TDU-DLNet: A transformer-based deep unfolding network for dictionary learning

Abstract

Deep unfolding attempts to leverage the interpretability of traditional model-based algorithms and the learning ability of deep neural networks by unrolling model-based algorithms as neural networks. Following the framework of deep unfolding, some conventional dictionary learning algorithms have been expanded as networks. However, existing deep unfolding networks for dictionary learning are developed based on formulations with pre-defined priors, e.g., ${\ell }_1$-norm, or learn priors using convolutional neural networks with limited receptive fields. To address these issues, we propose a transformer-based deep unfolding network for dictionary learning (TDU-DLNet). The network is developed by unrolling a general formulation of dictionary learning with an implicit prior of representation coefficients. The prior is learned by a transformer-based network where an inter-stage feature fusion module is introduced to decrease information loss among stages. The effectiveness and superiority of the proposed method are validated on image denoising. Experiments based on widely used datasets demonstrate that the proposed method achieves competitive results with fewer parameters as compared with deep learning and other deep unfolding methods.