Bayesian Decoding for Improved Random Access in Compressed Video Streams
Independent thesis Advanced level (degree of Magister), 20 points / 30 hpStudent thesis
A channel change in digital television is usually conducted at a reference frame, which are sent at certain intervals. A higher compression ratio could however be obtained by sending reference frames at arbitrary long intervals. This would on the other hand increase the average channel change time for the end user. This thesis investigates various approaches for reducing the average channel change time while using arbitrary long intervals between reference frames, and presents an implementation and evaluation of one of these methods, called Baydec.
The approach of Baydec for solving the channel switch problem is to statistically estimate what the original image looked like, starting with an incoming P-frame and estimate an image between the original and current image. Baydec gathers statistical data from typical video sequences and calculates expected likelihood for estimation. Further on it uses the Simulated Annealing search method to maximise the likelihood function.
This method is more general than the requirements of this thesis. It is not only applicable to channel switches between video streams, but can also be used for random access in general. Baydec could also be used if an I-frame is dropped in a video stream.
However, Baydec has so far shown only theoretical result, but very small visual improvements. Baydec produces images with better PSNR than without the method in some cases, but the visual impression is not better than for the motion compensated residual images. Some examples of future work to improve Baydec is also presented.
Place, publisher, year, edition, pages
Institutionen för teknik och naturvetenskap , 2005. , 52 p.
Programmering, algoritmer och datastrukturer, Video compression, video streams, likelihood, television, broadcast, multicast, digital signal processing, random access, channel change, channel switch, IPTV, digital video, bayesian statistics
Programmering, algoritmer och datastrukturer
IdentifiersURN: urn:nbn:se:liu:diva-297ISRN: LiTH-ISY-EX--05/3540--SEOAI: oai:DiVA.org:liu-297DiVA: diva2:20313