Plug-and-Play Image Restoration with Locally Structured Unitary Network Denoiser

This study proposes an image restoration method based on a Plug-and-Play (PnP) algorithm. A contribution of this work is to eliminate the need for supervised learning for the plugin denoiser. While recent advances in measurement technologies have enabled the acquisition of large-scale physical data, such data are often affected by noise, necessitating effective restoration techniques. A previous work has shown that PnP frameworks incorporating denoising convolutional neural networks (DnCNN) can achieve high restoration accuracy. However, they require extensive training data and may suffer from overfitting when such data are insufficient. To overcome this limitation, we propose to adopt Locally Structured Unitary Network (LSUN), a training-free denoiser proposed by Godage et al., within the PnP framework. The effectiveness is demonstrated in two-dimensional image restoration by showing that the proposed method achieves performance comparable to supervised approaches, even without supervised learning.