3D Segmentation-based Video Compression

3D Segmentation-based Video Compression

A 3D segmentation-based approach for the compression of video sequence is proposed. The sequence can be considered as a block of 3-dimensional data points with uniform volume regions bounded by contour surfaces. The segmentation method used to form this representation combines,

1. Segmentation-based Sequence Representation

A sequence is represented in terms of contours and regions.

2. Objectives and Significance of the Proposed Research

The primary objective of this research is to develop algorithms for image sequence compression based on the region growing methods in volume segmentation.

The end results can bring about a wide range of applications in condensed storage and archive of image sequences. Applications to video broadcasting, multimedia ,image database storage, image archive etc. would be appropriate.

It is expected that high quality image sequence with high compression ratio can be achieved by this compression technique. The encoding process is complex and computationally expensive while the decoding process is simple, fast and inexpensive. For example, in a video broadcasting, the encoding can be complex, expensive and computationally intensive but the decoding should be fast, simple and inexpensive. This compression method can meet the requirements of such situation.

3. Segmentation Using Gibbs-Markov Random Fields (GMRF)

In the segmentation method which forms part of the basis of this work, a GMRF is used for the region boundaries (contours) and a stationary Gaussian model for the grey level information or texture inside the regions. The steps involved in the segmentation process are,

4. Contour Surfaces Viewed in AVS

For example, the following is a sequence of circles shifting along a diagonal axis across the cube. The sequence simply contains only two regions separated by a round tube. If we can encode the partition boundary efficiently, we can achieve very high compression ratio for this sequence.

Figure 1: The segmented 'Circle'
sequence seen in AVS.

5. What We Have Already Done

In this section we would describe what we have already done so far. We have developed the algorithm to extract the contour surface of a sequence.

As a first step, individual frames are read in sequentially and stored in an array. The three steps, namely preprocessing, region growing and contour relaxation are performed on the sequence. The following is the picture showing the result of the Miss America sequence performed by our algorithm and viewed in the AVS 3D visual software package. The hollow holes in the image indicate the fact that those areas of the consecutive frames belong to a common region. Those regions are like tunnels passing through a block of volume data, the sequence.

Figure 2: The segmented 'Miss
America' sequence.

6. Encoder and Decoder

The following block diagrams described the encoding and decoding processes of our segmentation approach. The approximation steps include preprocessing, region growing, contour relaxation and contour surface encoding. Errors obtained from the difference between the approximation and the original images are retained for reconstruction.

Figure 3: The approximation steps.

The encoding process is described in the following block diagram.

Figure 4: The encoder block diagram.

The decoding process is described in the following block diagram.

Figure 5: The decoder block diagram.

7. Works in Progress

We are in the process of developing the algorithm to encode and decode the contour surface efficiently.

8. Concluding Remarks

This work described a segmentation-based approach for the compression of video sequences. The video sequence was approximated as a set of uniform regions. The segmentation method combines a Gaussian texture model and Gibbs-Markov contour model to approximate the interiors and the boundaries of objects in the sequence.

Gene K. Wu