Résumé:
This master's thesis investigates the implementation of single image super-resolution (SISR) algorithms. Two categories of image super resolution are considered: conventional methods based on interpolation and methods based on deep learning networks. Three interpolation methods are used, namely: Bicubic, bilinear and nearest algorithms. The implemented deep learning SR networks are: Very deep learning SR (VDSR), Enhanced Deep Super-Resolution Network Algorithm (EDSR) and Enhanced Super Resolution Generative adversarial network (ESRGAN). The considered SR algorithms are applied on the same and well known datasets, for the sake of comparison. The performance assessments are accomplished in terms of PSNR and SSIM in addition to the visual quality of the processed images. The obtained results indicate that deep learning SR algorithms offer significant improvements in LR image quality, outperforming other algorithms in terms of both PSNR and SSIM.
