Super-Resolution with Vision Transformers (SwinIR) for Low-Resolution CCTV Image Enhancement

The use of low-quality CCTV recordings is a drawback to the rising demand for video surveillance for higher-quality video. Traditional super-resolution methods cannot reconstruct high details accurately and will have inaccuracies when used for security purposes. In this study, we implement a Super-Resolution using SwinIR (Swin Transformer-Based Super Resolution) to super-resolve low-quality CCTV footage to produce much higher-quality images without losing key structural details. The approach uses deep learning feature extraction with self-attention to improve upon traditional CNN-based super-resolution techniques with a better-resolution reconstructions. The DIV2K dataset was used to train the models so that they can be robust and generalizable. The experimental results demonstrated how SwinIR had better performance than SRCNN and state-of-the-art techniques in terms of Peak Signal-To-Noise Ratio (PSNR) and Structure Similarity Index (SSIM) between restored and high-quality images for images that were super-resolved.