CDLNet: Robust and Interpretable Denoising Through Deep Convolutional Dictionary Learning

Deep learning based methods hold state-of-the-art results in image denoising,but remain difficult to interpret due to their construction from poorlyunderstood building blocks such as batch-normalization, residual learning, andfeature domain processing. Unrolled optimization networks propose aninterpretable alternative to constructing deep neural networks by derivingtheir architecture from classical iterative optimization methods, without useof tricks from the standard deep learning tool-box. So far, such methods havedemonstrated performance close to that of state-of-the-art models while usingtheir interpretable construction to achieve a comparably low learned parametercount. In this work, we propose an unrolled convolutional dictionary learningnetwork (CDLNet) and demonstrate its competitive denoising performance in bothlow and high parameter count regimes. Specifically, we show that the proposedmodel outperforms the state-of-the-art denoising models when scaled to similarparameter count. In addition, we leverage the model's interpretableconstruction to propose an augmentation of the network's thresholds thatenables state-of-the-art blind denoising performance and near-perfectgeneralization on noise-levels unseen during training.