Optical Flow-Guided Generative Adversarial Network For Video Super-Resolution

Abstract

Video Super-Resolution has recently gained significant attention due to its importance in surveillance, medical imaging, multimedia streaming and video restoration applications. However, reconstructing high-resolution video frames from low-resolution sequences remains challenging because the network must recover fine spatial details while also preserving temporal consistency between consecutive frames. This paper presents a Flow-Guided Generative Adversarial Framework for Video Super-Resolution that integrates optical flow-based temporal alignment, multi-frame feature fusion, and adversarial perceptual learning to enhance video reconstruction quality. The proposed framework employs a pretrained RAFT network to estimate inter-frame motion and align neighboring frames to a reference frame. The aligned frames are subsequently fused and processed through a residual convolutional reconstruction network to generate high-resolution outputs. To improve perceptual quality and texture fidelity, a PatchGAN discriminator is incorporated together with perceptual and adversarial loss functions. The framework achieves a PSNR of 28.58 dB and SSIM of 0.8672 on REDS validation data, while obtaining 25.42 dB PSNR and 0.8253 SSIM on VID4. Ablation studies further confirm the contributions of flow-guided alignment, multi-frame feature fusion, and adversarial perceptual learning to overall reconstruction performance. Under a fixed, modest training budget, the proposed framework achieves competitive perceptual quality and structural similarity relative to larger state-of-the-art architectures, indicating a favorable trade-off between reconstruction accuracy and perceptual realism for lightweight video super-resolution.