Deep Learning Methods for Sonar Image Despeckling: Advances and Challenges

Sonar image despeckling remains a significant challenge due to complex underwater propagation, with traditional methods struggling to maintain a favorable balance between effective noise suppression and fine detail preservation. Inspired by the advancements of deep learning in optical image processing, researchers have begun adapting deep learning models for sonar enhancement. However, the task faces unique obstacles, including the intricate physical mechanisms of sonar noise, the scarcity of public datasets, and the absence of clean ground-truth references. To our knowledge, this paper presents the first comprehensive survey dedicated to deep learning-based methods for sonar image despeckling. We systematically review both supervised and self-supervised learning paradigms, analyzing representative network architectures and their performance across various benchmarks. Furthermore, we catalogue widely utilized public datasets and their associated evaluation metrics, while providing reproducibility guidance for advanced sonar image simulation tools to enable the synthesis of highly realistic imagery. Finally, we discuss existing limitations, highlight key open challenges, and outline future research directions essential for developing robust and practical despeckling techniques required for reliable sonar image interpretation.