New version Lavtools code.

yuvdeinterlace.c

Double frame rate de-interlacer

A non destructive deinterlacer. Converts to a double frame rate, progressive yuv stream for further processing by temporal based non interlace aware filters. Then re-interlaced or frame rate converted before encoding.

This program uses an experimental detection algorithm to detect interlaced pixels and then uses cubic interpolation to replace pixels that are suffering comb effect. Image data is not lost as the frame rate is doubled. This adaptive algorithm is quite effective, however not ideal. Noisey artefacts may be apparent in (some) most material.

I have been looking into implimenting a 4 point DFT on the data and looking at the amplitude of the high (/2) frequency component. This has lead to better results but appears to still have issues at the boundary of interlace and non interlace material.

The above code could be extended to an arbitary number of pixels, comparing the /2 (interlace) frequency with the /4 frequency.

This release of the code I have played with comparing both the high frequency and the next frequency component. The high frequency shows areas of interlace however also detects edges. The second frequency component detects edges. By balancing these two results, by trial and error, I beleive to have come up with a better detection algorithm. However non interpolated interlace artefacts are still apparent. This is the best results so far.

Just when I thought I tuned the DFT algorithm to the best parameters. I came up with another method that simply looked for a "greater than, smaller than, greater than" pattern and as a result are now detecting interlace much more accurately than before. I also added an additional test on the AC value to eliminate non noticable interlace (or noise). This version is now available.

I am extremely happy with the results. Although not perfect, there are false positives, and a few false negatives. Of course I will be fine tuning this filter even more over time. For now I would use it as part of a yuv filter chain for my production videos.

TODO

Write a pixel merging algorithm for conversion of 25i material to 25p (rather than 50p).
Look at ways to perform anti-aliasing for fixing up nearest neighbour de-interlace filters. These sorts of filters should be shot, as it's quite easy to write a linear interpolator, not to difficult to write a cubic interpolator, and look I just wrote an adaptive type de-interlacer. I can't believe I saw this kind of de-interlace filter used on an episode of Top Gear the other night. Disgraceful!
Provide a screenshot...

Ideas

I have been experimenting with training a neural net to *learn* what the missing pixel may be from the surrounding pixels. My first attempts have not worked, if it is trained from images with many uphill lines, then uphill interpolation looks fantastic but downhill is very jaggered.

I am thinking about increasing the number of hidden layers in the neural net to see if it improves interpolation. The code is in java, I'll just need some encouragement to look at it again.

This part of the code would replace the cubic interpolator. Detection would still be separate algorithm, however comparing the interpolated pixel with the real pixel and replacing the real pixel with the interpolated if it is different by a threshold amount. May prove a better algorithm.

• 29th April 2008 Tuned the DFT parameters to optimal. However artefacts are still apparent.
• 13th April 2008 Implemented a 4 point DFT interlace detection algorithm.
• 10th April 2008 Implemented a cubic interpolation algorithm. Do you know how hard it is to find programmatic information about this on the web?
• 6th April 2008 Optimised the code for better performance.
• October 2007 Changed the code to an adaptive deinterlace filter. With linear interpolation. The re-interlace code is probably not needed as my yuvafps does a better job.
• 26th April 2005. Fixed a bug which incorrectly detected the end of file, creating more frames in the output.
• 9th April 2005 New features which now produces a full height frame, by line duplication. Also contains code from yuvdeprogress, to re-interlace the stream.

Example

by default yuvdeinterlace will double the frame rate and convert fields into full height frames. No arguments needed.