Coding of Guiding Data for VideoTranscoding

Abstract: In today’s and tomorrow’s internet traffic videos play an increasingly larger role. Thisis true for both live transmissions like video conferences, and videos stored oninternet platforms which the users can access at any time, the so-called Video onDemand (VoD). This thesis focusses on the latter case. Here the provider has severaloptions how to make the content available. Possibilities are for example simulcast,transcoding, or guided transcoding. While each of them has advantages, there are alsodrawbacks. These canbe storage requirements, computational complexity, or a loss of quality.A different approach named deflation tries to minimize the drawbacks by using lessstorage than simulcast, being not as computationally expensive as transcoding, andproviding a higher quality than guided transcoding. The first step is to estimate thetransform coefficients based on the encoded version of the original video, which isdecoded and downsized to the wanted resolution. Deflation then calculates thedifference between the estimated and original transform coefficients, which arecreated by encoding the original video at the required resolution. These coefficientscontain the actual picture data. Theyare then written into the bit stream, along with the control data containinginformation like prediction mode or how the picture is divided into Transform Blocks(TBs). In the inverse operation called inflation these details are parsed from the bitstream andthe delta is added to the estimated coefficients, recreating the original values.However, since the deflation uses the identical encoding methods as the most recentvideo encoding standard HEVC, there is potential for improvement as the encodercan be optimizedfor deflation and its special cases and coefficient layout. After a thorough analysis todetermine the layout of these delta-coefficients several new syntax elements wereadded and existing ones modified.These changes were evaluated in various configurations. A key difference to theoriginal scheme is the usage of knowledge gained from the estimated coefficients,which can be applied due to correlations between positions and magnitudes of bothcoefficient groups. The introduced changes reduce the storage requirements fordeflated files by between 1.5 and 3 percentage points compared to the originaldeflation scheme, with the exact values depending on the configuration and settings ofthe representation. While there are still many options to improve the deflationscheme, the additions made and described in this thesis proved to be quite successful.