Intra Prediction

In the intra prediction, it is one of the important prediction engines in the video coding stan-dard. Since in the H.264/AVC standard, the intra prediction is adopted to improve the spatial redundancy. When the previous frame changed abruptly, like scene change, therefore, the tem-poral domain prediction will not guarantee the prediction accuracy. Unfortuntely, it may cause the bit-rate higher and would transmit the prediction error. The purpose of the intra prediction is to help the temporal domain prediction to compromise the scene change. Normally, the period of the intra prediction in the group of pictures is the first one every IPPBIPPB or randomly in-sert an I frame to shut down the error propagation induced by the temporal prediction. With the help of the intra prediction, the H.264/AVC can achieve the hottest video applications during the last decade. In HEVC, the call for proposal during the meeting aims to enhance the bit-rate savings. Therefore, the intra prediction is urgently to further improve the prediction accuracy than previous standard in H.264/AVC. The mode of intra prediction is an important issue to be discussed. In H.264/AVC, the specification defines 9 modes for the 4x4 block, intra 4x4, 4 modes for 16x16 block, intra 16x16 as shown in Figure 2.3. The concept of the intra prediction is to use neighboring pixels to predict the current block. Therefore, the modes in H.264/AVC are sufficiently to predict frames from top, left, to right top corners. However, if the required resolution is Ultra-HD or more, the prediction accuracy for the intra prediction apparently is in-sufficient. Accordingly, the MPEG team aims to add more modes to further improve the coding efficiency and gain more bit-rate savings. In the specification of HEVC, the modes are defined as 35 more than previous standard as shown in Figure 2.4. Certainly, the results of the predicted frames are more precisely than before. The basic concept of the spatial domain prediction is to utilize the neighboring pixels including top and left reference pixels. The Table 2.1 shows that intra mode number corresponding to the associated names.

In the mode number 1, Intra DC is to average the top row and left column reference pixels same as the H.264/AVC. In addition, the HEVC standard defines the DC post filtering in order to smooth the blocky effect as shown in Figure 2.5(d). In the direct mapping mode such as vertical,

1

0 3 4

6 8

5 7

Figure 2.3: Direction of intra prediction in H.264

17 16 1 5 1 4 1 3 1 2 11 10 9 8 7 6 5 4 3 2

18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34

A

A

B

Figure 2.4: Direction of intra prediction in H.265

horizontal, diagonal from top right, bottom left and top left are the same as H.264/AVC. In the horizontal-8 mode, if the pixels are unavailable, the pixels are the padding. Also, in order to reduce the blocky effect, after the vertical and horizontal prediction, the post filter between boundary is filtered with predicted pixels and neighboring pixels.In the mode range of 2-34, is

Table 2.1: Intra Mode in HEVC Intra Prediction Mode Associate Names

0 Intra Planar

1 Intra DC

Others Intra Angular

35 Intra from Luma (chroma)

(a) Vertical Mode (b) Horizontal Mode (c) Vertical+8 Mode

Average

(d) DC Mode

(e) Vertical-8 Mode

PADDING

(f) Horizontal-8 Mode

Figure 2.5: Intra Mode

named as Angular Directional Intra (ADI). In the ADI, directions can be classified into two groups. As shown in Figure2.4, the group A uses positive angles while group B uses negative angles. The positive prediction angles have the range from [2, 5, 9, 13, 1, 7, 21, 26], 0 and 32 angles are the direct mapping in the vertical+8, vertical-8 and horizontak-8. The angle is defined as the displacement of the current pixel and top reference pixel in the vertical prediction.

Also, it is defined as the right current pixel and left column pixel in the horizontal prediction.

We first describe the group A of intra prediction in detail. In Figure 2.6, take mode 27 for example, the current 4x4 block only needs top reference pixels. Different with H.264/AVC, the current predicted pixel needs two inputs to do the linear interpolation. According to the specification, the choice of the two pixels is adopted the angle calculation. As shown in Figure 2.6, the parameter P OS₀ will be added with angles which would be achieved by the look-up

table to get P OS₁, then P OS₁ will next to be added with the same angles to get the P OS₂. After finishing angle operations, the output of the intra prediction pixel is calculated in 32-tap filtering. If the angles are negative, then the Table2.2 shows the negative angels corresponding to the inverse angles. The processing step to do the intra prediction in negative angles is to first flip the pixels from the side to the main as shown in Figure 2.7(a). Also, the pixel decision is through the look-up table and then do the interpolation as the same as Figure 2.6.

C

A B

POS0

POS1

POS2

POS³

angle

A C B

32-POS0 POS0

angleangle

interpolate

Figure 2.6: 4x4 block interpolation of intra prediction

Table 2.2: Intra Inverse Angles in HEVC

Intra Prediction Mode 2 3 4 5 6 7 8 9

inv Angle -256 -315 -390 -482 -630 -910 -1638 4096

Intra Prediction Mode 19 20 21 22 23 24 25 26

inv Angle -315 -390 -482 -630 -910 -1638 4096

(a) flip the pixels (b) inverse angle filtering

Figure 2.7: Inverse Angle

To gain the benefit of the homogeneous region in the I frame, HEVC proposed the new planar mode interpolation. As shown in Figure 2.8, the planar operation requires two steps. In the left filtering, pixel a is copied to the rightmost of the current pixel, the leftmost of the pixel

is linear operation with pixel a. In the right filtering, the pixel b is also put in the bottommost of the current pixel, the topmost pixel is linear operation with the pixel b. Finally, the bilinear filtering will get the planar output. In concept of the intra prediction algorithm, it requires top row and left column pixels for the input filtering. Further, in the hardware-based design, the coding order follows the raster scan order. Because the next line of LCUs are not available when decoding the first line of LCUs. Accordingly, the intra predicted pixels after the recon-struction with transform data should be stored into the line memory. As the next line of LCUs are ready to decode, the LCU would access the line buffer for the filtering. For not harming the external memory bandwidth, the conventional design requires 1-line internal buffer for storing the temporal pixels. However, the line buffer depends on the frame width would be higher if the H.265 is targeted for the high resolution. How to reduce the line buffer is a big issue in the decoder design discussed in next chapter.

Figure 2.8: Planar Prediction in H.265

Current block Reconstructed block Width

Pixels need to be stored

Figure 2.9: Line pixels should be stored into libe buffer