Reproduce “"Double-DIP": Unsupervised Image Decomposition via Coupled Deep-Image-Priors” 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
Yossi Gandelsman, Assaf Shocher, Michal Irani
Researchers at The Weizmann Institute of Science developed "Double-DIP," an unsupervised framework leveraging coupled Deep-Image-Priors to decompose a single image or video into constituent layers. The method achieved broad applicability in tasks such as segmentation, dehazing, and watermark removal, notably ranking second on the O-HAZE dehazing dataset with a PSNR of 18.815 and effectively handling non-uniform airlight.
Key insights
- The inherent inductive bias within convolutional neural networks (as utilized by DIP) is powerful enough to capture the simpler internal statistics of individual image layers.
- Coupling multiple DIPs, along with a multi-component loss function, allows for the unsupervised disentanglement of an image's constituent layers.
- A wide range of seemingly disparate computer vision tasks can be unified and addressed through this general unsupervised image decomposition framework.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.