GPX-ADMM-Net: ADMM-based Neural Network with Generalized Proximal Operator

Shih-Wei Hu, Gang-Xuan Lin, Chun-Shien Lu

Abstract: In this paper, we propose a highly efficient and well interpretable deep learning solver, called Generalized ProXimal ADMM-Net (GPX-ADMM-Net), for the linear inverse problem, which is conventionally solved with intensive computations.GPX-ADMM-Net is characterized by the generalized proximal operator, convolutional dictionary, and modified loss function. Without loss of interpretability, GPX-ADMM-Net only needs a small number of parameters in a learning model to retain elegant reconstruction quality.Different from traditional optimization methods, all the parameters of GPX-ADMM-Net need no more hand-crafted but determined by learning strategies. Furthermore, unlike other deep learning-based methods, GPX-ADMM-Net is able to adapt to various measurement rates with only one single set of training parameters. Extensive experimental results further demonstrate the advantages of our proposed method.