ErA network beats state-of-the-art on defocus deblurring benchmarks
Corrects kernel estimation errors with alternating updates and ResUNet denoisers.
ErA (Error-Aware Deep Unrolling Network) is a new end-to-end framework for removing defocus blur from a single image. Unlike prior methods that treat kernel estimation as a black box, ErA jointly learns a compact kernel basis and per-pixel weights. It incorporates an error-aware term into the Augmented Lagrangian unrolling process, which alternates between updating kernel weights and applying ResUNet denoisers to correct estimation errors. This iterative refinement leads to sharper deblurring results.
On standard benchmarks, ErA sets new state-of-the-art performance: 34.2 dB PSNR on DPDD, 31.8 dB on RealDOF, and 29.5 dB on RTF, with corresponding SSIM scores above 0.94. It also shows strong cross-dataset generalization on CUHK without fine-tuning. The method is particularly effective for scenes with spatially varying defocus, a common challenge in smartphone photography and microscopy. The authors have released both the paper and code via arXiv.
- ErA uses an error-aware term in Augmented Lagrangian unrolling to iteratively correct kernel estimation errors.
- Achieves record PSNR/SSIM on three defocus deblurring datasets: DPDD, RealDOF, and RTF.
- Generalizes to CUHK dataset without ground truth, demonstrating robustness across real-world defocus blur.
Why It Matters
Sharper photos from any camera: ErA fixes defocus blur in challenging conditions, improving smartphone, security, and medical imaging.