Comparing the Performance of Interpolative Algorithms with Neural Networks for Image Super-Resolution

Sammanfattning: The importance of graphics cannot be understated in the current era. Yet, due to technological limitations some sacrifices have to be made in order to save bandwidth, primarily the compression and downsampling of images. Unfortunately, these measures also entail loss of information and lower graphical fidelity when restoring and upscaling images to their original resolution. To mitigate this problem, there are graphical interpolation algorithms that are able to reduce visual inaccuracies caused by re-upscaling. However, these solutions are flawed as they do not restore lost information but merely smoothen the transitions between the remaining data points. In recent decades, neural networks have gained prominence in the field of image super-resolution, with one of their most promising features being the ability to reconstruct lost information. We believe that the potential of neural networks will usher in a new age of graphical upscaling. Thus, we decided to further investigate the capabilities of neural networks by conducting a comparative study between traditionally used interpolation algorithms and neural networks. Our research shows that neural networks have undeniable potential with superior efficacy and accuracy compared to interpolation algorithms, provided they are trained to process a designated category of imagery. Otherwise, some of the more advanced interpolation algorithms appear to be on par with certain neural networks in terms of quantitative measurements, but their qualitative performances fall short.