Sobel-Guided Patch-Replacement Autoencoding for Fundus Enhancement with Unpaired Real Data

Abstract

Real aligned low-and high-quality (LQ–HQ) fundus pairs provide direct restoration supervision but are scarce; synthetic pairs introduce a degradation gap, whereas real unpaired data lack correspondence for structure-preserving restoration. To retain authentic degradations while recovering restoration supervision, we propose Sobel-guided patch-replacement autoencoding (SPRAE), a two-stage framework for learning and transferring a restoration prior from real unpaired data. Stage I learns the restoration prior by reconstructing HQ content from Sobel-matched real LQ patches transplanted into HQ images. Stage II transfers the frozen prior to real LQ images through cross-attention over a structurally indexed HQ feature bank, while a reconstruction-derived structure-identity objective suppresses unsupported retinal changes. SPRAE achieves the best image-quality and downstream-task performance across in-domain and out-of-distribution evaluations.