A Progressive Meta-Algorithm to Synthesize Very Large Super Resolution Images with Fine Details

Abstract

This work presents a Meta-Algorithm based on Progressive GANs that is capable of generating very large super resolution images with fine details. Prior works for patch based super resolution lack contextual information which limits the size of the overall image that can be synthesized with adequate context. We present a Meta-Algorithmthat is capable of generating super resolution images of any size with adequate context.Our Progressive Meta-Algorithm is based on prior works in progressive GANs but is reformulated in a way that allows for upsampling of any size image by overcoming the theoretical memory limitations while providing contextual cues across varying scales. Our novel Progressive Meta-Algorithm pipeline allows patch-based image upsampling using context across multiple scales by progressively training multiple GAN models with established connection between one model to another model. In this paper, the experimental design essentially focuses on data preprocessing, training & testing the models obtained from our novel progressive Meta-Algorithm pipeline along with postprocessing experiments to address the minor image shifts and coloring issues.We have successfully addressed the limitations of generating very large images using GANs by generating images of size 4096x4096 by employing our novel Progressive Meta-Algorithm by overcoming the theoretical memory issues, contextual issues or missing details in the generated images.