Abstract

Video super-resolution (VSR) aims to reconstruct high-resolution video frames from their low-resolution counterparts while maintaining spatial details and temporal consistency. Unlike single-image super-resolution, VSR must exploit information distributed across neighboring frames, where motion, occlusion, blur, compression artifacts, and illumination variations can complicate reconstruction. Recent transformer-based restoration models have improved long-range temporal modeling, but their computational complexity and memory requirements remain barriers to real-time deployment. Existing VSR methods often face a trade-off between reconstruction quality and computational efficiency. Convolutional recurrent approaches provide efficient temporal propagation but may inadequately capture long-range dependencies, whereas transformer-based approaches model temporal relationships effectively but frequently require substantial computational resources. This study proposes a novel Temporal-Window Motion-Aware Transformer for Real-Time Super-Resolution (TWMAT-SR). The proposed framework combines lightweight convolutional feature extraction, motion-guided temporal alignment, localized temporal self-attention, and progressive feature reconstruction. A temporal-window attention mechanism restricts attention computation to informative neighboring frames, while a motion-aware alignment module dynamically compensates for inter-frame displacement. A recurrent feature-refinement pathway further propagates high-frequency information across successive frames without repeatedly processing the complete temporal sequence. The framework is optimized using a combined reconstruction, perceptual, and temporal-consistency objective. The experimental evaluation demonstrates that the proposed TWMAT-SR framework achieves 33.48 dB PSNR, 0.937 SSIM, and 0.110 LPIPS at the 50th training epoch on the REDS dataset for 4× video super-resolution. The model reaches an inference speed of 28.9 FPS, outperforming BasicVSR, VRT, and RVRT by 2.5, 12.8, and 9.2 FPS, respectively. The proposed method also obtains a temporal consistency error of 0.037, representing reductions of 30.2%, 22.9%, and 14.0% compared with BasicVSR, VRT, and RVRT. These results indicate that TWMAT-SR improves spatial reconstruction and perceptual fidelity while maintaining efficient inference and reduced temporal instability.

Authors

Mahesh Maurya¹, B. Sathananth²
St. John College of Engineering and Management, India¹, Cape Institute of Technology, India ²

Keywords

Video Super-Resolution, Transformer, Temporal Attention, Motion-Aware Alignment, Real-Time Video Restoration