Enhancing Video Frame Interpolation using Modified GAN Architecture: A Deep Learning Approach

Video frame interpolation (VFI) leverages Generative Adversarial Networks (GANs) to enhance interpolation quality, resulting in smoother transitions and slow-motion effects. Our modified Down-Up GAN architecture improves temporal connections, yielding higher-quality results. Experimental findings on the Vimeo-90K and UCF101 datasets, along with real-world CCTV footage, demonstrate substantial improvements in PSNR and SSIM parameters. Specifically, our proposed method achieves a PSNR of 38.89 and SSIM of 0.97 on Vimeo-90K, outperforming previous methods. Additionally, on UCF101, we attain a PSNR of 35.49 and SSIM of 0.905. Evaluation on CCTV footage yields a PSNR of 41.53 and SSIM of 0.982, highlighting robustness across datasets and real-world scenarios. These results indicate promising avenues for future research, including optimizing for real-time applications and exploring attention processes and recurrent neural networks (RNNs) to enhance accuracy.