|M.Sc Student||Assaf Tzabari|
|Subject||Compression at the Source for Video Coding|
|Department||Department of Electrical Engineering||Supervisor||Professor Emeritus Feuer Arie|
|Full Thesis text|
This work examines the idea of improving video compression by designated sampling of the video during the encoding and then performing reconstruction as part of decoding. Such system implements compression at the source, where the compression of the data starts before the encoding. The goal is to reduce the number of pixels significantly by using sampling pattern that is designed to exploit the frequency characteristics of typical video signals. Such sampling should allow high quality reconstruction, thus removing redundancy efficiently, while preserving video quality.
Previous work, that analyzed typical video characteristics, showed that for natural video signals the main energy in frequency domain is concentrated along the plane of spatial-frequencies and the axis of temporal frequencies. Based on that observation it was shown that 3D non-uniform recurrent sampling preserves more information than regular uniform sampling of the same rate. We examine the contribution of the non-uniform sampling to the compression, using the H.264 standard as a reference for comparison.
The system that is introduced here performs video compression in two steps. First, the video is down-sampled using the pattern that fits to our prior regarding the frequency characteristics of the signal. This should leave a minimal number of pixels that allow high quality reconstruction. In the second step, the system encodes the sampled video using a modified H.264 encoder.
The results show that the sampling significantly improves the compression as long as the video comply with our prior assumptions regarding its band limits in frequency domain. We show that H.264 standard is not efficient in removing the redundancy that exists in over-sampled signals. And therefore reducing pixel-rate through sampling, before the encoding, contributes to the compression. In such case, the performance of the system depends on the sensitivity to errors in the reconstruction. It also depends on the capability of the modified H.264 encoder to exploit correlations between the remaining samples.
Although in general video signals may exceed the spectral-support band limits that our prior assumption presents. We suggest using Compression at the source as a feature that upgrades the standard H.264 encoder and can be used selectively. The compression system should adjust every video as much as possible to comply with the prior, in order to benefit from the Compression at the source evident capabilities.