Deep learning for image decensoring or how the search for the implementation of a certain function in TensorFlow led to some thoughts on reversing censored images.
Well, I didn’t came across anything helpful for my initial problem but I came across DeepCreamPy. WTF? Hentai (or manga porn) is censored in Japan. I’m kind of curious if mentioning DeepCreamPy puts this blog post or even this website on some kind of “adult warning/filter lists”. While I don’t think it is rather challenging to “complete” painted pictures using deep learning (this goes for image restoration in general), there are some other use cases that came to my mind rather quickly:
- every interview with people wearing a balaclava (considering 4k or even 8k video streams it might even be possible to recover exact keypoints for biometric identification)
- license plates restoration (they have to be censored for GDPR compliance in Germany (7 ways to do so ;)).
- in general every image with parts missing
- censored documents (given enough input documents; I remember a case where authors (names) were censored in German research reports on deep geologic repositories (for nuclear spent fuel) but the the titles of research reports were published. However, in an English text book on the same topic, there was no censorship of the same references.)
In general, this is nothing new. We as humans do it all the time (part of our vision system is to interpolate reality). However, the tasks I mentioned above are nothing new either. However, they are still relatively time consuming and therefore expensive. However, with deep learning it is easier and much faster. Similar to cryptography, censorship will be broken after x years it was introduced. This is simply something everyone should remember at all time.