Reproduce “Blind Motion Deblurring with Pixel-Wise Kernel Estimation via Kernel Prediction Networks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Guillermo Carbajal, Patricia Vitoria, José Lezama, Pablo Musé
This paper introduces J-MKPD, a learning-based blind motion deblurring method that explicitly estimates dense, pixel-wise blur kernels using a Kernel Prediction Network (KPN) prior to non-blind deconvolution. This approach, trained on a specially synthesized dataset, yields improved generalization and restoration quality on real-world blurry images compared to end-to-end deep learning methods.
Key insights
- Explicitly leveraging a precise forward degradation model within a deep learning framework improves generalization and interpretability in blind motion deblurring.
- Representing dense, spatially-varying kernel fields via a small set of image-adaptive basis kernels and pixel-wise mixing coefficients is an efficient and effective strategy.
- Joint training of the kernel prediction network and the deconvolution network is crucial for minimizing artifacts and achieving superior restoration quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.