DDD2Self: Self-Supervised Image Denoising with D ynamic D ata-driven D ropout

Self-supervised denoising has gained traction for real-world applications where paired clean-noisy data is unavailable. While recent blind-spot masking approaches have shown promise, they often rely on fixed or heuristic settings such as uniform dropout rates and small-scale masking which limit stochastic variation and hinder adaptation to diverse image structures, ultimately affecting robustness and overall generalization. Additionally, focusing on isolated pixels or patches can neglect coarse-level image structures, reducing reconstruction quality. To overcome these limitations, we propose DDD2Self, a novel self-supervised denoising framework based on a dynamic, data-driven column-dropping masking strategy. Unlike prior methods, DDD2Self applies Gaussian-sampled dropout probabilities derived from column statistics, introducing adaptive stochasticity while preserving smooth dropping rate and ensuring stable training. By removing entire columns, DDD2Self prevents identity mapping while preserving the coarse structure of objects, promoting holistic learning for improved reconstruction. Extensive experiments on both synthetic and real-world noise datasets demonstrate that DDD2Self achieves superior denoising performance with faster convergence compared to existing methods. The Code is publicly available at: https://github.com/jubyrea/DDD2Self.