An Optimal Control Approach for Inverse Problems with Deep Learnable Regularizers

This paper introduces an optimal control framework to address the inverse problem using a learned regularizer, with applications in image reconstruction. We build upon the concept of Learnable Optimization Algorithms (LOA), which combine deep learning with traditional optimization schemes to improve convergence and stability in image reconstruction tasks such as CT and MRI. Our approach reformulates the inverse problem as a variational model where the regularization term is parameterized by a deep neural network (DNN). By viewing the parameter learning process as an optimal control problem, we leverage Pontryagin's Maximum Principle (PMP) to derive necessary conditions for optimality. We propose the Method of Successive Approximations (MSA) to iteratively solve the control problem, optimizing both the DNN parameters and the reconstructed image. Additionally, we introduce an augmented reverse-state method to enhance memory efficiency without compromising the convergence guarantees of the LOA framework.