EE 5359 MULTIMEDIA PROCESSING SPRING 2011 Final Report IMPLEMENTATION AND ANALYSIS OF DIRECTIONAL DISCRETE COSINE TRANSFORM IN H.264 Under guidance of DR K R RAO DEPARTMENT OF ELECTRICAL ENGINEERING UNIVERSITY OF TEXAS AT ARLINGTON Presented by PRIYADARSHINI ANJANAPPA [email protected] A popular scenario in image blocks is the occurrence of directional edges. By recognizing such characteristics, the video coding standard H.264/advanced video coding (AVC) [2] has developed a number of directional predictions in the coding of all intra blocks- called intra predictions. But it is still the conventional discrete cosine transform (DCT) [3] that is used after each intra prediction. Conventional DCT [3] The 2D DCT of a square or a rectangular block is used for almost all block-based transform schemes for image and video coding. The conventional 2D DCT is implemented separately through two 1D transforms, one along the vertical direction and the other along the horizontal direction, as shown in Fig.1. These two processes can be interchanged, as the 2D DCT is a separable transform. The conventional DCT seems to be the best choice for image blocks in which vertical and/or horizontal edges are dominating. Fig. 1. 2D DCT implementation: A combination of 1D DCTs along horizontal and vertical directions.Forward 2D DCT (NXM) XC2 (k,l) = C(k) C(l) cos [ ] cos [ ] k = 0,1, … ,N-1 C(p) = , p=0 p=k,l l = 0,1, … ,M-1 C(p) = 1 , p=0 x(n,m) = samples in the 2D data domain XC2 (k,l) = coefficients in the 2D DCT domain Inverse 2D DCT (NXM) = ( C(l) XC2(k,l) cos [ ] ) cos [ ] n = 0,1, … , N-1 m = 0,1, … , M-1 The transform used by H.264/AVC to process both intra and inter prediction residuals [4] is related to an integer 2D DCT, implemented using 1D DCTs horizontally, followed by 1D DCTs vertically (This can be interchanged). It has been found that the coding efficiency can be improved by using directional transforms [1][6], since the residuals often contain textures that exhibit directional features. A directional DCT (DDCT) framework [1] has been developed which provides a remarkable coding gain as compared to the conventional DCT. The encoder block diagram representing the basic coding structure for H.264/AVC for a macroblock is shown in Fig. 2. The illustration of H.264 profiles is shown in Fig.3.Fig. 2. H.264 encoder block diagram [2] Fig. 3. Illustration of H.264 profiles [14] Intra coding in AVC and where DDCT fits in [8] The H.264 encoder forms a prediction of the current macroblock – One based on the current frame using intra prediction/spatial prediction technique. Intra prediction is an important technique in image and video compression to exploit spatial correlation within one picture. It has 4 prediction modes for16x16 blocks (shown in Fig. 4), 9 prediction modes for 8x8 blocks and 9 prediction modes for 4x4 blocks (shown in Fig. 5) [4][8]. a) 16x16 (for Luma) Mode 0 (vertical): extrapolation from upper samples (H). Mode 1 (horizontal): extrapolation from left samples (V). Mode 2 (DC): mean of upper and left-hand samples (H+V). Mode 3 (Plane): a linear “plane” function is fitted to the upper and left-hand samples H and V. This works well in areas of smoothly-varying luminance. Fig. 4. 16x16 luma intra prediction modes [5] b) 4x4 (for Luma) Mode 0 - Vertical Mode 1 - Horizontal Mode 2 - DC Mode 3 - Diagonal-down-left Mode 4 - Diagonal-down-right Mode 5 - Vertical-Right Mode 6 - Horizontal-down Mode 7 - Vertical-left Mode 8 - Horizontal-upFig. 5. 4x4 luma intra prediction modes [5] A-H -> they are the previously coded pixels of the upper macroblock and are available both at encoder/decoder. I-L -> they are the previously coded pixels of the left macroblock and are available both at encoder/decoder. M -> it is the previously coded pixel of the upper left macroblock and available both at encoder/decoder. For each intra prediction mode, an intra prediction algorithm is used to predict the image content in the current block based on decoded neighbors. The intra prediction errors are transformed using a 4x4 integer DCT. An additional 2x2 Hadamard transform is applied to the four DC coefficients of each chroma component. If a macroblock is coded in intra- 16x16 mode, a similar 4x4 transform is performed for the 4x4 DC coefficients of the luma signal, as shown in Fig. 5a. [2]. In this framework, DDCT is to replace the AVC transforms by a set of transforms taking into account the prediction mode of the current block. Hence, DDCT provides 9 transforms for 4x4, 9 transforms for 8x8, and 4 transforms for 16x16, although many of them are the same or can be simply inferred from a core transform. For each transform, the DDCT also provides a fixed scanning pattern based on the quantization parameter (QP) and the intra prediction mode to replace the zigzag scanning pattern of DCT coefficients in AVC.Fig. 5a. 4x4 DC coefficients for intra 16x16 mode Transforms [8] DDCT provides 9 transforms for 4x4, 9 transforms for 8x8, and 4 transforms for 16x16 [4][5][8]. For each intra prediction mode, DDCT consists of two stages: - Stage 1 – along the prediction direction: pixels that align along the prediction direction are grouped together and 1-D DCT is applied. Note that, in cases of prediction modes that are neither horizontal nor vertical, the DCTs used are of different sizes. - Stage 2 – across the prediction direction: another stage of DCT is applied to the transform coefficients resulted in the first stage. Again, the DCTs may be of different sizes. Six directional modes of DDCT are shown in Fig. 6. Stage 1 is illustrated in Fig. 7. and stage 2 is illustrated in Fig. 8. To make the transform sizes more balanced, the DDCTs group pixels in the corners together in order to use DCT of longer size, hence more efficient in terms of compression.Fig. 6. Six directional modes in DDCT defined in a similar way as in H.264 for the block size 8x8. [1] (The vertical and horizontal modes are not included here) Fig. 7. NXN image block in which 1D DCT is applied along diagonal down left direction (mode 3)[1] Fig. 8. Arrangement of
View Full Document
Unlocking...