Digital Imaging & Data Compression

 

Video Compression Based on MPEG-2

Our research on video compression based on MPEG-2 and its applications are mainly focused on the following directions:

(i)  Towards removing the encoding of motion vectors;

(ii) Content adaptive quantization;

(iii) Variable block sized encoding;

(iv) Multi-resolutional coding with Wavelets.

The first direction is based on the fact that MPEG-2 achieves motion estimation and compensation in terms of macro-blocks (1616), yet the transform coding and quantization etc. is carried out in terms of blocks (88). For each macro-block, if we take the block at the bottom right out and regard this block as the one to be encoded, the rest of the macro-block still reserves ¾ of its area. Thus, if this ¾ macro-block is used to search the reference frame for its best match, it is likely that the best match would be of the same location as that in MPEG-2 (the full area of the same macro-block). The following table illustrates our initial experimental results showing that the number of times with the same location as that of MPEG-2 reach 50% for all video clips tested, and in some cases, such as Claire, the value has reached 85%, which is indeed encouraging. The significance of so doing is that, if we conduct the motion estimation and compensation in terms of those 88 blocks, the ¾ macro-block will become available at the decoding end and hence, the same ¾ macro-block can be used to get the same location of the best match. Therefore, the bottom right block of the best matching ¾ macro-block will produce the same motion estimation and compensation as that in MPEG-2. Compared with MPEG-2, no bits will be required to encode such a motion vector since the ¾ macro-block is readily available at the decoding end.

Video Samples

Percentage of Same Matching Location as MPEG-2

Susie

76%

Missa

49.4%

Mobile

65%

Claire

85%

Salesman

74.4%

Trevor

76%

The proposed scheme can be illustrated in Figure 1.

            wpe1.jpg (10802 bytes)

The second direction is to connect the choice of quantization tables or redesign of quantization tables more tightly with content analysis or visual activity analysis within each individual frame in terms of blocks encoded. The visual activity analysis can be designed in various ways include:

(i) Local texture analysis;

(ii) Statistical parameter estimations, such as mean value, variance and other higher order moments;

(iii) contour and edge detections.

The above visual activity analysis can either be conducted in pixel domain or transform domain.

The visual activity analysis can also be directly exploited for our investigation along the third direction of research. The variable sized blocks can be obtained by one of the following techniques:

(i) Tree-based decomposition, such as binary tree, quad-tree etc.

(ii) Statistical parameter analysis;

(iii) Triangular mesh segmentation;

(iv) Free form regions comprising squared blocks.

Our research in fouth direction is established on zero-tree encoded wavelets coefficients, which has the same route as that of MPEG-4.

Back to Research Activities