Patch-based Neural Video Compression without Blocking Artifact

다국어 초록 (Multilingual Abstract)

While Learned Video Coding (LVC) models have demonstrated superior compression performance over conventional codecs, their high peak memory usage limits application to high- resolution videos. This paper presents a novel patch-based decoding method that addresses this limitation without compromising coding efficiency and visual quality. We introduce two key contributions; 1) asymmetric optimal patch-overlap based on a quantitative analysis of CNN’s receptive fields and its relation to blocking artifacts and 2) maximum optical flow magnitude- based temporal information overlapping. We validate our approach on three distinct LVC models—MobileCodec, Scale-Space Flow, and DCVC-DC—representing different complexity levels and operational characteristics. Experimental results demonstrate significant peak memory reduction without BD-rate performance degradation and blocking artifacts near patch boundaries, establishing a practical pathway for high-resolution learned video compression. Also, our method enables the first successful application of DCVC-DC to 4K video sequences on desktop GPU while maintaining compression performance.

목차 (Table of Contents)

• 1. Introduction 1
• 2. Related works 3
• 2.1. Learned Image Coding (LIC) 4
• 2.2. Block-based and Patch-based LIC models 5
• 2.3. Learned Video Coding (LVC) 7
• 3. Choosing the asymmetric optimal patch-overlap size 9
• 3.1. Single convolutional kernel and transposed convolutional kernel’s receptive field 10
• 3.2. Multi-layer transposed convolutional kernel’s receptive field 11
• 4. Design of patch-based learned video coding 12
• 4.1. Selection of components for patch-based operation 12
• 4.2. Asymmetric patch-overlap 13
• 4.3. Maximum optical flow magnitude-based temporal information overlapping 13
• 5. Experimental results 17
• 5.1. Test environment and settings 17
• 5.2. Performance comparison of different coding approaches on 4K video 18
• 5.3. Comparative analysis of DCVC-DC performance on JVET Class A sequences 21
• 5.4. Ablation studies 22
• 5.4.1. Impact of asymmetric patch overlap 23
• 5.4.2. Effect of maximum optical flow magnitude-based temporal information overlapping 24
• 6. Conclusion 24
• 7. References 26
• Appendix 29
• A. VTM experimental settings 29
• B. Subjective quality analysis 30