An Unpaired Learning-Based Method for Image Despeckling

Speckle noise removal, or despeckling, is a crucial component of coherent imaging systems. In this work, we extend BD-QMAP [1], a theoretically grounded despeckling algorithm developed for stationary stochastic processes, to images and develop an image despeckling algorithm. Extensive experiments demonstrate that our approach outperforms classic Bayesian and MMSE despecklers. Furthermore, BD-QMAP, with its simple implementation, achieves performance close to that of classic image processing techniques and shows potential for further improvements to compete with supervised learning-based despeckling methods trained on large paired datasets—despite not requiring any paired ground truth and speckled data.