Image deblurring algorithms can be broadly categorized into iterative algorithms that solve mathematical models and deep learning algorithms. Iterative algorithms regularize sparsity, with prior knowledge typically set manually, leading to limited applicability. In contrast, deep learning algorithms design networks based on empirical experience, resulting in poor interpretability. To address these issues, this paper proposes an interpretable deep learning network for image deblurring, named ISTA-DeblurNet. Inspired by the Iterative Shrinkage-Thresholding Algorithm, ISTA-DeblurNet employs nonlinear transformations to replace the L1 norm regularization term in deregularized models. Parameters such as nonlinear transformations, shrinkage thresholds, and regularization coefficients in ISTA-DeblurNet are learned end-to-end rather than set manually. This algorithm combines iterative methods with neural network structures, enhancing the interpretability of deep learning networks while improving the performance of iterative algorithms.
Dong et al. (Tue,) studied this question.