JVET-W0113 AHG11: Neural network based in-loop filter [L. Wang, W. Jiang, X. Xu, S. Liu (Tencent)]
In this contribution, a neural network based in-loop filter is exploited to replace the Deblocking and SAO modules. Compared with EE1 anchor (VTM11+V0056), the proposed method reports {4.73%, 10.73, 11.51%} and {3.54%, 10.99%, 10.24%} BD-rate savings with AI and RA configurations, respectively.
In order to improve the quality of reconstructed image, neural network models are trained for I slice and B slice, separately. In the in-loop filter process, Deblock and SAO are replaced by the proposed NN filter.
This is operated before ALF.
It has a similar architecture as in JVET-W0111, but larger (621 kMAC/pixel).
The performance is better for AI than RA, which is different from other proposals.
An interesting aspect could be the 5x5 convolution in the parallel path. It would be interesting if that is an element which gives higher intra gain.
It was agreed to investigate this in an EE.
NN related HLS signalling (0)
No contributions were noted in this area.
AHG12: Enhanced compression beyond VVC capability (31)
General (2)
Contributions in this area were discussed in session 22 at 0610–0700 UTC on Thursday 15 July 2021 (chaired by JRO).