Supplementary Experimental Results

 

Ablation Study

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.

 

CLASSRDN [1]PRNPRN + 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%

 

Comparison with Previous SOTA methods

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.

 

CLASSVRCNN [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%

 

 

Reference

[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