L1-STDF: L1-Norm Based Filter Pruning for Efficient Compressed Video Quality Enhancement

Abstract

The Spatio-Temporal Deformable Fusion (STDF) network has shown strong performance in video quality enhancement (VQE) by exploiting multi-frame information through spatiotemporal deformable convolutions. However, its high computational complexity often restricts real-time processing on resourceconstrained devices. While previous pruning approaches reduced theoretical complexity, they frequently introduced practical implementation bottlenecks, such as the need to decompose filters, which limited actual inference acceleration on standard hardware. To address these challenges, we propose L1-STDF, a hardwarefriendly variant of the STDF architecture that employs L1-norm structured filter pruning. By removing entire convolutional filters while respecting structural dependencies, L1-STDF preserves the original dense structure for direct deployment. Experimental results on the MFQEv2 dataset demonstrate that L1-STDF achieves a 68.27% reduction in parameters and a consistent 66.40% decrease in GFLOPs, with a marginal PSNR drop of only 0.1 dB. Hardware-level evaluations on an NVIDIA RTX A6000 GPU reveal that these theoretical savings translate into an average increase of 65.8% in frames per second (FPS) and an average energy saving of 40.7% per frame across all tested resolutions. These results validate L1-STDF as an efficient solution for enabling real-time and energy-efficient VQE on edge devices.