PROJECT TITLE :

Multiple Cycle-in-Cycle Generative Adversarial Networks for Unsupervised Image Super-Resolution

ABSTRACT:

The single picture super-resolution problem has been extensively investigated with the aid of convolutional Neural Networks (CNN). To map low-resolution (LR) images into high-resolution (HR) images, most of these CNN-based approaches rely on learning a model to downsample an HR picture with an already-known model. However, when the down-sampling method is unclear and the LR input is deteriorated by sounds and blurring, it is impossible to obtain the LR and HR image pairings for classical supervised learning. We propose a multiple Cycle-in-Cycle network structure based on the recent unsupervised imagestyle translation applications using unpaired data, inspired by the recent unsupervised imagestyle translation applications using unpaired data. New network cycles are inserted sequentially in order to super-resolve the intermediate output of the first cycle, which has a well-trained x2 network model in place. The total number of up-sampling cycles varies depending on the various elements (x2, x4, x8). End-to-end training ensures that the desired HR output is achieved. A comparison of our suggested method's quantitative and qualitative outcomes shows that it is on par with the most current supervised models.