LowAux-RDNet beats reflection removal benchmarks with 27.5 dB PSNR
New model uses low-pass filters and scene-balanced data to outperform rivals on 5 datasets.
Jizhong Li presents LowAux-RDNet, a new approach for single-image reflection removal that recovers a clean transmission layer from photos taken through glass. The method builds on the existing RDNet decomposition pipeline and introduces LowAux, a training-only auxiliary loss that applies low-pass filtering symmetrically to both the predicted residual and the target. This stabilizes the learning of low-frequency components while keeping the original high-frequency residual supervision intact, leading to more balanced performance across diverse reflection types. To improve real-world coverage, the training incorporates scene-balanced real reflection pairs from the RRW dataset, reducing dataset bias and boosting cross-dataset generalization.
On a unified benchmark spanning CEILNet, Real20, Postcard, Objects, and Wild datasets, LowAux-RDNet achieves a five-dataset macro-average PSNR of 27.546 dB, SSIM of 0.9220, NCC of 0.9751, and LMSE of 0.004760—the highest PSNR, SSIM, and NCC, and the lowest LMSE among all compared public checkpoints and internal variants. Per-dataset analysis shows the main gain is more consistent performance across reflection distributions, though strong semantic reflections (e.g., text on glass in Postcard) remain difficult. The code is available on GitHub.
- LowAux-RDNet adds a low-pass residual supervision objective during training to stabilize reflection removal learning.
- Scene-balanced real-world pairs from RRW improve generalization across five benchmark datasets: CEILNet, Real20, Postcard, Objects, Wild.
- Achieves top macro-average PSNR (27.546 dB), SSIM (0.9220), and LMSE (0.004760) among compared methods on a unified evaluator.
Why It Matters
More robust reflection removal for real-world photos, enabling clearer images for photography, surveillance, and autonomous driving systems.