Real-World Image Deblurring via Unsupervised Domain Adaptation

Hanzhou Liu, Binghan Li, Mi Lu, Yucheng Wu

Abstract: Most deep learning models for image deblurring are trained on pairs of clean images and their blurry counterparts, where the blurry inputs are artificially generated. However, it is impossible for these synthesized blurry images to cover all the real-world blur. Even in two synthetic datasets, the blur type, illumination, and other important image parameters could be different. Consequently, the performance of most existing deblurring models decreases when applied to real-world images and artificial blurry images from a different synthetic dataset. Very few previous deblurring works consider the gap among blurry images from different domains. Inspired by the current success of unsupervised domain adaptation (UDA) on image classification tasks, we develop, UDA-Deblur, a novel deblurring framework that utilizes domain alignment to attenuate effects of the aforementioned gap. In our work, channel attention modules are adopted to exploit the inter-channel relationship for features; multi-scale feature classifiers are designed to discriminate domain difference. UDA-Deblur is trained adversarially to align the feature distributions of the source domain and the target domain. We provide adequate quantitative and qualitative analysis to show the state-of-the-art performance of UDA-Deblur. Firstly, we evaluate the proposed UDA-Deblur on synthesized datasets related to real-life scenarios, which achieves satisfying deblurring results. We further demonstrate that our approach also outperforms prior models on real-world blurry images. For a persuasive comparison, we carefully design experiments on GoPro, HIDE and ReaBlur datasets. More importantly, this is the first work considering real-world image deblurring from a feature-level domain adaptation perspective.