A twostage algorithm for the early detection of zeroquantized discrete cosine transform coefficients in High Efficiency Video Coding
 WeiGang Chen^{1}Email authorView ORCID ID profile,
 Xun Wang^{1} and
 Yan Tian^{1}
https://doi.org/10.1186/s1364001702052
© The Author(s) 2017
Received: 28 December 2016
Accepted: 7 August 2017
Published: 16 August 2017
Abstract
For High Efficiency Video Coding (HEVC) using blockbased prediction, the discrete cosine transform (DCT), and quantization, a large number of DCT coefficients in a transform block (TB) are commonly found to be quantized to zero. A twodimensional transform in HEVC is usually implemented by first applying a butterflybased onedimensional (1D) DCT to each row of the residual block, followed by a 1D DCT to each column. Accordingly, we propose a twostage method for the early detection of zeroquantized DCT coefficients in this paper. In the first stage, a distributionbased method, which uses the sum of absolute differences as the threshold criterion and compares it to threshold values obtained for all columns, is employed to detect zeroquantized columns in the pixel domain prior to conducting the actual DCT. After conducting the 1D row DCT, the second detection stage, which uses the intermediate matrix resulting from the row transform as the input, is applied in the transform domain only to those columns predicted to contain nonzero coefficients. As an orthogonal transform, the 1D row DCT has a tendency to pack a large fraction of the signal energy into a relatively few coefficients. Therefore, the second stage of our algorithm is more effective than methods that conduct detection in the pixel domain, particularly for TBs greater than 8×8. Experimental results demonstrate that an HEVC encoder with our proposed twostage algorithm can dramatically reduce unnecessary 1D row and column DCT operations compared with a standard encoder and therefore exhibits applicability to practical HEVC encoder implementation.
Keywords
High Efficiency Video Coding Discrete cosine transform Quantization Zeroquantized coefficients1 Introduction
High Efficiency Video Coding (HEVC) design follows the classic blockbased hybrid video coding approach [1]. For HEVC using blockbased intra or interprediction, the discrete cosine transform (DCT), and quantization, a substantial number of DCT coefficients of the residual block are commonly found to be quantized to zero, particularly when the quantization parameter (QP) is large. Therefore, considerable computation can be avoided if the transform blocks (TBs) containing DCT coefficients that are uniformly zero (i.e., allzero blocks (AZBs)) and TBs where all the DCT coefficients in a specified subblock are zeros (i.e., partialzero blocks (PZBs)) [2] can be detected prior to conducting the DCT operations and quantization.
The problems involved with the early detection of AZBs and zero subblocks for 4×4 DCT in H.264/AVC [3–6] and 8×8 DCT in MPEG2/4 and H.263 [2, 7–10] have been extensively studied. However, the methods developed for these standards are not well suited for HEVC. The main reason is that HEVC supports larger TB sizes (i.e., 16×16 and 32×32) than those employed in prior standards, and the detection efficiency will be degraded when applying methods developed for 8×8 or 4×4 blocks to larger TBs. More recently, various schemes have been developed for detecting variable sized AZBs in HEVC [11–14]. These methods typically predict AZBs by setting the thresholds according to the sum of absolute differences (SAD) [11, 13, 14], sum of absolute transformed differences (SATD) [13], and sum of squared differences (SSD) [13]. However, we have observed in experiments that the number of AZBs relative to the total number of TBs decreases with increasing TB size, as would be expected. Therefore, even though AZB detection algorithms are efficient, their capacity to reduce the computational burden is limited.
In HEVC, the twodimensional (2D) forward DCT of a TB is implemented by first applying a butterflybased onedimensional (1D) DCT [15] to each row of the residual block, followed by applying a 1D DCT to each column. As an orthogonal transform, 1D DCT has a tendency to pack a large fraction of the signal energy into a relatively few elements of the intermediate matrix. In this paper, we propose a twostage method that capitalizes on this property for the early detection of zeroquantized columns in HEVC. In the first stage, a distributionbased method, which uses the SAD as the threshold criterion and compares it to threshold values obtained for all columns, is employed to detect zeroquantized columns in the pixel domain before conducting the actual DCT. After conducting the 1D row DCT, the second detection stage, which uses the intermediate matrix resulting from the row transform as the input, is applied in the transform domain only to those columns predicted to contain nonzero coefficients. Therefore, in contrast to most existing techniques, the proposed algorithm performs detection not only in the pixel domain but also in the transform domain after conducting 1D row DCT. Furthermore, the proposed method regards AZB detection as a special case of zeroquantized column detection, where an AZB is a TB whose columns are all zeroquantized columns. Experimental results demonstrate that the proposed twostage algorithm can detect zeroquantized columns more effectively than existing methods, particularly for TBs greater than 8×8.
The remainder of this paper is organized as follows. Section 2 discusses previous studies related to the early detection of AZBs and PZBs for 4×4 and 8×8 DCT. Section 3 presents our proposed twostage method for detecting zeroquantized columns in HEVC. Experimental results are given in Section 4. We conclude our work in Section 5.
2 Related work
Existing methods for the early detection of zeroquantized DCT coefficients typically aim at detecting the entire residual block as an AZB or as a PZB. In addition, most existing detection methods are thresholdbased, and the SAD of the motion compensation block, which can be obtained during motion estimation, has been the most widely used threshold selection criterion. Some variants of the SAD have also been employed. Examples include rowbased SAD [2] and the maximum value of the sum of positive residuals and the absolute value of negative residuals [10]. Methods for deriving the thresholds can be roughly classified into three categories: upper boundbased [7, 9, 10, 16], distributionbased [8, 17], and hybrid methods that employ multiple thresholds derived using methods in the first two categories [5, 18].
where α is determined by a theoretical analysis of the properties of the DCT. Among these methods, Zhou et al. [7] provided a larger value α (i.e., \(\frac {1}{4}\) for an 8×8 DCT), which provided a more stringent condition for detecting AZBs. Sousa [9] refined the model proposed by Zhou et al. and derived a more precise α (i.e., \(\alpha = \frac {1}{4}\cos ^{2} \left ({\frac {\pi }{{16}}} \right)\)). Ji et al. [2] refined Sousa’s model using a combination of SAD and rowbased SAD as the threshold criterion. Their method not only derived a more precise sufficient condition for detecting AZBs than Sousa’s but also provided conditions for the early detection of PZBs with 34 and 16 zeroquantized coefficients in an 8×8 block. After the H.264 standard was launched to address a new generation of multimedia applications, the implications of the employment by the standard of an integer 4×4 DCT and scaling multiplication to avoid division for quantization were studied [3, 4], and an upper boundbased method was extended to AZB detection in H.264/AVC. The advantage of upper boundbased methods is that they do not incur a false positive error. However, these methods determine the thresholds by approximating the individual elements of a transform matrix by a theoretical maximum value at different frequency positions. The deviation between the upper bound and the actual maximum value of the DCT coefficients increases as the TB size increases. Therefore, extending these methods to TBs greater than 8×8 reduces the detection ratio. Taking a 16×16 TB, for example, the SAD value is on average four times the SAD value of any of its four quadrants (i.e., 8×8 blocks). Using the method proposed in [9], the sufficient condition for AZB detection is about twice that for an 8×8 block.
Distributionbased methods employ Gaussian or Laplacian distributions to model the residual pixels and derive multiple thresholds based on a theoretical analysis of the relationship between the distribution of the residual pixels and the distribution of the DCT coefficients [8, 17]. These methods derive AZB and PZB thresholds under various assumptions, e.g., the residual pixels in a motioncompensated frame follow a Laplacian distribution [8] or a Gaussian distribution [17], and the distribution has a zero mean and a separable covariance. Furthermore, distributionbased methods discriminate DCT coefficients using a probability framework based on threesigma limits, where sigma (σ) denotes the standard deviation of the distribution. Based on the assumptions of the method, the residual pixels follow a zeromean Laplacian or Gaussian distribution and the distribution of the DCT coefficients at a given frequency position will fall within the range (−3σ,3σ) with a very high probability. Because these assumptions may fail, distributionbased methods may induce false positive detection and result in video quality degradation.
where C _{ N } is an N×N core transform matrix. The elements of C _{ N } are derived by approximating the scaled DCT basis functions. Details regarding the core transform matrix for different transform sizes have been presented in a previous study [19]. The 2D forward DCT in (2) is implemented by first applying a butterflybased 1D DCT [15] to each row of the residual block, that is, Y=C _{ N } X ^{T}, followed by the 1D column DCT F=C _{ N } Y ^{T}. In such an implementation, even if only a few coefficients are predicted as nonzero, the 1D row and column transforms cannot be skipped. As a result, the capacity of these methods to reduce the computational burden are limited (a detailed illustration of this phenomenon and its cause can be found in [17]). Moreover, HEVC supports larger TB sizes (up to 32×32) than those supported in previous standards.
The above analysis of past work indicates that it is necessary to develop a method of AZB and PZB detection that conforms well to the separated 1D row and column transform structure and larger TB sizes employed in HEVC.
3 The proposed method
In the first stage, the distributionbased methods presented in [8, 20], which were developed for an 8×8 DCT, are extended to zeroquantized column detection for variable TB sizes ranging from 4×4 to 32×32. After conducting the firststage detection process, a butterflybased 1D DCT is applied to each row of the residual block, and an intermediate matrix is obtained, which serves as the input of the second stage of the detection process. The second stage seeks to detect zeroquantized columns in the transform domain and applies 1D DCT only to those columns predicted to contain nonzero coefficients.
3.1 Firststage detection
In blockbased video coding, the pixels of an N×N residual block are often statistically modeled by a Laplacian distribution with a zero mean, and as having a separable covariance [20]. Because the pixels within a residual block are often assumed to be identically distributed, the weighted summation in (2) for a given F(u,v) can be approximated as having a Gaussian distribution according to the central limit theorem [21].
Here, 0<ρ<1 is the onestep correlation coefficient of the residual pixels and is typically assumed as an a priori value [8, 20].
Thus, the distribution of the DCT coefficients can be directly estimated without actually applying the DCT.
Example thresholds obtained for a 16×16 block (β=3,ρ=0.6)
i  0  1  3  4  6  8  10  11 

TH _{ i }/qStep  4.84  6.13  10.68  14.54  24.76  36.82  48.88  54.33 
3.2 Secondstage detection
3.3 Summary of the algorithm
We summarize the entire twostage scheme as follows, where steps 1 and 2 represent the first stage and steps 3 and 4 represent the second stage.
4 Experimental results and discussion
To evaluate the performance of the proposed twostage algorithm, the algorithm was implemented on the HEVC Test Model (HM) 12.0 reference software [23]. However, we must first provide some commentary regarding the implementation of the algorithm on this platform. First, because only four 1D column transforms are employed for a 4×4 TB and the secondstage detection requires additional computations for S _{ v } and f _{0,v }, the performance gain of the second stage is limited for 4×4 TBs. In practice, the second stage is bypassed under this condition, and experimental results for 4×4 TBs are therefore not presented. Second, we employ a FLAG vector, each element of which indicates whether or not a column in the coefficient matrix is a zeroquantized column. The 1D DCT operation in the HM 12.0 software (i.e., partialButterflyXX()) was slightly modified to skip the 1D DCT of a column if it is labeled by the FLAG vector as a zeroquantized column. Third, HEVC makes use of an integer DCT. The core transform matrices are finite approximations of the original realvalued DCT matrices, and all the matrix elements are scaled by a factor of 2^{(6+M/2)}, where M= log2N. After application of the 1D row DCT, elements in Y are right shifted by M−9+B, where B is the bit depth of the video [19]. Therefore, f _{0,v } and S _{ v } are scaled by a factor of 2^{−(15−B−M/2)}.

Maximum and minimum coding unit (CU) sizes were respectively 64×64 and 8×8.

Maximum and minimum transform unit (TU) sizes were respectively 32×32 and 4×4, and maximum TU depth was 3. That is, for 64×64 and 32×32 luma CUs, the supported TU sizes were 32×32, 16×16, 8×8, and 4×4. For 16×16 and 8×8 CUs, the TU tree structure had its root at the CU level, and the minimum TU size was 4×4.

Fast motion estimation was enabled, and the search range was 64 in the horizontal and vertical directions.

The QP value was set to 22, 27, 32, or 37.
Three Class B (BasketballDrive, BQTerrace, Parkscene), four Class C (BasketballDrill, BQMall, Keiba, PartyScene), three Class D (BasketballPass, BQSquare, BlowingBubbles), and three Class E (Fourpeople, KristenAndSara, Stockholm) image sequences were tested. For each sequence, 200 frames were coded. Simulations were run on a personal computer with an Intel Core i54430 CPU and 8GB RAM. The operating system was Microsoft Windows 7 64bit Enterprise edition.
Resulting Δ Z (%) and Δ T (%) values for the tested image sequences using the proposed twostage method
QP  Stage  Class B  Class C  Class D  Class E  

N=8  16  32  N=8  16  32  N=8  16  32  N=8  16  32  
22  Stage 1  0.45  0.71  0  4.88  6.88  3.57  4.78  8.13  4.22  24.9  17.4  6.92 
Stage 2  23.3  7.24  5.72  22.8  12.1  7.98  23.1  12.6  8.12  15.8  15.6  7.61  
Δ Z _{Total}  23.75  7.95  5.72  27.68  18.98  11.55  27.88  20.73  12.34  40.7  34.0  14.53  
Δ T  −12.3  −11.6  −13.5  −14.3  
27  Stage 1  13.3  12.8  0.15  18.2  17.1  11.3  17.7  18.6  12.5  17.6  30.6  17.4 
Stage 2  27.6  18.5  9.5  23.6  18.1  14.4  24.1  20.4  16.7  23.1  26.3  15.6  
Δ Z _{Total}  40.9  31.3  9.65  41.8  35.2  25.7  41.8  39.0  29.2  40.7  56.9  34.0  
Δ T  −14.0  −13.8  −14.3  −17.5  
32  Stage 1  27.2  26.9  15.3  27.1  26.7  23.1  28.1  27.2  23.1  28.4  50.9  24.9 
Stage 2  22.6  21.6  20.2  21.7  17.1  15.3  23.4  19.9  17.0  26.1  20.6  15.8  
Δ Z _{Total}  49.8  48.5  35.5  48.8  43.8  38.4  51.5  47.1  40.1  54.5  71.5  40.7  
Δ T  −16.6  −15.4  −17.0  −18.8  
37  Stage 1  33.8  31.1  27.2  38.8  32.2  28.3  39.9  33.1  28.5  47.8  62.4  31.8 
Stage 2  16.8  15.7  17.1  15.6  15.3  15.5  15.7  16.1  16.4  15.8  18.2  16.1  
Δ Z _{Total}  50.6  46.8  44.3  54.4  47.5  43.8  55.6  49.2  44.9  63.6  80.6  47.9  
Δ T  −18.9  −16.5  −19.1  −20.7 
Resulting η (%) values of the proposed twostep method for a subset of the tested image sequences
Class  Sequence name  QP = 22  27  32  37  

N=8  16  32  N=8  16  32  N=8  16  32  N=8  16  32  
B  BasketBallDrive  58.2  44.5  36.7  64.9  52.1  40.4  87.8  81.3  76.5  98.5  97.9  93.2 
BQTerrace  57.1  46.6  50.1  70.5  68.0  53.2  89.6  84.1  80.2  98.1  98.4  93.9  
ParkScene  59.4  48.8  49.2  71.1  67.3  48.9  89.1  82.6  79.1  98.2  96.7  92.1  
C  BasketballDrill  90.1  73.1  37.6  90.4  83.5  62.8  94.3  87.2  81.7  98.7  91.9  85.1 
BQMall  91.5  72.8  39.0  91.8  84.2  62.5  94.5  88.1  82.2  98.9  92.2  86.9  
Keiba  88.7  50.3  48.1  86.9  75.1  51.9  90.1  79.5  71.2  96.8  95.3  90.6  
PartyScene  91.8  71.1  41.7  92.1  83.9  61.8  94.6  88.0  82.5  98.1  92.4  87.1  
D  BasketballPass  60.1  46.4  37.6  65.2  52.0  41.5  87.9  82.1  79.2  98.7  95.9  94.2 
BQSquare  87.6  72.4  35.5  92.7  85.1  75.9  95.9  89.2  96.1  97.2  93.8  98.1  
BlowingBubbles  88.7  74.3  48.8  92.1  86.6  76.1  95.2  90.0  95.2  97.4  94.1  96.5  
E  FourPeople  87.6  71.0  30.5  92.3  84.0  74.9  95.5  89.1  96.0  98.2  93.4  98.6 
KristenAndSara  88.5  70.8  30.9  91.7  85.5  74.2  96.1  88.7  96.2  98.7  93.0  98.2  
Stockholm  63.1  59.3  33.8  56.1  46.9  41.4  72.4  69.2  48.4  93.8  92.4  71.2 
BDRate increase of the proposed twostep method for a subset of test image sequences, β=3.0
Class  Sequence name  Y  Cb  Cr  Δ T _{QP=22}  Δ T _{QP=37} 

B  BasketBall Drive  0.41  0.58  0.70  −11.6  −18.4 
BQTerrace  0.33  0.47  0.60  −12.6  −19.2  
ParkScene  0.27  0.34  0.54  −12.7  −19.1  
C  BasketballDrill  0.23  0.52  0.46  −11.4  −16.3 
BQMall  0.43  0.46  0.55  −11.9  −16.7  
Keiba  0.29  0.62  0.63  −11.0  −16.6  
PartyScene  0.19  0.41  0.33  −12.1  −16.4  
D  BasketballPass  0.25  0.46  0.41  −12.7  −16.6 
BQSquare  0.17  0.36  0.31  −14.0  −17.1  
BlowingBubbles  0.14  0.32  0.26  −13.8  −17.3  
E  FourPeople  0.44  0.6  0.71  −14.5  −21.2 
KristenAndSara  0.12  0.34  0.31  −14.6  −21.0  
Stockholm  0.25  0.55  0.72  −13.8  −19.9 
BDRate increase of the proposed twostep method for a subset of test image sequences, β=3.5
Class  Sequence name  Y  Cb  Cr  Δ T _{QP=22}  Δ T _{QP=37} 

B  BasketBall Drive  0.06  0.09  0.10  −10.5  −16.6 
BQTerrace  0.05  0.08  0.10  −11.1  −17.2  
ParkScene  0.04  0.06  0.08  −10.9  −16.3  
C  BasketballDrill  0.03  0.07  0.07  −9.9  −14.3 
BQMall  0.06  0.07  0.08  −10.2  −15.1  
Keiba  0.05  0.08  0.09  −9.8  −14.9  
PartyScene  0.02  0.06  0.05  −10.5  −14.0  
D  BasketballPass  0.04  0.07  0.07  −11.0  −14.5 
BQSquare  0.01  0.06  0.05  −12.1  −15.0  
BlowingBubbles  0.01  0.05  0.04  −11.9  −15.3  
E  FourPeople  0.06  0.06  0.10  −12.9  −19.0 
KristenAndSara  0.00  0.04  0.05  −13.0  −19.0  
Stockholm  0.02  0.05  0.09  −12.2  −17.5 
Resulting Δ Z (%) values when employing the method proposed by Sousa [9] in the first stage
QP  Stage  Class B  Class C  Class D  Class E  

N=8  16  32  N=8  16  32  N=8  16  32  N=8  16  32  
22  Stage 1  0.4  0  0  1.5  0.3  0  1.9  0.1  0  4.8  0.01  0 
Stage 2  23.3  7.13  5.72  23.1  14.6  9.24  23.9  14.8  10.1  31.5  32.6  12.2  
Δ Z _{Total}  23.7  7.13  5.72  24.6  14.9  9.24  25.8  14.9  10.1  36.3  32.61  12.2  
27  Stage 1  5.6  0  0  5.3  0.4  0  5.7  0.6  0  21.6  0.5  0 
Stage 2  32.8  30.3  9.5  31.9  32.4  20.4  32.1  33.7  20.8  23.5  42.7  31.8  
Δ Z _{Total}  38.4  30.3  9.5  37.2  32.8  20.4  37.8  34.3  20.8  45.1  43.2  31.8  
32  Stage 1  19.5  1.6  0  17.7  1.3  0  19.1  1.2  0  27.7  17.6  0 
Stage 2  27.7  42.8  35.3  30.0  38.6  28.7  32.0  37.7  32.9  26.1  46.0  40.3  
Δ Z _{Total}  47.2  44.4  35.3  47.7  39.9  28.7  51.1  38.9  32.9  53.8  63.6  40.3  
37  Stage 1  32.4  5.8  0  31.8  5.4  0  39.1  14.5  0  47.0  22.8  0.2 
Stage 2  17.0  39.3  42.2  20.3  40.4  35.2  17.8  44.7  40.2  12.4  45.2  44.1  
Δ Z _{Total}  49.4  45.1  42.2  52.1  45.8  35.2  56.9  59.2  40.2  59.4  68.0  44.3 
Resulting Δ Z (%) values when employing the method proposed by Xie et al. [16] in the first stage
QP  Stage  Class B  Class C  Class D  Class E  

N=8  16  32  N=8  16  32  N=8  16  32  N=8  16  32  
22  Stage 1  0  0  0  0  0  0  0  0  0  0  0  0 
Stage 2  23.4  7.13  5.72  23.7  14.7  9.24  24.1  15.1  9.21  32.7  32.6  12.2  
Δ Z _{Total}  23.4  7.13  5.72  23.7  14.7  9.24  24.1  15.1  9.21  32.7  32.6  12.2  
27  Stage 1  0  0  0  0.1  0  0  0.1  0  0  0  0  0 
Stage 2  37.8  30.3  9.5  36.8  32.5  20.4  37.4  33.1  22.5  43.5  42.8  31.8  
Δ Z _{Total}  37.8  30.3  9.5  36.9  32.5  20.4  37.5  33.1  22.5  43.5  42.8  31.8  
32  Stage 1  0  0  0  0.8  0  0  0.9  0  0  1.0  0  0 
Stage 2  45.5  43.8  35.3  43.3  38.6  28.7  44.5  38.9  31.3  47.4  56.3  40.3  
Δ Z _{Total}  45.5  43.8  35.3  44.1  38.6  28.7  45.4  38.9  31.3  48.4  56.3  40.3  
37  Stage 1  1.5  0.1  0  5.6  1.0  0  5.8  0.2  0  6.8  0  0 
Stage 2  47.3  43.2  42.2  45.2  42.2  35.2  45.1  44.1  38.9  45.5  51.6  44.2  
Δ Z _{Total}  48.8  43.3  42.2  50.8  43.2  35.2  50.9  44.3  38.9  52.3  51.6  44.2 
5 Conclusions
A twostage method for the early detection of zeroquantized DCT coefficients in HEVC was proposed in this paper. In the first stage, a distributionbased method is employed in the pixel domain to detect zeroquantized columns prior to conducting 1D row DCT operations. Using the intermediate matrix resulting from the row transform as the input, the second stage performs zeroquantized column detection in the transform domain only for the nonzero columns determined in the first stage. As an orthogonal transform, the 1D DCT has a tendency to distribute a large fraction of the signal energy into a relatively few coefficients. Therefore, the second stage of our algorithm is more effective than existing methods that conduct PZB detection in the pixel domain, particularly for TBs greater than 8×8.
Declarations
Acknowledgements
The authors would like to thank the anonymous reviewers for their helpful comments, and we thank LetPub (www.letpub.com) for the linguistic assistance during the preparation of this manuscript.
Funding
This work was supported in part by the National Natural Science Foundation of China (Grant Nos. 61672460, 61602407, and U1609215) and the Natural Science Foundation of Zhejiang Province (Grant No. LY14F020001).
Authors’ contributions
WC carried out the main part of this manuscript. XW is a supervisor of this research. YT has assisted in the experimental part of the work. All authors read and approved the final manuscript.
Competing interests
The authors declare that they have no competing interests.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License(http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
Authors’ Affiliations
References
 GJ Sullivan, JR Ohm, WJ Han, T Wiegand, Overview of the High Efficiency Video Coding (HEVC) standard. IEEE Trans. Circ. Syst. Video Technol. 22(12), 1649–1668 (2012).View ArticleGoogle Scholar
 X Ji, S Kwong, D Zhao, H Wang, CCJ Kuo, Q Dai, Early determination of zeroquantized 8×8 DCT coefficients. IEEE Trans. Circ. Syst. Video Technol. 19(12), 1755–1765 (2009).View ArticleGoogle Scholar
 YH Moon, GY Kim, JH Kim, An improved early detection algorithm for allzero blocks in H.264 video encoding. IEEE Trans. Circ. Syst. Video Technol.15(8), 1053–1057 (2005).View ArticleGoogle Scholar
 H Wang, S Kwong, CW Kok, Efficient prediction algorithm of integer DCT coefficients for H.264/AVC optimization. IEEE Trans. Circ. Syst. Video Technol.16(4), 547–552 (2006).View ArticleGoogle Scholar
 H Wang, S Kwong, Hybrid model to detect zero quantized DCT coefficients in H.264. IEEE Trans. Multimed. 9(4), 728–735 (2007).View ArticleGoogle Scholar
 M Zhang, T Zhou, W Wang, Adaptive method for early detecting zero quantized DCT coefficients in H.264/AVC video encoding. IEEE Trans. Circ. Syst. Video Technol. 19(1), 103–107 (2009).View ArticleGoogle Scholar
 X Zhou, Z Yu, S Yu, Method for detecting allzero DCT coefficients ahead of discrete cosine transformation and quantisation. Electron. Lett. 34(19), 1839–1840 (1998).View ArticleGoogle Scholar
 IM Pao, MT Sun, Modeling DCT coefficients for fast video encoding. IEEE Trans. Circ. Syst. Video Technol. 9(4), 608–616 (1999).View ArticleGoogle Scholar
 LA Sousa, General method for eliminating redundant computations in video coding. Electron. Lett. 36(4), 306–307 (2000).View ArticleGoogle Scholar
 S Jun, S Yu, Efficient method for early detection of allzero DCT coefficients. Electron. Lett. 37(3), 160–161 (2001).View ArticleGoogle Scholar
 H Wang, H Du, W Lin, S Kwong, OC Au, J Wu, Z Wei, Early detection of allzero 4×4 blocks in High Efficiency Video Coding. J. Vis. Commun. Image R. 25(7), 1784–1790 (2014).View ArticleGoogle Scholar
 K Lee, HJ Lee, J Kim, Y Choi, A novel algorithm for zero block detection in High Efficiency Video Coding. IEEE J. Sel. Topics Signal Process. 7(6), 1124–1134 (2013).View ArticleGoogle Scholar
 J Su, Y Huang, L Sun, S Sakaiday, T Ikenaga, in Asia Pacific Signal and Information Processing Association—Annual Summit and Conference (APSIPA ASC 2011), Online Proceeding. Low complexity quadtree based all zero block detection algorithm for HEVC (APSIPAXi’an, 2011).Google Scholar
 H Wang, H Du, J Wu, in Proc. 2014 IEEE Conf. Multimedia and Expo. Predicting zero coefficients for high efficiency video coding (IEEE, Chengdu, 2014).Google Scholar
 WH Chen, CH Smith, SC Fralick, A fast computational algorithm for the discrete cosine transform. IEEE Trans. Commun. 25(9), 1004–1009 (1977).View ArticleMATHGoogle Scholar
 Z Xie, Y Liu, J Liu, T Yang, A general method for detecting allzero blocks prior to DCT and quantization. IEEE Trans. Circ. Syst. Video Technol. 17(2), 237–241 (2007).View ArticleGoogle Scholar
 J Li, M Gabbouj, J Takala, Zeroquantized inter DCT coefficient prediction for realtime video coding. IEEE Trans. Circ. Syst. Video Technol. 22(2), 249–259 (2012).View ArticleGoogle Scholar
 H Wang, S Kwong, CW Kok, Efficient predictive model of zero quantized DCT coefficients for fast video encoding. Image Vis. Comput. 25(6), 922–933 (2007).View ArticleGoogle Scholar
 M Budagavi, A Fuldseth, G Bjøntegaard, V Sze, M Sadafale, Core transform design in the High Efficiency Video Coding (HEVC) standard. IEEE J. Sel. Topics Signal Process. 7(6), 1029–1041 (2013).View ArticleGoogle Scholar
 MT Sun, IM Pao, Statistical computation of discrete cosine transform in video encoders. J. Vis. Commun. Image R.9(2), 163–170 (1998).View ArticleGoogle Scholar
 EY Lam, JW Goodman, A mathematical analysis of the DCT coefficient distributions for images. IEEE Trans. Image Process.9(10), 1661–1666 (2000).View ArticleMATHGoogle Scholar
 AK Jain, Fundamentals of digital image processing (Prentice Hall, New Jersey, 1988).MATHGoogle Scholar
 HM 12 Software. https://hevc.hhi.fraunhofer.de/svn/svn_HEVCSoftware. Accessed 5 June 2017.
 G Bjø ntegaard, in Technical report VCEGM33, ITUT SG16/Q6. Calculation of average PSNR differences between RDCurves, (2001).Google Scholar