Image super-resolution using artificial intelligence

• Main Author: Sendjaja, Dewi
• Other Authors: Yap Kim Hui
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/157462
• Institution: Nanyang Technological University

Cameras have undoubtedly grown more advanced but there are still some scenarios where the perceptual quality and resolution of the captured image is less than ideal. That is where super- resolution comes into play, which is the process of upscaling the resolution and improving the quality of an image. While traditional methods of super-resolution such as bicubic and bilinear interpolation exist, the field has gotten much more advanced with the help of deep learning. This project aims to explore deep learning super-resolution models and adapt them to a surveillance camera setting, which can be used in workspaces and even street views to observe blurry details. Initially, various state-of-the-art models were reviewed in order to determine the base model used for training. To better suit the application of surveillance super-resolution, a dataset was built using a combination of street scenes, factory CCTV footage as well as self-recorded footage using a CCTV camera. After training and extensive experiments using these datasets, a PSNR result of around 23-25dB is obtained using ESRGAN and SWINIR, with its SSIM ranging around 0.7. To improve the resulting image, post-processing in the form of applying bilateral filters was done. Efforts were also made to try and improve the model’s performance by tweaking the validation metrics and loss function of the model so as to include additional perceptual loss during training. Overall, although perfect restoration is not achieved, the resulting image still has promising quality and its visual pleasantness is vastly improved.