(Edited from the original in order to (1) name the isovideo products used in the tests, and (2) to make available (in Section 4) links to some Demeler-deinterlaced yuv4mpeg2 files for confirmation of our results. More results are available in the long version of this paper at Deinterlacing Interlaced Video Prior to H.264 and HEVC Compression

The HEVC and AVC compression efficiency of interlace content coded with interlace modes is informally compared against the same content deinterlaced and coded as progressive sequences. For example, each raw source 1080i 60 fields/sec interlace sequence is coded as a level 4.1 bitstream, while a copy of the source is also deinterlaced to 1080p 60 frames/sec and then coded as level 4.2 progressive bitstream. Each deinterlaced progressive-coded sequence has the same temporal sample rate, but twice the lines and hence twice the pixel rate of the original 1080i interlace source sequence. On the test set selected of content known to be challenging for deinterlacers, the bitstreams produced by encoding with HEVC HM 8.0 in Main Profile RA fixed-Qp on deinterlaced progressive frame sequences average ~17% smaller according to the bdrate() macro than bitstreams on the same content coded as interlaced field sequences. In a separate experiment comparing the effect of deinterlacing before AVC encoding vs. deinterlacing after decoding, bdrate() suggests that better results are yielded by the former path of deinterlacing prior to encoding: AVC tests performed with x264 default MBAFF mode (also fixed-QP “rate control”), the deinterlaced progressive coded bitstream average 18% smaller than their MBAFF bitstream counterparts.

The unobserved fields in between each known field in the interlace source sequence is interpolated by a 2500+ operation per pixel prediction engine in the Demeler deinterlacer that blends motion compensated pixels from the adjacent known fields of opposite parity with the current field. The quality of the deinterlacer is judged by observers to fall into the “high” professional category of commercial products.

As illustrated in the accompanying PDF slides to this document, the first test compares HEVC coding of deinterlaced frame sequences against interlaced field sequences on the same content. For simulation speed reasons, AVC is applied in the second test which compares the effect of deinterlacing prior to encoding vs. deinterlacing after decoding.

As summarized in Table 1 below, the average size of the HEVC coded deinterlaced frame sequences average 17% smaller than HEVC fields sequence bitstreams. The list of sequences [1] below are the subset in the study that were captured from interlaced cameras or computer generated to challenge deinterlacers (BBC_sequence3, ants).

Table 1 shows the results of deinterlaced frame sequences vs. interlaced field sequences with HM 8.0 RA common configurations [2].

As shown in Table 2, a different test set (from Table 1) of progressive source material is first interlaced (by filtering and scaling each source frame into a simulated field using Meler), and then encoded with x264’s default mode [3] settings including MBAFF.

Table 2: pre-deinterlacing BD rate reductions over post-deinterlaced

Though overall pleasing to the viewer, interpolation of missing fields in deinterlacing introduces small amount of artificial spurious decision noise and jaggy edges at the pixel level that trigger HEVC HM to code more coefficients and smaller blocks than it would otherwise for native progressive sources. Suppression of such artifacts either by pre-processing or various encoder optimizations such as coefficient thresholding/decimation will likely improve deinterlace-before-encoding results. The sequences which show gains from deinterlacing prior to encoding are likely due to the number of motion-compensated interpolation steps tallying over 2500 operations per output pixel that provide smooth continuity for spatial and temporal video coding tools. Preliminary results not included here also suggest that adaptive quantization is a more appropriate mode for AVC and HEVC deinterlaced coding.

In order allow independent confirmation of our compression test results, isovideo is making the Demeler-deinterlaced output available for some well-known interlaced-captured sources. The links below provide access to files that were deinterlaced and then lossless-compressed using x264, in the matroska contrainer format. Lossless-compressed files are still very large, and must be downloaded before viewing. They can be converted using ffmpeg/avconv, or viewed with the freely available vlc and mplayer (with versions available for all major operating systems). Some HD sources (*pip_isovideo.mkv) were originally progressive, so they were interlaced before being deinterlaced.

Note: Since this paper was originally published in October 2012, the old test video results are no longer available for download.

• [1] Interlaced source sequences:https://media.xiph.org/vqeg/TestSeqences.
  
• [2] F. Bossen, “HM 8 Common Test Conditions and Software Reference Configurations”, JCTVC-J1100, July 2012.
• [3] x264 reference software:http://x264.nl/.

Follow:                                   Share:  
Copyright © 2018 isovideo LLC. All rights reserved.