Comparison of H.266 with H.265, AV1, H.264

The composition of an image sequence of H265: VPS+SPS+PPS+SEI+one I frame + several P frames. VPS, SPS, PPS, SEI, one I frame, one P frame can all be called
It is a NALU.

2. Differences in codec frameworks

H.265 still uses hybrid codec, and the codec structure is basically the same as H.264

H.265 classic framework:

3. Block division structure:

H.264 is a 16x16 (sub-block size can be 8X16, 16X8, 8X8, 4X8, 8X4, 4X4 is very flexible) macro block,

VP9 can be sampled in 64×32 or 4×8 blocks, supports the use of 64×64, and supports segmentation of frames into regions with specific similarities; compared to H.265, VP9 supports horizontal or vertical subdivision

H.265 is a recursive structure using CU (CodingUnit), PU (PredictionUnit) and TU (TransformUnit), quadtree division (prediction block brightness 64x64-8x8, chroma 32X32-4X4, transform block 32x32->4x4), and H.265 adds an asymmetrical partition mode; the specific segmentation process is marked by two variables: the depth of segmentation (Depth) and the segmentation flag (Split_flag). The H.265/HEVC standard breaks through the previous standard for prediction blocks and transform blocks Restrictions on size relations. Since the PU and TU are directly divided by the CU, there is no definite relationship between the sizes of the two. A PU may contain multiple TUs, and a TU may span multiple PUs.

On this basis, in addition to quad-tree division, H.266 adds tri-tree and binary tree division.

Thinking: Can it be divided into irregular shapes? Such as triangles, circles, ellipses, hexagons and so on.

4. Intra prediction:

The 4×4 and 8×8 blocks in H.264 contain 9 prediction modes, and the 16×16 block contains 4 prediction modes;

VP9 has 10 intra prediction modes;

H.265 has 33 intra-angle prediction modes + DC (top and left averaging) + planer; compared with H.264/AVC, H.265/HEVC increases the use of the boundary pixels of the lower left square as the reference for the current block ;

H.266 has 65 intra-frame brightness angle prediction modes, in fact there are 65+10+10=85, which are selected according to the aspect ratio; increase ISP (further division technology for blocks); PDPC technology, combined with unfiltered For reference pixels and filtered reference pixels, add MIP mode; CCLM mode;

Note: Planar mode is suitable for areas where the pixel value changes slowly. It uses two linear filters in the horizontal and vertical directions, and uses the average of the two as the predicted value of the current block of pixels. The DC mode is suitable for large flat areas. The current block prediction value can be obtained from the average value of the reference pixels on the left and above. Angle mode is mainly used for textures in different directions in video content.

5. Inter-frame prediction:

Frame type structure: H.265 uses HIERACLE-B structure

mv precision: H.265 is pixel precision (chroma), and uses more adjacent pixels for sub-pixel precision interpolation. Prediction modes: SKIP, DIRECT, MERGE (5 candidate MVs), AMVP (2 candidate MVs).

Pixel accuracy improved by H.266;

VP9 inter-frame prediction uses ⅛ pixels for motion compensation. There are non-displayable frames as reference frames, and the non-displayable frames have average bidirectional prediction.

Reference list:

H.265 uses two reference lists, each with 16 reference items, but the maximum number of unique pictures is 8.

There are 6 candidates in the merger mode of H.266. Compared with H.265, TMVP and HMVP are changed.

6. Transform

H.264 integer DCT 4X4 8X8; Hadamard transform

Both VP9 and HEVC support transform block sizes of 4x4-32x32. DCT In intra-coded macroblocks, one or both of the vertical and horizontal transformation paths will be DST

HEVC 4X4 DST; Transform_skip mode: transform_skip_flag, this mode has a good effect on text desktop video; RQT technology is based on quad-tree adaptive transformation technology; there is no Hadamard transform

HEVC internal bit depth increase: In order to ensure the internal bit accuracy in the intermediate prediction, transformation and quantization process, in order to achieve better compression performance

HEVC only adopts 4-point DST7 for intra-frame prediction residual transformation, and DCT2 is still used for other sizes and inter-frame prediction residuals;

H.266 has an inseparable secondary transform lfnst; MTS (Multiple Transform Selection), which uses multiple candidate transforms for the prediction residuals, can better adapt to the statistical characteristics of the dynamic changes of the prediction residuals, and significantly further improve the transform gain. For inter-frame transformation technology, sub-block transformation technology (Sub-block Transform, SBT)

7. Entropy coding:

H.264 uses Integer Discrete Cosine Transform (DCT), CABAC compression (lossless, CABAC is also short code for high-frequency data, long code for low-frequency data. It will also be compressed based on contextual relevance), between two I frames is A picture sequence GOP.

VP9 supports four transform sizes: 32x32, 16x16, 8x8 and 4x4. These transformations, like most other codes, are approximate integers of DCT. In intra-coded macroblocks, one or both of the vertical and horizontal transformation paths will be DST (Discrete Sine Transform).

The entropy coding of HEVC uses two arithmetic codes: CABAC and CAVLC. CAVLC is mainly used to encode SEI, parameter sets, film headers, etc., and all remaining data and syntax elements are encoded using CABAC.

H.265: zigzag scan: ACS technology, vertical scan, horizontal scan, diagonal scan.

8. Filtering:

H.265 adds SAO

ALF is added in H.266, brightness 7x7, chroma 5x5

Three different sub-pixel interpolation filters can be selected for each block of VP9:

Normal 8th pixel/smooth 8th pixel, can be smooth or fuzzy prediction/sharp 8th pixel, can be sharp prediction

9. Acceleration technology

H.265 adds parallel toolsets such as Tile and WPP to improve encoding speed

Tile divides the image into rectangular areas. A Tile block is a basic parallel unit. There can be multiple tiles in some slices and multiple slices in some tiles.

WPP: The full name is wavefront parallel process, which is the basic coding unit of LCU behavior.

One line of LCU block is the basic parallel unit, and each line of LCU is a sub-stream

10. Other

VP9 optimizes the 8th pixel accuracy of the motion vector, three switchable sub-pixel interpolation filters, reference motion vector, entropy coding, loop filtering, ADST, DCT, etc.

H.264 Level: The description of the video, the higher the Level, the higher the bit rate, resolution, and fps of the video

H.266: Chroma Joint Coding JCCR

HEVC IBDI technology