Deep Blind Un-Supervised Learning Network for Single Image Super Resolution

• Published in: 2021 IEEE International Conference on Image Processing (ICIP) pp. 1789 - 1793
• Main Authors: Yamawaki, Kazuhiro, Han, Xian-Hua
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01-01-2021
• Subjects: Benchmark testing; Deep learning; Degradation; degradation model; generative network; Image resolution; Network architecture; Performance gain; Single image super resolution; Training; un-supervised learning; unknown kernel
• ISSN: 2381-8549
• DOI: 10.1109/ICIP42928.2021.9506783

Deep learning based methods have recently made significant progress in image super resolution (SR) field and lead to great performance gain in terms of both effectiveness and efficiency. Most of the current methods have struggled to design more complicated and deeper network architectures and aimed to learn a good LR-to-HR mapping with the previously prepared training sample pairs under a fixed degradation model (Blurring and down-sampling operations) such as bicubic dawn-sampling. However, these methods are generally implemented in a fully-supervised way with largescale training dataset, and are hardly generalized to most real scenarios with unknown and complicated degradation model. This study proposes a blind un-supervised learning network for automatically estimating the degradation operations in single SR problem, where the blurring kernel (operation) is unknown. Motivated by the considerable possessed image priors in the network architectures, we construct a generative network for simultaneously learning the inherent priors of the latent high resolution (HR) image and the degradation operations with the under-studying low-resolution (LR) observation only. Specifically, we exploit a general depth-wise convolutional layer for both approximating a special degradation and automatically learning any complicated blurring kernel in a general SR framework, and propose an end-to-end HR image learning network from its LR observation. Experimental results on two benchmark datasets validate that our proposed method achieve promising performance under the unknown degradation model.