This is a short working example of face blurring that seems to produce far less double-jpg'ing artifacts in regions that are not blurred compared to alternate methods. This script is not meant as a turnkey solution to be used for other projects, but rather is a demonstration for developers to re-purpose. I'm posting this in the hopes that it is useful, but will not be monitoring github issues.

If you want to use this in your own project, I'd suggest pulling out the following two functions:

• blur (which is based on Kaiyu Yang's version here: https://github.com/princetonvisualai/imagenet-face-obfuscation/blob/main/experiments/blurring.py)
• handle(...), which demonstrates how to mitigate double jpg'ing. handle is also a few small modifications away from being able to be easily parallelized with multiprocessing

Do not use:

• handleBad -- this demonstrates a data processing that causes double jpging artifacts.

You can try the script on either calder.jpg or crop5.jpg. The script will generate the blurred version. If showDiff is set to true, the script will also generate a version that does not try to avoid double jpg'ing artifacts as well as generate difference maps.

This script mitigates double jpg artifacts, but doesn't totally eliminate them. The re-decoding/re-encoding step produces slight differences even with the same quantization scheme. I suspect this is due to differences in rounding modes or other decisions that are left up to the library, but I'm not a jpg expert.

If totally eliminating double jpg-ing artifcats is important (or you need a more turn-key solution), please use David Pinto's jpegblur

Thanks to:

• Kaiyu Yang, whose code was the starting point for this
• David Pinto, who gave a lot of helpful advice and who provided a fix that preserves metadata

Dependencies:

• Python Image Library
• Numpy