Performance Benchmarking of State-of-the-Art GAN-Based Video Super-Resolution Algorithms Using PSNR and SSIM Metrics

2026-01-19T15:23:27+00:00

Video super-resolution (VSR) has emerged as a critical research domain with extensive applications spanning surveillance, medical imaging, entertainment, and remote sensing. This study presents a rigorous and comprehensive evaluation of four state-of-the-art Generative Adversarial Network (GAN) architectures for video super-resolution: GFPGAN (Generative Facial Prior GAN), ESRGAN (Enhanced Super-Resolution GAN), TecoGAN (Temporally Coherent GAN), and RRDB-ESRGAN (Residual-in-Residual Dense Block ESRGAN). We conduct exhaustive experiments on the Low-Dose Video (LDV) benchmark dataset, employing a multi-faceted evaluation framework encompassing both distortion-based metrics (Peak Signal-to-Noise Ratio and Structural Similarity Index) and perceptual quality metrics (Learned Perceptual Image Patch Similarity and Natural Image Quality Evaluator). Additionally, we introduce temporal consistency analysis using optical flow warping error and inter-frame similarity metrics to assess motion coherence in reconstructed video sequences. Our experimental findings reveal that GFPGAN achieves the highest PSNR (34.052 dB) and SSIM (0.952), while TecoGAN demonstrates superior temporal consistency with the lowest temporal warping error (0.0234). Furthermore, we present comprehensive ablation studies examining the impact of architectural components, loss function configurations, and training strategies on reconstruction quality. Computational complexity analysis reveals significant variations in inference time and memory requirements across algorithms, providing practical guidance for deployment scenarios. This research contributes valuable insights for researchers and practitioners seeking optimal GAN-based solutions for video enhancement applications.

International Journal of Intelligent Systems and Applications in Engineering

Performance Benchmarking of State-of-the-Art GAN-Based Video Super-Resolution Algorithms Using PSNR and SSIM Metrics