Abstract
Compressed videos frequently present artifacts that compromise visual quality. Deep learning models have demonstrated significant effectiveness in mitigating such distortions. In this study, we introduce 3D-STDF, an architecture based on the well-known Spatio-Temporal Deformable Fusion (STDF) and augmented with 3D convolutions to more effectively model temporal dependencies across video frames. Furthermore, we refine the Quality Enhancement (QE) module by integrating residual blocks, thereby enabling the extraction and representation of more intricate spatial features. Experimental results indicate that models based on the 3D-STDF architecture achieved an overall average improvement of up to 0.607 dB in PSNR, clearly outperforming previous STDF-based solutions.