Issues in multi-view autostereoscopic image compression

Multi-view auto-stereoscopic images and image sequences require large amounts of space for storage and large bandwidth for transmission. High bandwidth can be tolerated for certain applications where the image source and display are close together but, for long distance or broadcast, compression of information is essential. We report on the results of our two-year investigation into multi-view image compression. We present results based on four techniques: differential pulse code modulation (DPCM), disparity estimation, three-dimensional discrete cosine transform (3D-DCT), and principal component analysis (PCA). Our work on DPCM investigated the best predictors to use for predicting a given pixel. Our results show that, for a given pixel, it is generally the nearby pixels within a view that provide better prediction than the corresponding pixel values in adjacent views. This led to investigations into disparity estimation. We use both correlation and least-square error measures to estimate disparity. Both perform equally well. Combining this with DPCM led to a novel method of encoding, which improved the compression ratios by a significant factor. The 3D-DCT has been shown to be a useful compression tool, with compression schemes based on ideas from the two-dimensional JPEG standard proving effective. An alternative to 3D-DCT is PCA. This has proved to be less effective than the other compression methods investigated.