Deep Unsupervised Blind Learning for Single Image Super Resolution
Kazuhiro Yamawaki, Xian‐Hua Han · 2022
This study proposes a deep unsupervised blind learning SR framework by leveraging the powerful representation capability of deep-learning but without the requirement of prior training using external data. There are two types of unknown factor, the latent high-resolution (HR) images to be estimated and the imaging conditions (also called the degradation operations, i.e., the blur kernel and downsampling), which greatly affect performance in the SR problem. Accordingly, we present two generative networks to model the deep priors of the latent HR image and blur kernel and a specially designed degradation block to implement the imaging procedure, which constructs an end-to-end blind learning network to jointly estimate a plausible SR image and the degradation information using only the LR observation. In particular, we exploit an encoder-decoder architecture to generate the latent HR image, a fully-connected network to learn the blur kernel, and a specific depth-wise convolutional layer to realize the degradation model. Experiments on several benchmark datasets with different imaging conditions demonstrate the superiority and generalization of our method over the state-of-the-art convolutional neural network-based SR methods, as well as previously unsupervised SR methods.