Add a deinterlacing model - Githubissues

MIR-MU / ffmpeg-tensorflow

Dockerfile for FFMpeg with Libtensorflow

35 stars 14 forks source link

Add a deinterlacing model #13

Open Witiko opened 3 years ago

Witiko commented 3 years ago

The dnn_processing filter of FFmpeg can do more than just superresolution. FFmpeg supports a number of built-in deinterlacing algorithms (bwdif, estdif, kerndeint, nnedi, w3fdif, yadif), but none of them can be pretrained. In https://github.com/MIR-MU/ffmpeg-tensorflow/issues/6#issuecomment-763350720, @dlueth mentioned the lszhuhaichao/Deep-Video-Deinterlacing model, which implements a paper by Zhu et al. (2017). The code is written in Tensorflow and pre-trained models are available.

We should:

[ ] write Python code that would load the pretrained models and convert them to the format expected by libtensorflow,
[ ] download and convert the pretrained models during script/produce-sr-models, and
[ ] add visual comparisons and code examples to the readme.

dlueth commented 3 years ago

@Witiko perhaps this repo can also be a little help: https://github.com/HENDRIX-ZT2/Deep-Video-Deinterlacing