9), we may get x̃Ω = D< X̃< . after
adopted in the ICDF algorithm to implement our proposed compression, the SSE distortion of x̃Ω may be calculated as
compression method.
To verify the convergence of ICDF, we apply it to the 16 × 2
kx̃Ω − x̂Ω k2 =(X̃< − X̂< )T D< T D< (X̃< − X̂< )
16 macro-blocks of two popular grayscale images of Lena
and Boat. The PSNR values of interpolated images versus the =(X̃< − X̂< )T W(X̃< − X̂< ) (15)
T
iteration number of ICDF are illustrated in Fig. 2. According =∆X̃ W∆X̃
1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806281, IEEE
Transactions on Image Processing
x 10
4 Lena x 10
4 Lena x 10
4 Lena
4 4 4
3 3 3
2 2 2
1 1 1
0 0 0
−800 −600 −400 −200 0 200 400 600 800 −800 −600 −400 −200 0 200 400 600 800 −800 −600 −400 −200 0 200 400 600 800
4
x 10 Boat 4
x 10 Boat 4
x 10 Boat
3 3 3
2 2 2
1 1 1
0 0 0
−800 −600 −400 −200 0 200 400 600 800 −800 −600 −400 −200 0 200 400 600 800 −800 −600 −400 −200 0 200 400 600 800
Lena Boat
36 32
31
35
30
34
PSNR(dB)
PSNR(dB)
29
33
28
32
27
JPEG coding JPEG coding
31 IDID−based coding IDID−based coding
26
ICDF−based coding ICDF−based coding
ICDF+ECSQ−based coding ICDF+ECSQ−based coding
30 25
0.15 0.2 0.25 0.3 0.35 0.4 0.15 0.2 0.25 0.3 0.35 0.4
Bitrate(bpp) Bitrate(bpp)
1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806281, IEEE
Transactions on Image Processing
7
Lena Boat
(only pixels used for interpolation) (only pixels used for interpolation)
36 32
35
31
34
30
PSNR(dB)
PSNR(dB)
33
29
32
28
31
ICDF−based coding ICDF−based coding
ICDF+ECSQ−based coding ICDF+ECSQ−based coding
30 27
0.15 0.2 0.25 0.3 0.35 0.4 0.15 0.2 0.25 0.3 0.35 0.4
Bitrate(bpp) Bitrate(bpp)
Fig. 6. Verification of the effectiveness for ECSQ (only pixels used for interpolation are included).
K−1
X 2
Ek = ϕk,i ∆X̃i . (19) Algorithm 2
i=k 1. Initialization:
It is found from Eq. (18) that many cross-terms are involved (1) Quantize X̃K−1 : X̂K−1 = Q(X̃K−1 );
to determine SSEΩ so that the uniform quantization as used in (2) Calculate the quantization error:
the JPEG baseline coding will no longer be the best strategy to ∆X̃K−1 = X̃K−1 − X̂K−1 .
get a minimal SSEΩ . However, SSEΩ cannot be minimized 2. Error compensation:
by simply taking partial derivatives on ∆X̃i because the For k = (K − 2) : −1 : 0
quantization is a non-linear operation and is always required (1) Calculate the compensation amount:
to implement the data compression. ∆k = (1/ϕk,k ) K−1
P
i=k+1 ϕk,i ∆X̃i ;
In fact, the quantization error can be positive and negative in
(2) Perform the compensation on X̃k : X̄k = X̃k + ∆k ;
a random fashion. When two quantization errors have different
(3) Quantize X̄k : X̂k = Q(X̄k );
signs, the resulting cross-term would contribute a reduction to
SSEΩ . Different combinations of such cross-terms will lead (4) Calculate the quantization error:
to different distortion and the minimal one may be found by ∆X̃k = X̃k − X̂k .
trying an exhausted-search on all possible distortion combina- End for
tions. However, such a search strategy becomes impossible in
practical applications and it is thus necessary to find a more
efficient way to reduce the distortion for x̃Ω .
According to Eq. (18), it also reveals that a reduced SSEΩ
may be achieved by minimizing each Ek separately.
PK−1 When k < based on all pixels used for interpolation is shown in Fig. 6.
K − 1, let’s use ∆k to represent (1/ϕk,k ) i=k+1 ϕk,i ∆X̃i . It is found from Fig. 6 that the reconstruction quality of those
As a result, Ek may be further represented as pixels involved in interpolation has been improved after ECSQ
Ek =ϕ2k,k ∆X̃k + ∆k
2 is carried out. As a result, an overall quality gain has been
2 (20) achieved accordingly, as proved by the better R-D performance
=ϕ2k,k X̃k − Q(X̃k ) + ∆k
offered by ECSQ in Fig. 5.
where X̃k ∈ X̃< , Q(·) stands for the quantization carried
out on the transform coefficient and ∆X̃k = (X̃k − Q(X̃k )).
Thus, no mater how X̃k ’s are quantized, the quantization errors C. Our CDTDDC-based coding
produced on them will contribute a compensation term ∆k to The proposed ICDF is used to implement the downward
X̃k . Based on Eq. (20), as long as this compensation term conversion in the transform domain for an efficient interpo-
is added into X̃k , the quantization
2 on X̄k = X̃k + ∆k will lation as well as a low bit-cost. Meanwhile, the proposed
2
yield an item of X̄k − Q(X̄k ) = ∆X̄k , which would be ECSQ makes a low compression distortion for those pixels
2
smaller (statistically) than ∆X̃k + ∆k . This finally leads used for interpolation. Coupling them together, we build up
to a new quantization algorithm, i.e., the proposed ECSQ, our CDTDDC-based coding for the compression of image
and it has been summarized in Algorithm 2. After ECSQ is signals. The framework for the CDTDDC-based coding has
performed on each X̃k (0 ≤ k < K −1), we may get a smaller been shown in Fig. 1. According to Fig. 1, it is found
SSE distortion SSEΩ through combining all Ek together. that most additional operations are carried out only at the
We verify the effectiveness of the proposed ECSQ algorithm encoder side, and we will make a detailed discussion about
by applying it to the test images of Lena and Boat. More the computational complexity in next section. In our work, the
specifically, it is performed on each 16 × 16 macro-block to ICDF algorithm is performed on each 16 × 16 macro-block if
make a further compression after the ICDF-based filtering. it is needed. We will further verify the compression efficiency
Firstly, the overall R-D performance for the whole image is of our proposed CDTDDC-based coding through applying it
presented in Fig. 5. Then, the R-D performance evaluated to compress both grayscale and color images.
1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806281, IEEE
Transactions on Image Processing
Lena Boat
38 36
37 35
36 34
33
35
32
PSNR(dB)
PSNR(dB)
34
31
33
30
32
29
31 28
With bilinear interpolation With bilinear interpolation
30 With Lanczos interpolation 27 With Lanczos interpolation
With bicubic interpolation With bicubic interpolation
29 26
0.1 0.2 0.3 0.4 0.5 0.6 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Bitrate(bpp) Bitrate(bpp)
Fig. 8. Comparison of using different interpolations in our proposed method for grayscale image coding.
38 35 36
37 34 35
36 33 34
35 32 33
PSNR(dB)
PSNR(dB)
PSNR(dB)
34 31 32
33 30 31
32 29 30
36 31
36
35 30
34 34 29
33 28
PSNR(dB)
PSNR(dB)
PSNR(dB)
32
32 27
30
31 26
30 28 25
Fig. 9. R-D performances for grayscale image coding by using different methods.
1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806281, IEEE
Transactions on Image Processing
1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806281, IEEE
Transactions on Image Processing
10
44 44
30
42 42
29
PSNR(dB)
PSNR(dB)
PSNR(dB)
40 40
28
38 38
27
JPEG coding 36 JPEG coding 36 JPEG coding
IDID−based coding IDID−based coding IDID−based coding
CDTDDC−based coding CDTDDC−based coding CDTDDC−based coding
26 34 34
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35
Bitrate(bpp) Bitrate(bpp) Bitrate(bpp)
House Flowers
35 38
34
36
33
32
34
31
PSNR(dB)
PSNR(dB)
30 32
29
30
28
27
JPEG coding 28 JPEG coding
26 ICDF+ECSQ on all components ICDF+ECSQ on all components
ICDF+ECSQ on Cb & Cr components ICDF+ECSQ on Cb & Cr components
25 26
0 0.2 0.4 0.6 0.8 1 1.2 1.4 0 0.2 0.4 0.6 0.8 1
Bitrate(bpp) Bitrate(bpp)
Fig. 15. Comparison of applying ICDF and ECSQ to different components in color image coding.
are performed on both the luminance component (Y) and methods by performing both ICDF and ECSQ on different
two chrominance components (Cb and Cr) to implement the combinations of luminance and chrominance components. The
compression. The luminance component is always filled with first one is implemented by applying them to both luminance
plenty of textures, while less textures exist in the chrominance component and chrominance components, and the second one
component. An example to compare such a difference has is achieved by just performing them on the chrominance
been illustrated in Fig. 13, where the luminance component components. The comparison results for these two methods are
and two chrominance components of the test image House presented in Fig. 15. Compared with the JPEG baseline cod-
(as shown in Fig. 7) have been presented after the RGB-to- ing, using the second coding strategy offers a consistent coding
YCbCr conversion. Apparently, the chrominance component gain over all bit-rates, while implementing the compression by
is much smoother than the luminance component, which also using the first method only gets the performance improvement
makes the interpolation more effective for the chrominance at the low bit-rate. Based on the above discussions and the
component to compose a full-resolution image from its sub- experimental results shown in Figs. 14 and 15, we propose
sampled counterpart. As a result, the chrominance components to build up the CDTDDC-based coding for color images
are always sub-sampled uniformly before compression and the by applying both ICDF and ECSQ only to the chrominance
image interpolation is then used to reconstruct a full-resolution components after the RGB-to-YCbCr conversion.
chrominance image after compression, leading to the 4:2:0 Although our CDTDDC-based coding is performed on the
chroma format coding. chrominance components to implement the data-size reduc-
Similar to the 4:2:0 chroma format coding, when we apply tion, it is quite different from the traditional 4:2:0 chroma
our CDTDDC-based coding to compress the converted YCbCr format coding. The 4:2:0 chroma format coding gets a reduc-
image, performing it on the chrominance components should tion of data-size in the pixel domain, while our CDTDDC-
be more effective than on the luminance component. This has based method achieves the same size reduction in the trans-
been verified by the experimental results shown in Fig. 14, form domain, offering a new solution to reduce the data-
where the R-D performances of the JPEG coding and the size for chrominance components. Moreover, the CDTDDC-
CDTDDC-based coding for the compression of each single based coding proposes a novel way to reduce the compression
component of test image House are all presented. Moreover, distortion, which potentially leads to a better reconstruction.
IDID [21] is also applied to each chrominance component to Before we make a further comparison of different com-
implement the spatial down-sampling for the compression. It pression methods, we firstly conduct a comparison of using
is also seen from Fig. 14 that using our proposed CDTDDC- different interpolation methods to implement the CDTDDC-
based method to compress chrominance components works based coding for color images. The bicubic, bilinear and Lanc-
more efficiently than using the IDID-based method. zos interpolations are still integrated in the ICDF algorithm
Then, we design two CDTDDC-based color image coding and also used to interpolate chrominance images. With these
1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806281, IEEE
Transactions on Image Processing
11
House Flowers
35 38
34 37
33 36
35
32
34
PSNR(dB)
PSNR(dB)
31
33
30
32
29
31
28 30
With bilinear interpolation With bilinear interpolation
27 With Lanczos interpolation 29 With Lanczos interpolation
With bicubic interpolation With bicubic interpolation
26 28
0.2 0.4 0.6 0.8 1 1.2 0.4 0.5 0.6 0.7 0.8 0.9 1
Bitrate(bpp) Bitrate(bpp)
Fig. 16. Comparison of using different interpolations in our proposed method for color image coding.
interpolation methods, some comparison results obtained by very close if evaluated by the PSNR values, but the latter one
using the CDTDDC-based coding to compress test images produces a smoother edge along the roof of house. Secondly,
House and Flowers (as shown in Fig. 7) have been presented it is seen from Figs. 19-20 that the CDTDDC-based coding
in Fig. 16. The results shown in Fig. 16 demonstrate again offers a better visual quality than the JPEG coding and the
that using the bicubic interpolation in our proposed method IDID-based coding, even than the ACT-based coding, when
gets the best compression performance, making it adopted in it is used to compress image Peppers. Moreover, the ACT-
our proposed CDTDDC-based coding for color images. CDTDDC-based method provides the best visual quality over
Several coding methods are used to compress color images all compared coding methods.
in this work, including our CDTDDC-based coding, the JPEG
baseline coding and the adaptive color-space transform (ACT) C. Discussion about computational complexity
based coding [47]. Meanwhile, IDID [21] is applied to both In our proposed CDTDDC-based coding, the extra com-
chrominance components to form the 4:2:0 format and it putation mainly results from running the FISTA algorithm
accordingly leads to the IDID-based coding method. to implement the ICDF, and it is important to note that
We make a comparison of these methods by conducting this additional computation only occurs at the encoder side.
a number of simulations. Moreover, the adaptive color-space Meanwhile, the ECSQ algorithm also contributes some ad-
transform proposed in [47] is also adopted in the CDTDDC- ditional calculations to implementing the encoding of image
based coding, generating the ACT+CDTDDC-based compres- signals. Moreover, when our proposed method is applied to the
sion for a further comparison. All compared methods are grayscale image coding, the RDO-based mode determination
applied to six color images as shown in Fig. 7. Particularly, occurring at the encoder side also leads to the increment
the 4:2:0 coding format is adopted in the last three methods. of computational complexity. On the other hand, the image
The R-D performances of all compared methods are shown interpolation is needed at the decoder side for the grayscale
in Fig. 17 in which the average PSNR of three color channels image coding, while it does not need any additional operations
is adopted. It is found from Fig. 17 that our CDTDDC-based at the decoder side for the color image coding.
coding method always outperforms the JPEG coding and the Compared with the traditional JPEG coding, the increased
IDID-based coding although it cannot exceed the ACT-based running time spent by using some compared methods is shown
method for most test images. The RGB-to-YCbCr conversion in Tables I-II, for grayscale image coding and color image
is optimally designed in the ACT-based method, which makes coding, respectively. In both coding scenarios, the quantization
it get a better compression performance. After we integrate quality factors (QFs) are all set to 50. Due to just adopting a
such an optimal conversion in our CDTDDC-based coding, new color transform in the traditional JPEG coding, the ACT-
the generated ACT+CDTDDC-based compression achieves the based coding does not cause any additional computations and
highest compression efficiency, and it significantly improves the evaluation for its complexity is not included in this work.
the coding performance for test images Airplane, Peppers Similarly, the complexity evaluation for the ACT+CDTDDC-
and Parrot, as proved by up to 1 dB gain over the ACT- based method is not included either, and one can refer the
based coding. This further verifies the effectiveness of our evaluation on the CDTDDC-based coding.
CDTDDC-based coding scheme. According to the experimental results presented in Tables
In addition, some visual results are presented in Figs. 18-20 I-II, it is found that using our proposed method to compress
to make a subjective comparison between various compression images costs much more time than using the other coding
methods. Firstly, according to the results shown in Fig. 18, methods. In fact, running the FISTA algorithm at the encoder
although the JPEG coding, IDID-based coding and CDTDDC- side contributes near 300% time increment to the grayscale
based coding offer similar objective results, the compressed image coding and near 350% time increment to the color
image portion by using the CDTDDC-based method shows image coding. However, using such a solution in our proposed
less blocking artifacts, especially in the wall of house. On coding will not cause any extra workload at the decoder side.
the other hand, the compression performances of the ACT- Based on the above discussions, we believe that our pro-
based method and the ACT-CDTDDC-based method are also posed CDTDDC-based coding is useful for the practical
1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806281, IEEE
Transactions on Image Processing
12
TABLE I. Time increment (%) for grayscale image coding. to compress all components cannot guarantee a high overall
SRDDS-based ITDDC-based CDTDDC-based compression efficiency. Therefore, when such a coding method
coding coding coding is used to compress color images, it is just performed on both
Lena 68.62 197.82 475.21 chrominance components and the luminance component is still
Man 66.41 211.23 502.07
Goldhill 65.41 199.73 490.54
compressed by the JPEG coding. Compressing color images
Boat 67.26 202.55 491.71 in this way not only guarantees a high coding efficiency,
Barbara 69.59 211.82 493.76 but also makes the CDTDDC-based color image coding fully
Baboon 64.75 207.81 483.91 comparable to the traditional 4:2:0 chroma format coding.
Average 67.01 205.16 489.53
Benefited from the excellent compression of chrominance
TABLE II. Time increment (%) for color image coding. components, our proposed method achieves much more gains
over the state-of-the-art methods, which makes it more suitable
IDID-based coding CDTDDC-based coding for practical applications.
House 4.41 426.08
Airplane 4.69 420.56
Flowers 4.43 424.02
V. C ONCLUDING R EMARKS
Peppers 4.62 434.10
Parrot 4.31 421.36 In this paper, we have proposed a new TDDC-based coding
Butterfly 4.59 423.98
Average 4.51 425.02
method to compress image signals. In our way, each 16 × 16
macro-block is converted into an 8 × 8 small-sized coeffi-
cient block by using our proposed ICDF algorithm, which
application due to its higher compression efficiency, although guarantees a better interpolation for the reconstruction of a
its encoding procedure is complicated. More specifically, we compressed macro-block as well as a lower compression cost.
may use the proposed ICDF and ECSQ algorithms to encode After that, the small-sized coefficient block is compressed
image signals with a powerful server, and just utilize a simple with our proposed ECSQ algorithm to reduce the compression
JPEG-based decoder to conduct the decoding for each client. distortion of some specific pixels collected to implement
This kind of applications has been very popular in today’s the interpolation for the whole macro-block. Based on the
cloud computing. ICDF and ECSQ algorithms, we have proposed a CDTDDC-
based coding to compress both grayscale and color images.
D. Further discussion about the CDTDDC-based coding When our CDTDDC-based coding is used to compress color
It is found that both our proposed grayscale image coding images, it not only offers a new way to reduce the data-
and color image coding have outperformed the state-of-the-art size of the chrominance components but also makes a high
methods. Moreover, applying the CDTDDC-based method to compression efficiency. Experimental results show that using
compress color images works more efficiently than applying our CDTDDC-based coding to compress images has achieved
it to compress grayscale images. In our proposed method, a significant gain over the existing methods, especially for the
we design two different solutions, one for grayscale image compression of color images.
and another for color image, to compress the image signals.
These two solutions are developed based on the corresponding R EFERENCES
characteristics of the grayscale and color images, thus leading
[1] G. K. Wallace, “The JPEG still picture compression standard,” Commun.
to different compression gains. ACM, vol. 34, no. 4, pp. 30-44, 1991.
More specifically, the grayscale image is normally filled [2] Information Technology—JPEG 2000 Image Coding System: Core
with various textures, which makes recovering a full-resolution Coding System, ISO/IEC 15444-1 and ITU-T Rec. T.800, 2000.
[3] B. Zeng and J. J. Fu, “Directional discrete cosine transforms: a new
image from its sub-sampled counterpart quite difficult. Any framework for image coding,” IEEE Trans. Circuits Syst. Video Technol.,
interpolation method, even coupled with an adaptive down- vol. 18, no. 3, pp. 305-313, Mar. 2008.
sampling, still cannot rebuild a high-quality image from a sub- [4] S. Muramatsu, D. D. Han, T. Kobayashi, and H. Kikuchi, “Directional
lapped orthogonal transform: theory and design,” IEEE Trans. on Image
sampled one. Such a limitation also impacts the compression Process., vol. 21, no. 5, pp. 2434-2448, May 2012.
efficiency of our proposed CDTDDC-based coding when used [5] G. Fracastoro, S. M. Fosson, and Enrico Magli, “Steerable discrete
to compress grayscale images. We adopt two coding modes cosine transform,” IEEE Trans. Image Process., vol. 26, no. 1, pp. 303-
314, Jan. 2017.
to improve the coding efficiency for this scenario. Although
[6] E. H. Yang, C. Sun, and J. Meng, “Quantization table design revisited
such an operation makes our proposed grayscale image coding for image/video coding,” IEEE Trans. on Image Process., vol. 23, no.
outperform the state-of-the-art methods, it still cannot work 11, pp. 4799-4811, Nov. 2014.
so efficiently as the proposed CDTDDC-based color image [7] T. Richter, “JPEG on STEROIDS: Common optimization techniques
for JPEG image compression,” in Proc. IEEE Int. Conference on Image
coding does. Process., pp. 61-65, Sept. 2016.
Based on the simulation results shown in Fig. 14, it is found [8] G. Lakhani, “Modifying JPEG binary arithmetic codec for exploiting
using the CDTDDC-based coding to compress chrominance inter/intra-block and DCT coefficient sign redundancies,” IEEE Trans.
on Image Process., vol. 22, no. 4, pp. 1326-1339, Apr. 2013.
components performs much better than using it to compress [9] D. Liu, X. Y. Sun, F. Wu, S. P. Li, and Y. Q. Zhang, “Image compression
luminance component. Compressing the luminance compo- with edge-based inpainting,” IEEE Trans. Circuits and Systems for Video
nent with CDTDDC only achieves the coding gain at the Technology, vol. 17, no. 10, pp. 1273-1287, Oct. 2007.
[10] Z. W. Xiong, X. Y. Sun, and F. Wu, “Block-based image compression
low bit-rate. Moreover, the R-D results presented in Fig. 15 with parameter-assistant inpainting,” IEEE Trans. Image Process., vol.
also demonstrate that applying the CDTDDC-based coding 19, no. 6, pp. 1651-1657, Jun. 2010.
1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806281, IEEE
Transactions on Image Processing
13
34 34
36
33 33
32 32 34
PSNR(dB)
PSNR(dB)
PSNR(dB)
31 31
32
30 30
29 29 30
JPEG coding JPEG coding JPEG coding
28 IDID−based coding 28 IDID−based coding IDID−based coding
ACT−based coding ACT−based coding 28 ACT−based coding
27 CDTDDC−based coding 27 CDTDDC−based coding CDTDDC−based coding
ACT+CDTDDC−based coding ACT+CDTDDC−based coding ACT+CDTDDC−based coding
26 26 26
0.2 0.4 0.6 0.8 1 1.2 0.2 0.4 0.6 0.8 1 1.2 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Bitrate(bpp) Bitrate(bpp) Bitrate(bpp)
Peppers Parrot Butterfly
33 40 38
32
38 36
31
36 34
30
PSNR(dB)
PSNR(dB)
PSNR(dB)
29 34 32
28
32 30
27 JPEG coding JPEG coding JPEG coding
IDID−based coding IDID−based coding IDID−based coding
ACT−based coding 30 ACT−based coding 28 ACT−based coding
26 CDTDDC−based coding CDTDDC−based coding CDTDDC−based coding
ACT+CDTDDC−based coding ACT+CDTDDC−based coding ACT+CDTDDC−based coding
25 28 26
0.2 0.4 0.6 0.8 1 1.2 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.2 0.4 0.6 0.8 1 1.2
Bitrate(bpp) Bitrate(bpp) Bitrate(bpp)
Fig. 17. R-D performances for color image coding by using different methods.
1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806281, IEEE
Transactions on Image Processing
14
[11] W. Hu, G. Cheung, A. Ortega, and O. C. Au, “Multiresolution graph [35] N. Ahmed, T. Natarajan, and K. R. Rao, “Discrete cosine transform,”
Fourier transform for compression of piecewise smooth images,” IEEE IEEE Trans. Comput., vol. C-23, pp. 90-93, Jan. 1974.
Trans. Image Process., vol. 24, no. 1, pp. 419-433, Jan. 2015. [36] A. K. Jain, “A sinusoidal family of unitary transforms,” IEEE Trans.
[12] G. Fracastoro, F. Verdoja, M. Grangetto, and E. Magli, “Superpixel- Pattern Anal. Mach. Intell., vol. 1, no. 4, pp. 356-365, Oct. 1979.
driven graph transform for image compression,” in Proc. IEEE Int. [37] E. Y. Lam and J. W. Goodman, “A mathematical analysis of the DCT
Conference on Image Process., Sept. 2015. coefficient distributions for images,” IEEE Trans. Image Process., vol.
[13] Y. P. Sun, X. M. Tao, Y. Li, L. H. Dong, and J. H. Lu, “HEMS: 9, no. 10, pp. 1661-1666, Oct. 2000.
Hierarchical exemplar-based matching-synthesis for object-aware image [38] T. H. Thai, R. Cogranne, and F. Retraint, “Statistical model of quantized
reconstruction,” IEEE Trans. Multimedia, vol. 18, no. 2, pp. 171-181, DCT coefficients: application in the steganalysis of Jsteg algorithm,”
Feb. 2016. IEEE Trans. Image Process., vol. 23, no. 5, pp. 1980-1993, May 2014.
[14] J. Xiao, R. M. Hu, L. Liao, Y. Chen, Z. Y. Wang, and Z. X. Xiong, [39] J. Q. Yang, G. P. Zhu, and Y. Q. Shi,“Analyzing the effect of JPEG
“Knowledge-based coding of objects for multisource surveillance video compression on local variance of image intensity,”IEEE Trans. Image
data,” IEEE Trans. Multimedia, vol. 18, no. 9, pp. 1691-1706, Sept. Process., vol. 25, no. 6, pp. 2647-2656, Jun. 2016.
2016. [40] T. Wiegand, G. J. Sullivan, G. Bjøntegaard, and A. Luthra, “Overview
[15] X. F. Zhang, S. Q. Wang, K. Gu, W. S. Lin, S. W. Ma, and W. of the H.264/AVC video coding standard,” IEEE Trans. on Circuits and
Gao, “Just-noticeable difference-based perceptual optimization for JPEG Systems for Video Technology, vol. 13, no. 7, pp. 560-576, Jul. 2003.
compression,” IEEE Signal Process. Letters, vol. 24, no. 1, pp. 96-100, [41] G. J. Sullivan, J.-R. Ohm, W.-J. Han, and T. Wiegand, “Overview of the
2017. high efficiency video coding (HEVC) standard,” IEEE Trans. on Circuits
[16] X. D. Song, X. L. Peng, J. Z. Xu, G. M. Shi, and F. Wu “Distributed and Systems for Video Technology, vol. 22, no. 12, pp. 1649-1668, Dec.
compressive sensing for cloud-based wireless image transmission,”IEEE 2012.
Trans. Multimedia, vol. 19, no. 6, pp. 1351-1364, Jun. 2017. [42] I. Daubechies, M. Defrise, and C. De Mol, “An iterative thresholding al-
[17] B. Zeng and A. N. Venetsanopoulos “A JPEG-based interpolative image gorithm for linear inverse problems with a sparsity constraint,” Commun.
coding scheme,” in Proc. IEEE Int. Conference on Acoustics, Speech, Pure Appl. Math., vol. 57, no. 11, pp. 1413-1457, 2004.
and Signal Processing, Apr. 1993. [43] J. M. Bioucas-Dias and M. A. Figueiredo, “A new TwIST: two-step
[18] W. Lin and L. Dong, “Adaptive downsampling to improve image iterative shrinkage/thresholding algorithms for image restoration,” IEEE
compression at low bit rates,” IEEE Trans. Image Process., vol. 15, Trans. on Image Process., vol. 16, pp. 2992-3004, Dec. 2007.
no. 9, pp. 2513-2521, Sept. 2006. [44] A. Beck and M. Teboulle, “A fast iterative shrinkage-thresholding
[19] T. Frajka and K. Zeger, “Downsampling dependent upsampling of algorithm for linear inverse problems,” SIAM J. Imag. Sci., vol. 2, no.
images,” Signal Process. Image Commun., vol. 19, no. 3, pp. 257-265, 1, pp. 183-202, 2009.
Mar. 2004. [45] S. Y. Zhu, J. L. Yu, M. Y. Li, C. Chen, L. Y. Zeng, and B.
[20] X. L. Wu, X. J. Zhang, and X. H. Wang, “Low bit-rate image Zeng, “Interpolation-directed transform domain downward conversion
compression via adaptive down-sampling and constrained least squares for block-based image compression,” in Visual Communication and
upconversion,” IEEE Trans. Image Process., vol. 18, no. 3, pp. 552-561, Image Procssing, Nov. 2016.
Nov. 2009. [46] ITU-R, “Parameter values for high definition television systems for
[21] Y. B. Zhang, D. B. Zhao, J. Zhang, R. Q. Xiong, and W. Gao, production and international programme exchange,” ITU-R Rec. BT.709-
“Interpolation dependent image downsampling,” IEEE Trans. Image 5, Apr. 2002.
Process., vol. 20, no. 11, pp. 3291-3296, Nov. 2011. [47] L. Zhang, X. Y. Xiu, J. L. Chen, M. Karczewicz, Y. W. He, Y. Ye, J. Z.
Xu, J. Sole, and W.-S. Kim, “Adaptive color-space transform in HEVC
[22] L. Ma, S. N. Li, and K. N. Ngan, “Perceptual image compression via
screen content coding,” IEEE Journal on Emerging and Selected Topics
adaptive block-based super-resolution directed down-sampling,” in Proc.
in Circuits and Systems, vol. 6, no. 4, pp. 446-459, Dec. 2016.
IEEE Int. Symposium on Circuits and Systems, May 2011.
[23] Y. B. Zhang, X. Y. Ji, H. Q. Wang, and Qionghai Dai, “Stereo
interleaving video coding with content adaptive image subsampling,”
IEEE Trans. on Circuits and Systems for Video Technology, vol. 23, no.
7, pp. 1097-1108, Jul. 2013.
[24] Z. Wang and S. Simon, “Low complexity pixel domain perceptual Shuyuan Zhu (S’08-A’09-M’13) received the Ph.D.
image compression via adaptive down-sampling,” in Data Compression degree from the Hong Kong University of Science
Conference, Apr. 2016. and Technology (HKUST), Hong Kong, in 2010.
[25] A. M. Bruckstein, M. Elad, R. Kimmel, “Down-scaling for better From 2010 to 2012, he worked at HKUST and Hong
transform compression,” IEEE Trans. on Image Process., vol. 12, no. Kong Applied Science and Technology Research
9, pp. 1132-1144, Sept. 2003. Institute Company Limited (ASTRI), respectively.
[26] G. Georgis, G.Lentaris, and D. Reisis, “Reduced complexity superreso- In 2013, he joined University of Electronic Science
lution for low bit-rate video compression,”IEEE Trans. on Circuits and and Technology of China (UESTC) and is currently
Systems for Video Technology, vol. 26, no. 2, pp. 332-345, Feb. 2016. an Associate Professor with School of Information
[27] R. Dugad and N. Ahuja,“A fast scheme for image size change in the and Communication Engineering. Dr. Zhu’s research
compressed domain,” IEEE Trans. Circuits Syst. Video Technol., vol. 11, interests include image/video compression, image
no. 4, pp. 461-474, Apr. 2001. processing and compressive sensing. He has over 60 research publications
[28] Z. Y. Wu, H. Y. Yu, and C. W. Chen, “A new hybrid DCT-Wiener-based and received the Top 10% paper award at IEEE ICIP-2014 and the Top 10%
interpolation scheme for video intra frame up-sampling,” IEEE Signal paper award at VCIP-2016. Dr. Zhu was the special session chair of image
Processing Letters, vol. 17, no. 10, pp. 827-830, Oct. 2010. super-resolution at IEEE DSP-2015. He served as the committee member for
[29] K.-W. Hung and W.-C. Siu, “Hybrid DCT-Wiener-Based Interpolation IEEE ICME-2014, VCIP-2016, and PCM-2017. He is a member of IEEE
via learnt Wiener filter,” in Proc. IEEE Int. Conference on Acoustics, CAS society.
Speech and Signal Processing, May 2013.
[30] K.-W. Hung and W.-C. Siu, “Novel DCT-based image up-sampling using
learning-based adaptive k-NN MMSE estimation,” IEEE Trans. Circuits
and Systems for Video Technology, vol. 24, no. 12, pp. 2018-2033, Dec.
2014.
Zhiying He received his BEng degree in electronic
[31] X. Li and M. T. Orchard, “New edge-directed interpolation,” IEEE engineering from University of Electronic Science
Trans. on Image Process., vol. 10, no. 10, pp. 1521-1527, Oct. 2001. and Technology of China (UESTC), Chengdu, Chi-
[32] X. J. Zhang and X. Wu, “Image interpolation by adaptive 2-D autore- na, in 2015. His research interest is image/video
gressive modeling and soft-decision estimation,” IEEE Trans. Image compression. He is currently pursuing the MEng
Process., vol. 17, no. 6, pp. 887-896, Jun. 2008. degree in the School of Information and Commu-
[33] X. M. Liu, D. B. Zhao, R. Q. Xiong, S. W. Ma, W. Gao, and H. F. nication Engineering at UESTC.
Sun, “Image interpolation via regularized local linear regression,” IEEE
Trans. on Image Process., vol. 20, no. 12, pp. 3455-3469, Dec. 2011.
[34] C. E. Duchon, “Lanczos filtering in one and two dimensions,” Journal
of Applied Meteorology, vol. 18, no. 8, pp. 1016-1022, Aug. 1979.
1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TIP.2018.2806281, IEEE
Transactions on Image Processing
15
1057-7149 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.