Trans-NLM Network for SAR Image Despeckling

Image despeckling is important to synthetic aperture radar (SAR) image restoration and related downstream tasks. Because existing SAR image denoising algorithms are difficult to simultaneously achieve high performance, efficiency, and interpretability, in this article, we propose a new image denoising method, namely, Trans-NLM, which incorporates the Transformer architecture into traditional nonlocal means (NLMs) filtering-based denoising algorithm. In doing so, the SAR image despeckling performance is enhanced, while the algorithm interpretability is retained. First, each pixel and its surrounding pixels are simultaneously mapped to a high-dimensional space as a neighborhood matrix through a shallow fully connected layer. Then, positional encodings are added to the neighborhood matrix, which is further mapped to the internal vectors Key, Query, and Value. The final vector representation of each pixel is also calculated according to the multihead attention mechanism. Finally, the representation vector is passed through a shallow fully connected layer for data dimension reduction to predict the corresponding pixel value. Moreover, layer normalization and residual learning are applied to accelerate the convergence. Experiments on both simulated and real SAR data demonstrate that compared with representative denoising models, for example, NLM, fastNLM, SAR-CNN, CNN-NLM, and MONet, the proposed Trans-NLM exhibits better performance in despeckling enhancement and efficiency simultaneously, along with more explainable inference process and transfer learning capability.