This article comes from the ICIP2020 paper "A FAST LOSSLESS IMPLEMENTATION OF THE INTRA SUBPARTITION MODE FOR VVC"
VVC mainly implements lossless encoding through Transform Skip Mode (TSM). Since VVC does not support the simultaneous use of TSM and Intra Subpartition (ISP), this article will study the combination of TSM and ISP in VVC lossless encoding.
The main configuration of the lossless coding in VVC is as follows:
-
Use TSM mode for all CUs.
-
For an 8-bit input signal, select QP as 4.
-
Disable loop filtering.
Introduction to ISP
ISP is to further divide the luma intra prediction block into sub-blocks. It divides the luma intra prediction block into n sub-blocks in the horizontal or vertical direction, and the number of sub-blocks n depends on the CU size.
-
If the CU contains more than 32 pixels, n=4
-
If the CU contains exactly 32 pixels, n=2
-
If the CU contains 16 pixels (4x4), ISP mode is not allowed
Fig.1 shows the ISP division when the CU contains more than 32 pixels.
Each sub-block divided by ISP uses the same intra-frame prediction mode. Therefore, it is only necessary to transmit the intra-frame mode once for a CU. Their intra-frame prediction process is the same as that of ordinary intra-frame prediction: first, the predicted value and residual value are generated, and then , The residual is transformed and quantized, and the non-zero coefficients are entropy encoded and transmitted to the decoding end. The result can be used to predict the next sub-block after reconstruction.
In VVC, the same CU does not allow ISP and TSM modes to be used at the same time, mainly because the combination of the two modes in lossy coding is not high. But this does not take into account the lossless coding situation, this article mainly completes the following two things:
-
Allow CU to use ISP while using TSM
-
Achieve the balance of coding time and coding gain by realizing fast search algorithm without loss
ISP in lossless encoding
Combination of ISP and TSM
In order to combine TSM and ISP in lossless encoding, the following improvements are needed:
-
Each sub-block of the CU is allowed to use ISP coding, and the TSM flag bit must be parsed when the corresponding CBF is non-zero. The context model of TSM in entropy coding also uses the one provided in VVC.
-
On the encoding side, the TSM mode is always selected for the sub-blocks generated in the ISP mode.
Coding end estimation
-
Intra prediction search in VTM7.0
The intra-frame prediction search process is used to select the optimal mode, including intra-frame prediction mode, MRL, MIP, ISP, etc. The mode with the smallest RD cost is selected, and the distortion is measured by the mean square error of the original pixel and the reconstructed pixel . However, the complexity of calculating RD cost for all modes is too high, so VTM7.0 builds the Most Promising Candidates (MPC) list, and only needs to calculate RD cost for the modes in MPC. The MPC list creation process is as follows:
-
Sort the RD cost of MRL, MIP and ordinary intra-frame candidate modes (excluding ISP), and the distortion is measured by the SAD of the predicted pixel and the original pixel .
-
Only K candidates are selected, and K is determined by the block size and the RD cost of the MIP mode.
-
Add candidates from Most Probable Modes (MPM) to the list.
-
Add the ISP mode to the end of the list.
It can be seen that there is no RD cost estimation for the ISP mode above. This is because the ISP mode cannot accurately calculate SAD in lossy coding. Except for the first sub-block, other sub-blocks need to be predicted using the reconstructed value of the previous sub-block, and this must wait for all RD cost to be calculated. obtain.
The intra-frame search of lossless coding in VTM7.0 is the same as the above, the only difference is that distortion is not calculated when calculating RD cost.
-
Lossless coding ISP fast search optimization
A feature of lossless coding is that all reconstructed pixels can be estimated because it is the same as the original pixel, so that SAD can be calculated for the ISP mode. This article proposes a fast algorithm to replace the fourth step of the above MPC list construction,
4.a Let C be equal to the second best RD cost in the MPC list (before the ISP mode is added).
4.b Calculate the ISP-SAD of all pre-selected common intra prediction modes to construct a sub-list, and calculate twice for each mode because ISP has two division methods, horizontal and vertical.
4.c Calculate the SAD-based RD cost for all elements of the ISP-SAD sublist.
4.d Add all ISP models with SAD-based RD cost less than C to MPC. In addition, every time an ISP mode is added, the worst non-ISP mode is removed. But at least one non-ISP mode is reserved.
The MPC construction process is shown in Fig.2,
Experimental results
The experiment was completed in VTM7.0 lossless mode. A total of two sets of experiments: TSM and ISP combined without search optimization (V1), TSM and ISP combined with search optimization (V2).
The experimental results are shown in the table, V1 coding gain 0.73% while coding takes 127%, V2 coding gain 0.71% while coding takes 111%. The results show that fast search optimization can achieve a better balance between coding time and coding gain. At the same time, both V1 and V2 save bit rate, which shows that combining TSM and ISP in VVC lossless coding can achieve good results.
If you are interested, please follow the WeChat public account Video Coding
