Recursive Valid-Event Recovery for Fidelity-Preserving Event-Stream Denoising

Event-stream denoising can irreversibly remove valid events around sparse contours, weak textures, and fast-moving structures when filtering decisions are treated as terminal. This letter proposes recursive valid-event recovery (RVER), a fidelity-preserving framework that retains filtered-out events as ambiguous candidates. RVER performs density-adaptive reliability screening, recomposes candidates with sparse reliable context, mines short-window structural correspondence, and recovers candidates through neighborhood-consistent verification. On DVSCLEAN, RVER achieves 94.97% valid-event retention, 99.01% noise suppression, and a 96.87% F1 score; on DVSNOISE20, it obtains an average RPMD of 65.78. In EDD-AB downstream validation, full RVER reaches 92.05% mAP, 0.91 percentage points above initial denoising. These results support improved valid-event preservation while maintaining strong noise suppression and structure-sensitive downstream utility.