ScrVSR: Screen Sharing Video Super-Resolution

This paper addresses the challenges of Video Super-Resolution (VSR) for screen sharing videos, focusing on 3× upscaling and compression artifact removal. The dataset includes screen recordings from various applications, simulating real-time user interactions. Input videos are low-resolution (LR) and compressed using the H.265 codec with six different Quantization Parameters (QP-17 to QP-37), reflecting typical screen sharing conditions. We propose an implicit transformer based model optimized using a character-aware, perceptual-quality-driven loss function, and trained on text-dense image patches. Specifically, during training, if a ground-truth patch contains more than 20% textual content, both the ground-truth and its corresponding low-resolution patch are included to enhance text-preserving performance. Experimental results show that our model consistently outperforms state-of-the-art baselines, achieving up to ${\color{Red}\text{+1.21 dB}}$ higher PSNR and significantly lower character error rates across all QPs. Notably, at QP-17, our method achieves a PSNR of ${\color{Red}\text{29.54}}$ and a CER of ${\color{Red}\text{0.1927}}$, compared to (27.18 / 0.2137) and (28.31 / 0.2724) for competing methods. These improvements result in visibly clearer text and more consistent screen content reconstruction.