Deep Learning-based Super-Resolution on the Cloud: Focus on Face and Text Enhancement

Real-ESRGAN and SwinIR are two deep learning models for Single-Image Super-Resolution (SISR), which attempt to address real-world scenarios for image enhancement. However, the pre-trained models do not effectively handle LR images containing human faces and text. An experiment is conducted to expand upon the training performed in their respective studies and improve the image enhancement using a cloud computing environment. Traditional image quality metrics, Peak Signal-toNoise Ratio (PSNR), and Structural Similarity (SSIM), are used to objectively evaluate the image quality. Three learning-based perceptual metrics, the Blind / Referenceless Image Spatial Quality Evaluator (BRISQUE), Naturalness Image Quality Evaluator (NIQE), and the Learned Perceptual Image Patch Similarity (LPIPS), are also incorporated to assess how images would be subjectively perceived based on human perception. To evaluate the model performance of Real-ESRGAN and SwinIR, specifically for face and text images, both traditional and perceptual metrics are taken into consideration, in addition to the cost associated with model training using Microsoft Azure. The findings show that with additional fine-tuning, SwinIR has slightly improved PSNR and SSIM values while taking less training time compared to Real-ESRGAN at the cost of perceptual quality.