Introduction

Although series of JPEG and MPEG standards represent worldwide efforts for developing image compression technology, they are not optimal for the compression of 3D stereo image pairs. In addition, due to the relatively large disparities between the images, the expected compression ratio is lower than that expected for motion video [1].

3D images are constructed by simulating the effect of human eyesight when observing objects through two horizontally separated positions. Correspondingly, two frames are produced for one 3D image, labelled as left frame and right frame . Since 3D images have the feature that the two frames are produced from observations of the same scene with two different positions in parallel, both inter-frame and intra-frame correlation should be fully exploited before any data compression algorithms are designed.

For the purpose of this project, a two-pronged approach was taken:

1.

Compressing the left image The first step was to compress a reference image as well as possible, given no other information. This is essentially a standard image compression problem. The left image was chosen as the reference image in this project.

Various methods were evaluated and compared for suitability for this purpose; different sets of requirements suggested, and were met by, different methods for compression.

2.

Compressing the right image

Once a reference image has been optimally compressed, the right image can be compressed given full knowledge of the left image. Again, different requirements yield different approaches to this problem.

Since it is not possible to proceed with the second step until the first step is complete, the 2 steps were viewed as fundamental to this project; both were focussed on. Another approach could have been to focus only on the second step, but it was felt that that would bypass a fundamental aspect of stereo image compression.

A number of methods and approaches to each problem were proposed. The more promising of these methods are evaluated and discussed in this report.

Five 8-bit greyscale stereo pairs were chosen as test images for this project. They were selected as representative of a wide range of potential stereo sources. Two images (``Bishop'' and ``Flowers'') are stereo still photographs. Two more (``Funfair'' and ``Tunnel'') are frames from a stereo video sequence. The last image (``Vista'') is a stereo computer-rendered scene.

 

Figure 1: ``Bishop'' ©1996 Boris Starosta, http://www.starosta.com. Used with permission.

 

Figure 2: ``Flowers'' ©1996 Boris Starosta, http://www.starosta.com. Used with permission.

 

Figure 3: ``Funfair''

 

Figure 4: ``Tunnel''

 

Figure 5: ``Vista'' ©1996 Boris Starosta, http://www.starosta.com. Used with permission.