Reading: VRCNN-BN — Variable-fi lter-size Residue-learning Convolutional Neural Network with Batch Normalization (Codec Filtering)
Outperforms VDSR, DCAD, VRCNN, and Residual-VRN, Over 13% BD-Rate Reduction
In this story, Variable-fi lter-size Residue-learning Convolutional Neural Network with Batch Normalization (VRCNN-BN), by Shanghai University, and Shanghai University of Engineering Science, is briefly described. It is a post-processing out-of-loop filter instead of an in-loop filter. This is a paper in 2020 ACCESS which is an open access journal with high impact factor of 4.098. (Sik-Ho Tsang @ Medium)
Outline
- VRCNN-BN Network Architecture
- Experimental Results
1. VRCNN-BN Network Architecture
- ReLU is used except the last layer.
- Different from VRCNN, VRCNN-BN uses Batch Normalization (BN) which is originated from Inception-v2.
- (It has been a discussion on video restoration whether BN should be applied or not. Some with BN obtains better results, some without BN obtains better results…)
- MSE is used as loss function.
- Separate models are used for luma and chroma respectively.
- For each QP, a separate model is also trained. Thus, there are in total 8 models.
2. Experimental Results
2.1. CS-PSNR as Evaluation
- Color-Sensitive PSNR (CS-PSNR) is used for evaluation to calculate BD-rate:
- where PY=0.685, PU=0.137, PV=0.178, which are the percentages of the color components.
- HM16.19 is used.
2.2. BD-Rate
- For luma, BD-rate reduction, on CS-PSNR, of 10.3%, 8.9%, 13.1% and 11.8% for RA, AI, LDP and LDB confi gurations are obtained.
2.3. Visual Quality
- Ringing artifacts are reduced by using VRCNN-BN.
2.4. SOTA Comparison
- As shown above, VRCNN-BN outperforms VRCNN by large margin.
- VRCNN-BN also outperforms DCAD and Residual-VRN as shown above.
- VRCNN-BN outperforms VDSR by large margin.
During the days of coronavirus, let me have a challenge of writing 30 stories again for this month ..? Is it good? This is the 27th story in this month. 3 stories to go. Thanks for visiting my story..
Reference
[2020 ACCESS] [VRCNN-BN]
A CNN-Based Post-Processing Algorithm for Video Coding Efficiency Improvement
Codec Filtering
JPEG [ARCNN] [RED-Net] [DnCNN] [Li ICME’17] [MemNet] [MWCNN]
HEVC [Lin DCC’16] [IFCNN] [VRCNN] [DCAD] [MMS-net] [DRN] [Lee ICCE’18] [DS-CNN] [RHCNN] [VRCNN-ext] [S-CNN & C-CNN] [MLSDRN] [Double-Input CNN] [B-DRRN] [Residual-VRN] [Liu PCS’19] [QE-CNN] [EDCNN] [VRCNN-BN]
3D-HEVC [RSVE+POST]
AVS3 [Lin PCS’19]
VVC [Lu CVPRW’19] [Wang APSIPA ASC’19]
