Ensemble Generative Adversarial Network with Self-Attention and Transfer Learning for Frame-Wise Video Super-Resolution

This paper introduces an ensemble generative adversarial network with self-attention and transfer learning for high-fidelity frame-wise video super-resolution (EGANSAT-VSR). The proposed model integrates a self-attention mechanism to capture long-range spatial dependencies and strengthen high-frequency detail restoration, while its hybrid GAN architecture enhances perceptual sharpness and visual consistency across reconstructed frames. A pre-trained DIV2K backbone is further refined using transfer learning to isolate and quantify the impact of self-attention within the generator. Experimental evaluations on Set14, Vimeo-90K, and Vid4 demonstrate that the EGANSAT-VSR model achieves competitive reconstruction quality without explicit temporal modeling, attaining 42.789 dB PSNR and 0.9866 SSIM on Vid4.