JVET-W0059 AHG11: A Deep In-Loop Filter Method [X. Zhang, C. Fang, D. Jiang, J. Lin (Dahua)]
This contribution presents a convolutional neural network-based in-loop filtering method with QP based models. Compared with VTM-11.0-NNVC, the proposed method reportedly shows on average {%}, {%}, and {%} BD-rate reductions for {Y, Cb, Cr}, under AI configuration.
Separate networks were used for luma and chroma
The luma input size is 160x160 (CTU plus 16 samples from neighboring blocks), chroma size half (420 downsampling).
The additional filter stage is placed between SAO and ALF.
The number of operations is reported in GMAC/CTU (and only for luma), should be converted to KMAC/Pixel for comparability with other proposals.
Only partial results were available by the time of presentation:
All Intra Main 10 | |||||
Over VTM-11.0-NNVC | |||||
Y | U | V | EncT | DecT | |
Class A1 | #NUM! | #NUM! | |||
Class A2 | #NUM! | #NUM! | |||
Class B | #NUM! | #NUM! | |||
Class C | -4.55% | -13.21% | -15.06% | 1731% | 412921% |
Class E | -5.64% | -14.76% | -15.05% | 4044% | 618997% |
Overall | #NUM! | #NUM! | |||
Class D | -5.23% | -12.27% | -15.56% | 2751% | 303132% |
Class F | 1.45% | -2.37% | 2.54% | 2956% | 609187% |