In this paper, we propose a fully convolutional networks for iterativenon-blind deconvolution We decompose the non-blind deconvolution problem intoimage denoising and image deconvolution. We train a FCNN to remove noises inthe gradient domain and use the learned gradients to guide the imagedeconvolution step. In contrast to the existing deep neural network basedmethods, we iteratively deconvolve the blurred images in a multi-stageframework. The proposed method is able to learn an adaptive image prior, whichkeeps both local (details) and global (structures) information. Bothquantitative and qualitative evaluations on benchmark datasets demonstrate thatthe proposed method performs favorably against state-of-the-art algorithms interms of quality and speed.
展开▼
