SAR-CDL: SAR image interpretable despeckling through convolutional dictionary learning network

Abstract

Deep learning-based approaches have shown advantages in the task of despeckling for SAR images. However, it is still difficult to explain due to the black-box nature of deep learning. Deep unfolding methods provide an interpretable alternative to building deep neural networks, which combines traditional iterative optimization methods with deep neural networks for image recovery tasks. In this paper, we propose an unfolded deep convolutional dictionary learning framework (SAR-CDL) for SAR image despeckling. A new variational model based on convolutional dictionary for removing multiplicative noise is proposed. The alternate direction multiplier method combining deep learning method are used to optimize the variational model, which can parameterize the model by deep learning in an end-to-end learning manner and avoid the large workload of the tuning process. The performance of the proposed SAR-CDL is validated on both simulated and real SAR datasets. The experimental results show that the proposed model outperforms many state-of-the-art methods in terms of quantitative metrics and visual quality, with a stronger ability to recover the fine structure and texture of the SAR images. In addition, the proposed SAR-CDL is robust to the size of the training set and can achieve appropriate results while reducing the training dataset.