Learning Face Deblurring Fast and Wide

Portrait images and photos containing faces are ubiquitous on the web and the predominant subject of images shared via social media. Especially selfie images taken with lightweight smartphone cameras are susceptible to camera shake. Despite significant progress in the field of image deblurring over the last decade, the performance of state-of-the-art deblurring methods on blurry face images is still limited. In this work, we present a novel deep learning architecture that is designed to be computationally fast and exploits a very wide receptive field to return sharp face images even in challenging scenarios. Our network features an effective resampling convolution operation that ensures a wide receptive field from the very first layers, while at the same time being highly computationally efficient. We also show that batch normalization prevents networks from yielding high-quality image results and introduce instance normalization instead. We demonstrate our architecture on face deblurring as well as other more general scenes. Extensive experiments with state-of-the-art methods demonstrate the effectiveness of our proposed network, in terms of run-time, accuracy, and robustness to ISO levels as well as gamma correction.