Abstract:Generic deep learning (DL) networks for image restoration like denoising and interpolation lack mathematical interpretability, require voluminous training data to tune a large parameter set, and are fragile during covariance shift. To address these shortcomings, for a general linear image formation model, we first formulate a convex optimization problem with a new graph smoothness prior called gradient graph Laplacian regularizer (GGLR) that promotes piecewise planar (PWP) signal reconstruction. To solve the posed problem, we introduce a variable number of auxiliary variables to create a family of Plug-and-Play (PnP) ADMM algorithms and unroll them into variable-complexity feed-forward networks, amenable to parameter tuning via back-propagation. More complex unrolled networks require more labeled data to train more parameters, but have better potential performance. Experimental results show that our unrolled networks perform competitively to generic DL networks in image restoration quality while using a small fraction of parameters, and demonstrate improved robustness to covariance shift.