As mentioned earlier, whether it is image or video in airspace, there is a strong correlation, that is, for a pixel value and the pixel value of its neighboring pixels are very close, using this to predict the encoding The diamond correlation compressed image or video volume can be removed.
The intra prediction process is approximately as follows:
1, reference pixel preparation
Due to null domain correlation, we always select the reference pixel of the current PU as the current PU as the current PU, in the HEVC, is a column of one line and the left side of the current PU. Accurately, for the NXN PU selecting N pixels, n pixels in the upper right, N pixels on the left, N pixels in the lower left, 1 pixels in the upper left corner, a total of 4N + 1 pixels as reference pixels.
However, in some cases, if some reference pixels are not available in image boundaries, Tile or Slice boundaries, in some cases A, the E area has not yet encoded their pixels. It is necessary to find pixels to fill.
If the A region pixel does not exist, the pixel of the bottom of the B region is filled with the lowermost pixel if the E area does not exist, filling the pixel on the rightmost right side of the D area, as shown in the figure above.
If the pixels in all regions do not exist, all reference pixels are filled with R = 1<< (BitDepth-1), with 128 filled with 8bit pixels, 10bit pixels are filled with 512.
At that, all reference pixels are prepared, and these reference pixels will be filtered to reduce the impact of noise, but since the filtering method is closely related to the PU size and prediction mode, this part is placed in Section 3 (o = ^ • ェ •) o ┏┏┓
2, intra prediction
HEVC provides 35 modes for intra prediction, which is DC mode, PLANAR mode, and 33 angular modes. The intra mode number and the corresponding name are as follows:
Mode 0 is a PLANAR mode, 1 is a DC mode, 2 to 34 correspond to 33 angular modes.
The 2-17 polarization direction, the 19-34 directional direction, the mode 10 is the horizontal direction, and the mode 26 is the vertical direction, the more densely distributed in the horizontal or vertical direction in these angular patterns, because the object in the nature is at level The vertical direction movement is more accurately predicted in these directions.
So how do you use the reference pixel to calculate the predicted value based on these patterns?
(1) PLANAR mode
The pixel P at the corresponding (x, y) position must be used to calculate its predicted value by using the Planar mode. The first pixel is first pixels and the first pixels in the upper right level of PU. Pixel. The calculation method is as follows:
This calculation method can be calculated from each pixel in the PU to calculate the predicted value of each pixel in the Planar mode.
(2) DC mode
The DC mode first needs to calculate the average of the current PU and the left pixel (excluding the upper left corner, the upper right and left below)
If the current PU is the brightness PU of the chroma PU or greater than 16x16, all pixel predictive values in the PU are dcValue, and the predicted values are calculated according to the following methods for other situations (brightness PU having a size equal to 16x16).
(3) angular mode
For angular mode, its angular delay may both relate to the left reference pixel, if these pixels are projected to a row or a column, such as the vertical category, the left pixel projection To above.
Because each angular direction is different, the reference pixels that need to be projected each angle is also different. Each angle has an offset value for the vertical direction (mode 26) or horizontal direction (mode 10), as shown below.
A one-dimensional reference pixel list Ref [] is constructed as an example of the vertical class mode M (corresponding offset value is OFFSET [M]).
1Offset [M] <0, ie M is 18 to 25
2Offset [M] ≥ 0, ie M is 26 to 33
The horizontal class one-dimensional reference list construction method is similar, the following is the 4X4PU pixel projection example
You can get a one-dimensional reference pixel list REF [],
The following is the location of the corresponding reference pixel in the REF for pixels (x, y).
POS = (Y * offset [M]). 1882.>5
Pixel (x, y) corresponding weight factor W:
W = (Y * Offset [M]) & 31, where & represents bits and operations
The predicted value of the (x, y) pixel is:
3, reference pixel filtering
When the left oriented adjacent pixels are used directly as the reference pixel, it is usually used to reduce noise, increase prediction accuracy, and need to smoothing the reference pixel when selecting certain prediction modes, usually uses a 3 tap. The low pass filter is performed. details as follows:
Smooth filtering is required for 4x4 PU. For all other size PUs, if the DC mode does not require smoothing, if it is a Planar mode, smooth filtering is required. For angular mode.
The 8x8PU is only used for routine smooth filtering of patterns 2, 18, 34. In addition to the mode 9, 10, 11, 25, 26, 27, the 16x16PU is required to perform conventional smooth filtering. Conventional smooth filtering or strong filtering should be performed for 32x32PU except mode 10, 26.
Conventional filtering: tap coefficient is [1, 2, 1] / 4
Strong filtering: Strong filter only on 32x32PU, and to satisfy
Strong filter calculation method is as follows:
4, PU boundary value smoothing
In order to remove the discontinuous effect of the boundary, for a PU of less than 32x32, when mode 1, 10, 26, the first row and the first column after the PU are predicted to filter processing.
5, intra mode encoding
When the intra prediction mode is selected, the prediction mode is to be encoded to the decoding end. Because 35 modes of intra prediction require 6bit to encode. HEVC defines three most powerful modes MPM [0], MPM [1], MPM [2] for the current PU. If the current mode only needs to encode its index in the MPM, if not in the MPM Only 5bit can be encoded in another 32 modes.
The configuration of the MPM needs to utilize the pattern information of the encoded PU adjacent to the left and above.
The MPM construction process is as follows:
(1) Modea and Modeb
1Modea and Modeb are DC mode or Planar mode
MPM [0] = Planar
MPM [1] = DC
MPM [2] = 26
2Modea and Modeb are all angular patterns
MPM [0] = Modea
MPM [1] and MPM [2] are two angle modes adjacent to MODEA
(2) Modea and Modeb are different
MPM [0] = Modea
MPM [1] = MODEB
1 If Modea and Modeb are not Planar mode, MPM [2] = Planar
2 When 1 is not satisfied, if Modea and Modeb are not DC mode, MPM [2] = DC
3 When 12 is not satisfied, MPM [2] = 26
Be
I am interested in scanning to pay attention to WeChat public number! (Micro signal: video code)
Our other product:
