Image Deblurring Using Deep Multi-Scale Distortion Prior
Irina Kim, Dongpan Lim, Youngil Seo, Jeongguk Lee, Wooseok Choi, Seongwook Song · 2022 IEEE International Conference on Image Processing (ICIP) · 2022
Deep neural networks have recently advanced state-of-the-art in motion deblurring. However, non-uniform non-blind image deblurring has not been studied in depth. State-of-the-art methods shows improvement over conventional algorithms, but they are still not feasible for mobile deployment. Having informative prior information could improve performance of non-uniform deblurring. In this work, we propose a new deep framework that allows extracting spatially variant latent feature to Distortion Prior map from a pair of calibration sharp-blur images, without having to capture or model training dataset. We propose to use multi-scale Distortion Prior map that can fully utilize spatially variant information in the further restoration via multi-scale attention mechanism. Unlike prior art, we use image pyramid at decoder side, by fusing its fine level with coarse level of feature map via level attention and by injecting Distortion Prior at various resolution levels. Experiments show that proposed network outperforms state-of-the-art deblur networks both in terms of image quality and inference time. We demonstrate that proposed framework can successfully deblur non-uniform, non-blind applications, such as defocus blur removal. Being computationally efficient, it is feasible for mobile deployment.