IRBFusion: Diffusion-Based Blind Image Super Resolution Using Unsupervised Learning and Bank of Restoration Networks

Image super resolution focuses on increasing the spatial resolution of low-quality images and enhancing their visual quality. Since the image degradation process is unknown in real-life scenarios, it is crucial to perform image super resolution in a blind manner. Diffusion models have revolutionized the task of blind image super resolution in view of their powerful capability of producing realistic textures and structures. Design of the condition network is a key factor for diffusion models in providing high image super resolution performances. In this regard, we develop an effective image restoration bank by using a three-stage learning algorithm based on the idea of unsupervised learning, and feed its results, wherein visual artifacts are remarkably suppressed, to the condition network. The use of the unsupervised learning in the design of our image restoration bank guarantees that both diverse contextual information of visual signals, as well as, different degradation operations are considered for the task of blind image super resolution. Further, we guide the feature generation process of the condition network in such a way that the fidelity of the feature tensors produced for the task of image super resolution remains high. The results of extensive experiments show the superiority of our method over the state-of-the-art blind image super resolution schemes in the case of various benchmark datasets.