Structure-Guided Self-Supervised Deep Despeckling of Coherent Images: From Polarimetric SAR to Medical OCT Images

Abstract

Deep despeckling networks reach excellent accuracy when clean references exist, but coherent imaging systems never observe a speckle free scene, and self-supervised alternatives discard the spatial organization of the scene that classical region based filters exploit so effectively. This paper proposes a structure-guided self-supervised despeckling framework in which a hierarchical region representation of the image is injected into a compact convolutional network at three points: as guidance channels carrying the region model image and the region boundary map, as the anchor of a prior-anchored residual architecture whose output at initialization equals the strongest classical prefilter, and as a structure-consistency loss that penalizes output variance inside homogeneous regions. Training requires no clean references: the network learns from independent speckle realizations with prefiltered targets, and the same recipe applies unchanged to polarimetric SAR covariance data and to medical OCT (retinal optical coherence tomography) scans, since both modalities share the multiplicative speckle mechanism. Under a K-distributed product model protocol on ESAR Oberpfaffenhofen and AIRSAR Flevoland scenes, the guided network attains -4.42 dB and -8.61 dB relative error, improving on its unguided counterpart by 10.9 dB and 11.5 dB and reaching the accuracy of its classical anchor from a fully self-supervised deep architecture. On the public OCTID retinal database the guided network reaches 25.87 dB PSNR and 0.623 SSIM, surpassing all evaluated baselines, and multiplies the background equivalent number of looks of raw scans by more than twenty