We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.CV

Change to browse by:

cs

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Computer Vision and Pattern Recognition

Title: Image Deblurring with Domain Generalizable Diffusion Models

Abstract: Diffusion Probabilistic Models (DPMs) have recently been employed for image deblurring. DPMs are trained via a stochastic denoising process that maps Gaussian noise to the high-quality image, conditioned on the concatenated blurry input. Despite their high-quality generated samples, image-conditioned Diffusion Probabilistic Models (icDPM) rely on synthetic pairwise training data (in-domain), with potentially unclear robustness towards real-world unseen images (out-of-domain). In this work, we investigate the generalization ability of icDPMs in deblurring, and propose a simple but effective guidance to significantly alleviate artifacts, and improve the out-of-distribution performance. Particularly, we propose to first extract a multiscale domain-generalizable representation from the input image that removes domain-specific information while preserving the underlying image structure. The representation is then added into the feature maps of the conditional diffusion model as an extra guidance that helps improving the generalization. To benchmark, we focus on out-of-distribution performance by applying a single-dataset trained model to three external and diverse test sets. The effectiveness of the proposed formulation is demonstrated by improvements over the standard icDPM, as well as state-of-the-art performance on perceptual quality and competitive distortion metrics compared to existing methods.
Subjects: Computer Vision and Pattern Recognition (cs.CV)
Cite as: arXiv:2212.01789 [cs.CV]
  (or arXiv:2212.01789v1 [cs.CV] for this version)

Submission history

From: Mengwei Ren [view email]
[v1] Sun, 4 Dec 2022 10:40:35 GMT (10408kb,D)

Link back to: arXiv, form interface, contact.