Enhancing Video Quality Using a Multi-Domain Spatio-Temporal Deformable Fusion Approach
Enhancing Video Quality Using a Multi-Domain Spatio-Temporal Deformable Fusion Approach
Garibaldi da Silveira Júnior,Gilberto Kreisler,2 Authors,Guilherme Corrêa
TLDR
The Multi-Domain Spatio-Temporal Deformable Fusion (MD-STDF) architecture is presented, which employs a multi-domain learning approach to enhance the quality of compressed videos across codecs like HEVC, VVC, AV1, and VP9.
Abstract
Video compression is essential for efficient data management but often introduces artifacts that degrade visual quality. This work presents the Multi-Domain Spatio-Temporal Deformable Fusion (MD-STDF) architecture, which employs a multi-domain learning approach to enhance the quality of compressed videos across codecs like HEVC, VVC, AV1, and VP9. By training on multiple domains, the model effectively adapts to diverse artifact patterns and compression scenarios. Experimental results show that MD-STDF achieves significant quality improvements, with average ∆PSNR gains of 0.764 dB for HEVC, 0.695 dB for VVC, 0.359 dB for VP9, and 0.228 dB for AV1. The model also demonstrated resilience under different compression rates, with BD-Rate values indicating that video quality can be efficiently restored in high compression scenarios for VP9 (-16.50%) and HEVC (-14.59%). Visual analysis shows a reduction in artifacts, resulting in perceptible improvements in subjective quality.