To verify that the inter-block skip-connection and the guidence of MM-CU indeed enchances the frame restoration, we do an ablation study to compare the performance of RDN baseline [1] , PRN baseline, and MM-CU guided PRN (PRN+M). It is worth mentioning that we train a new model of RDN with 10 RDBs for comparison. We calculate the BD-rate of the sequences from class B to class E according the JCT-VC common test conditions under AI configuration. The result is shown in Table 1.
| CLASS | RDN [1] | PRN | PRN + M |
|---|---|---|---|
| CLASS B | -5.7% | -6.0% | -6.6% |
| CLASS C | -9.7% | -10.1% | -10.7% |
| CLASS D | -8.9% | -9.2% | -9.6% |
| CLASS E | -12.0% | -12.5% | -13.3% |
| Average | -8.7% | -9.1% | -9.6% |
We compare our proposed method with three previous SOTA methods under the AI configuration as well. The result is shown in Table 2 . It is shown that our method largely outperforms the three compared methods.
| CLASS | VRCNN [2] | DCAD[3] | DRN [4] | PRN + M |
|---|---|---|---|---|
| CLASS B | -4.3% | -3.4% | -3.8% | -6.6% |
| CLASS C | -5.0% | -4.6% | -7.5% | -10.7% |
| CLASS D | -5.4% | -5.2% | -7.3% | -9.6% |
| CLASS E | -6.5% | -7.8% | -10.7% | -13.3% |
| Average | -8.7% | -5.0% | -6.9% | -9.6% |
[1] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu, “Residual dense network for image super-resolution,” in Proc. IEEE Int’l Conf. Computer Vision and Pattern Recognition, 2018.
[2] Yuanying Dai, Dong Liu, and Feng Wu, “A convolutional neural network approach for post-processing in HEVC intra coding,” in Proc. International MultiMedia Modeling Conf., 2017.
[3] Tingting Wang, Mingjin Chen, and Hongyang Chao, “A novel deep learning-based method of improving coding efficiency from the decoder-end for HEVC,” in Proc. Data Compression Conf., 2017.
[4] Yingbin Wang, Zhu Han, Yiming Li, Zhenzhong Chen, and Shan Liu, “Dense Residual Convolutional Neural Network based In-Loop Filter for HEVC,” in Proc. IEEE Int’l Conf. Image Processing, 2018