Efficient and Effective Super Resolution for Single Images using Deep Learning

Abstract

Super resolution is a class of techniques that aims to increase the resolution of an image. Commissioned by the company Stapes IT this thesis aims to create an effi- cient and effective super resolution method for single images using deep learning. To tackle this issue, a comparison between different deep learning super resolution solutions had to be made to find the best framework to improve upon. It was found Real-ESRGAN is the best starting framework for its fast inference and high quality output. To improve this framework to the needs of Stapes IT the option of con- tinuous floating point upscaling factors was added. This was done by cascading the existing factor 4 and factor 2 models to a factor 8. To quantitatively measure the quality of the models an image data set relevant to the company was created and tested using the Peak Signal to Noise Ratio (PSNR) and Multi-Scale Structural Similarity Index Measure (MS-SSIM) metrics. For both these metrics, a higher score indicates a better result. Comparison to the existing SDMD model on the custom data set shows our cascaded model performs well with PSNR and MS-SSIM scores of 20.794 and 0.836 compared to 18.042 and 0.730 for SDMD. Using a combination of Lanczos4 interpolation and linear image blending the model is able to upscale to any floating point value between 1 and 8 from the existing 2 models. Tested on the same custom data set the models show a slowly declining average MS-SSIM score linearly correlated to the height of the upscaling factor starting with 0.984 for an upscaling factor of 1.5 till 0.836 for factor 8. We conclude that our method proves that creating super resolution images with continuous upscaling factors is practical, delivering decent high resolution images and can be implemented in a digital environment for commercial and practical use.