科研之友
微信
新浪微博
Facebook
分享链接
摘要
Nonblind image deconvolution restores the clear image from a blurred one under a known blur kernel, whose recent development has been boosted by supervised deep learning. Motivated by the inaccessibility of ground-truth images for supervised learning in many application domains, such as scientific imaging, this paper studies the unsupervised knowledge transfer problem for nonblind image deconvolution, which aims at adapting a deep model pre-trained on a source domain, to a ground-truth-scare target domain where image contents or blur kernels are distinct from that of the source domain. We propose to conduct the knowledge transfer regarding both images and kernels, by leveraging the model being adapted itself to generate pairs of a pseudo ground-truth image and a blurred image for self training. The proposed method neither accesses source-domain data, which avoids privacy issues, nor accesses target-domain ground-truths, which avoids ground-truth collection. Its effectiveness is demonstrated with the experiments on three deblurring tasks in different domains.
