In H.265 / HEVC, 35 prediction modes are defined based on the PU, while the implementation of the specific intra prediction process is in units of TUs. Compiling the provisions of the PU can divide the TUs in the form of a quadrush, and all TUs in the PU share the same prediction mode form division TU, and all TUs within a PU share the same prediction mode. H.265 / HEVC intra prediction can be divided into the following three steps: ** 1. Judge whether or not the current TU is available and corresponding to the corresponding processing 2. Filter the reference pixel 3. Calculate the reference pixel according to the filtered reference pixel Current TU predictive pixel value **
1. Acquisition of adjacent reference pixels
As shown in the figure, the current TU size is NXN, and the reference pixel can be divided into five parts: the lower left (a), the left (B), the upper left (c), above (d) and upper right (e), a total of 4N + 1 point. If the current TU is located at the image boundary, or the boundary of the slice (H.265 / HEVC is specified in intra coding, adjacent Slice or Tile cannot be referenced to each other), the adjacent reference pixels may not exist or not. In addition, the blocks of A or E in some cases may have not yet coded, and these reference pixels are not available.
Adjacent reference pixel position
When the reference pixel does not exist or not available, the H.265 / HEVC standard is filled with the nearest pixel. For example, if the reference pixel of the area A does not exist, the region A is filled with the pixels at the bottom of the area B; if the reference pixel of the region E does not exist, all reference pixels of the area E use D. The right side Pixel is filled. It should be noted that if all reference pixels are not available, the reference pixels are filled with a fixed value, the fixed value size is
R = 1 << (BitDepth -1)
Source code parsing: FillReferenceSamples This function is that the main function is to use the reconstructed YUV image to assign the current PU adjacent point to the next angle prediction to provide a reference sample value for the next angular prediction to the next angular prediction. Source code analysis can be used: http://blog.csdn.net/hevc_cjl/Article/details/8175721
Inframe prediction, a very important function initipaln, which has three main functions, (1) Detection of adjacent spots of the current PU include availability of orthoped value on the left, upper, upper right, left, left neighboring domain, or checks these Whether the point is existing; (2) Replacement process of reference spots, the main implementation is the content of JCTVC-J1003 is Draft 8.4.4.2.2, mainly due to function FillReferenceSamples; (3) adjacent samples Smooth filtering, mainly implementing the contents of Draft 8.4.4.2.3. Source code analysis can be referenced: http://blog.csdn.net/hevc_cjl/Article/details/8184276
2. Filtering of reference pixels
1. Different size TUs need to perform reference pixel filtering mode DC mode and 4x4 size TUs do not need to perform reference pixel filtering. Other situations are described as follows: 32x32TU: In addition to all angular patterns outside mode 26 (vertical) and Planar mode. 16x16TU: Four modes closest to horizontal vertical direction in 32x32TU - Mode 9, 11, 25, 27 8x8TU: Type 3 45 degree tilt direction (pattern 2, 18, 34) and Planar mode performs reference pixel filtering.
2. Conventional filtering and strong filtering methods are part of the need to meet several conditions for conventional filtering or strong filtering.
2. Predicting the calculation of pixels
Compared with H.264 / AVC, the H.265 / HEVC increases the boundary pixel of the lower left square as a reference for the current block. This is because the H.264 / AVC is encoded in a fixed size macroblock, and when the current block is performed intra prediction, the lower left square may have not yet encoded, and it cannot be used for reference; and H.265 / HEVC The coding structure of the four trigemoid makes this region become available pixels. In addition, the use of this area pixel also provides more possible prediction directions, in some cases, such as textures in the tilt direction, or the like), can greatly increase the prediction accuracy.
H.264 4x4 LuMa Prediction Modes
H.265 / HEVC frame reference diagram
H.265 / HEVC Brightness Component Accidentally Supports 5 Sizes of PU: 4x4, 8x8, 16x16, 32x32, 64x64 Each size PU has 35 prediction modes: 1. Planar mode 2. DC mode 3. 33 The angle mode All prediction modes use the same template.
1.Planar mode (solve the gradient smooth texture area)
The Planar mode is developed by the Plane model in the H.264 / AVC, which is suitable for areas where the image value is slow. Planar mode uses two linear filters in horizontal and vertical directions, and the average of both is used as the predicted value of the current block pixel.
Be
Interface PLANAR mode prediction
a) Calculation formula B) of corresponding pixels B) Its calculation effect diagram
2.DC mode (suitable for large-scale flat areas)
The DC mode is suitable for large-area flat areas, which are basically the same as H.264 / AVC. The current block predictive value can be obtained from its left and above (note the upper left corner, the upper left and upper right) reference pixels are obtained.
Non-edge blocks in 4x4, namely blue transparent part
3. Angle mode
H.264 / AVC uses different prediction directions (4x4 size), H.265 / HEVC further refines these prediction directions, specifies 33 angular prediction modes to better adapt to different ways of video content Texture.
The figure below gives the specific direction of 33 angular modes, where V0 (mode 26) and H0 (mode 10) are represented as vertical and horizontal directions, and the prediction direction of the remaining mode can be seen in the vertical or horizontal direction. An offset, the size of the offset angle can be calculated from the digital calculations below the mode.
Prediction direction of 33 angular modes
Be
θ is positive to the prediction direction to the left offset, θ is a negative representation of the prediction direction; for the horizontal class mode, θ is positive indicating the forward direction of the prediction direction, θ is negatively exposed downward offset downward.
For mode 1125, the prediction of the current block requires simultaneous use of reference pixels above and left. In order to be able to use a unified form, the H.264 / HEVC standard uses a "projected pixel" method. For mode 1825 (vertical class mode), the left reference pixel must be used in a given direction. The left side (horizontal arrangement) of the above reference pixels, and for the mode 11 to 17 (horizontal mode), the above reference pixels must be projected above the left reference pixel in a given direction (vertical arrangement).
Be
STEP:
Be
For a given vertical class mode M, the reference pixel that needs to be used in the use of "projection pixels" is used as the one-dimensional form, which is reiced as REF. 2. Calculate the position of the current pixel corresponding to the reference pixel, recorded as POS: POS = (Y * Offset [M]). 1882.> 5 3. Calculate the weighting factor W: W = (Y * Offset [M]) & 31 where the current pixel corresponds to the reference pixel, where, & reply with the operation . 4. Calculate the predicted value of the current pixel: PX, Y = ((32-W) REF [POS] + WREF [POS + 1] +16). 1882.>5 Need to note that for mode 26 (vertical mode) , Predicted pixel values were changed in the formula: PX, Y = RX, O + ((RO, Y - RO, O). 1882.> 0)
Source code parsing: There are functions of the above mode under the tcomprediction.cpp file in the hm. Specific parsing can be referring to: http://blog.csdn.net/hevc_cjl/article/details/8175721 http: //blog.csdn.net/hevc_cjl/ Article / details / 8184276 http: //blog.csdn.net/hevc_cjl/Article/details/8200793http: //blog.csdn.net/hevc_cjl/article/details/8216065 http: //blog.csdn.net/hevc_cjl/Article/ Details / 8259844HTTP: //blog.csdn.net/hevc_cjl/Article/details/8259844
Tcomprediction.cpp
Author: Persistently link: https://www.jianshu.com/p/d19d7eb3844a Source: The master of the brief review belongs to the author. Any form of reprints please contact the author to obtain authorization and indicate.
Our other product:
